Day 2 - Infrastructure and Ecologies | Future Cities Symposium & Exhibition 2025

Today we go right into the heart of what happens at fcl, which means the research. And personally this is what I enjoy most. Especially also in the realm of fcl, where so many disciplines come together to tackle, analyze and also propose solutions for urban issues.

Usually or mostly, we think of research as a practice where ideas take shape, where discipline brings new insights to its field. Yet, as has been continuously emphasized over the past day and also this morning, and as practiced throughout the past 15 years of FCL, FCL also questions.

Unfold across and between the supply and I think it is precisely in these intersections that insights emerge and also insights that can help decision makers and ultimately find their way into policy. A point widely discussed yesterday and one we will also turn, I think later during the discussion.

And fittingly, we also gather here at create a building and an ecosystem that FCL has been instrumental, but also in bringing to life. But it also has shaped FCL over the last 15 years. In this first parallel session, we'll explore three teams at the core of FCL's research. First we will talk about or our colleagues will talk about mobility, land use and infrastructure, how it can adapt under deep uncertainty. Then we have a presentation on how coastal cities face and respond to interacting climate stresses. And third, how ecology of cities shapes biodiversity, resilience and human well being. And together these perspectives show how science and design and engineering can guide more sustainable and future or positive or sustainable urban development. Just some logistics. Each presenter will have around 10 to 15 minutes to share their work. And once all presentations are completed, we will gather together here in front for panel discussion in the second half of the session. So I encourage you to listen with curiosity and without further ado, I would like to invite Prateek Mansel, professor at nus, and Brian Ad, professional at eth, who will talk about their research.

Thank you very much. That's a bit loud.

Okay.

Years of work in 10 minutes.

My name is Brian Ady from the ETH, as was just said, this is Pratik Banzal from the National University of Saint Singapore.

That picture up there is the Yurong Lake District, an area in Singapore that's undergoing large adaptation. Singapore does a fabulous job of adapting to meet its needs and we set ourselves the task in our module to figure out if we can help do that even better.

Our module is called Adaptive Mobility Infrastructure and Land Use and is looking at the combination of those three things.

We had a bit of a challenge to set ourselves an overarching goal, so we did. We said it is improving infrastructure planning processes. So what you do with everything that's in Singapore. So that organizations, so in this case organizations like the URA and lta, obviously we're not just in Singapore, so it's also the part in Switzerland with the AJA and ASTRA and sbb. But can they better meet the needs of their inhabitants or their stakeholders? And can they do it faster? Reaching things like carbon neutrality or improving travel times or helping to make people even healthier with active modes of transportation?

That's a big ask. We covered a large area from high up down to low down. In a way, we summarized them in these three points here that we are trying to better understand how mobility needs may change. So social behavior models, can we make the models better of mobility, infrastructure and land use and the trade offs between them? And then how would that help change the planning processes themselves inside the organizations or even across organizations?

I have two big pictures or we have two big pictures on the left hand side. And these were two areas that acted as inspiration for us. One is really the implementation of the cycling routes and the park connectors across Singapore and the other one is the fact that Singapore is trying to. One of the aspects of reaching carbon neutrality is bringing in electric vehicles and rolling out the appropriate infrastructure to make that happen.

To give you an impression of the people, there's a lot of us. We are 4, 5, 6, 7, 8, 9, 10, 12 people, us two. We have three co investigators that are spread around the world. We have our module coordinator, Che Ming Yi, which kind of made sure everything was working on track. Can do Kua Dang and our four PhD students, two here based in Singapore and two based in Zurich.

Just before. How we thought we would do it is we have then four slides to give you an overview of the research that we've been doing. And Pratik will start with the first two. So Pratik.

Thanks Brian. Audible. Okay. Okay, great. So as Brian said, the broader aim of the module is to think about the adaptive infrastructure planning. When we are doing any infrastructure planning, the demand modeling becomes very important. How do you model the mobility demand we have been building in the earlier phases the travel demand models for Singapore and around the world using activity based models. But in the last three years we have mainly touched on two key dimensions. And the first one is how do we bring the social networks into the activity based models? So what do we mean by bringing social networks to the activity based models? So the first aspect of having the social networks in the activity based model is generating a synthetic population which has the social networks where people are connected within the households and across the households, basically friends. Second aspect of developing these models is generating the activity schedules of the people where we can understand how people are willing to negotiate their travel times, time of the day for the activity and the location of the activities. So how do you account for these trade offs? When we develop these digital twins, these are like high resolution activity based simulations.

So this is the work led by Johan Jin, who is a PhD student at NUS.

She not only started thinking about how do you start collecting data for such things because there is no existing framework, how do we really collect the data. So it's not just a modeling problem, it's also thinking about how do we collect data for such things. So first part is collecting data for generating synthetic population with social networks and the activity schedules. And second part is the modeling which is equally challenging because it's a high dimensional feature space. So how do we bring that high dimensional feature space in the synthetic population and in the activity schedules? So these are the two key aspects that we have worked on. It's a work in progress as Zuhan is going to finish PhD end of next year. But we have the frameworks ready, we have the data collection completed.

And the first part which is the synthetic population has also been completed. I mean it's a bit of a.

I would say claim I would like to make here that there is no existing social network data across the globe for the mobility. So this is kind of a first social network based mobility data that we have collected in Singapore and it's fairly representative because it's for the around 1200 people. And we have tried to maintain the representativeness of the sample across several demographic characteristics.

The second part is also very important when we are generating these activity based models which have not been explained in the past 15 years, I would say 10 years of FCL and many other attempts to develop the demand models is how do we use these large scale data sets that we have been talking about yesterday as well. Like we have now our cellular traces connected, we have transit smart card, tap in, tap out data which has been collected. But how do we use these data sets in transport planning? Because you're just collecting tap in, tap out information or the traces information, but you don't observe the social demographic characteristics or the many other information which is important to generate the activity schedules. So we have on one hand household travel survey data which is extremely rich, but the sample size is too small. But we have very detailed characterization of the activity schedules.

Okay, we are back. And on the other hand we have transit Smart car data, cellular trace data, where we don't have these type of information but very high resolution spatiotemporal information. So there is an. And the second issue with the survey data set which we have experienced during the COVID 19, that there is no way to update your mobility patterns if there is some immediate shock happens to the system. Right. Because household travel survey data is collected every four years. And.

If, let's say in 2020 something happened, we didn't have any information about what's going on. So we said, okay, how about we think about fusing these two data sets together, which is household travel survey data and transit smart card data, and generate the time dependent origin destination matrices of the traffic flow flow and then from that start creating the full activity schedules of the people such that we can maintain those spatial temporal distributions that we observed in the smart card data and all the nice characteristics that we observed in household travel survey data. And that leads to two important things. First, you create very high resolution synthetic population and the activity schedules which can be updated every day basically. You can also see how the mobility pattern is changing from Monday to Sunday on the weekends and weekdays and which brings lot of advantages when we get to the planning.

Great. Okay. All right. So that's the first two slides. The next two research slides give the next one moves over to what we've been doing in the area of modeling. And here I said we're trying to improve the considerations of the mobility infrastructure and land use together. And part of that's really related to the fact that if you're trying to plan infrastructure, there's almost an infinite number of things that you can do with it. And it depends a lot where you build or make infrastructure depends on how you're going to use the lands where people are going to live, how you want them than to move with the mobility. So things like car light like we showed in the beginning with the new Jurong district.

I've got numerous things listed there and I'll highlight a couple of them. The first is to take it back to Switzerland for a minute, was really looking at how the feedback loops happen when you build infrastructure, how you actually change the behavior of people and you change the land use. And there we zoomed in on the city with the the smallest city in the world with a metro, and that is Lausanne, and looked at its growth over the last.

Bit. Challenged to say exactly when it started, but I'd say the last 50 years probably of when they were running it. And they're on huge expansion plans. So it Gave a nice natural experiment to allow us to identify what was going on so that we could talk a bit more about about what it means in the future and where even future expansions might be required. That's one thing. Another thing was these infinite number of combinations. There are lots of wonderful models out there like agent based simulations of traffic for all of Singapore for example. People can wish them to be a little bit better. They can always be a little bit better. That's why people love research. And we're doing our research and seeing if we can better model the social networks and how that affects choice behavior of how and where you travel to. But from a planning perspective, one of the things is these things take a long time to run. Agent based models of traffic take a long time to run. If you want to vary the land use and see what happens to transport, that takes even more time. And then if you want to throw in, okay, I'd like to do a thousand different ways to model infrastructure. What is that going to mean? Pretty soon you get to the point that like you might not even get to retirement before your models are finished. So we zoomed in a lot on how can we massively speed these up and we did that with surrogate models or machine learning based on some models that we already had developed and we found that we could massively speed up the time. We're not quite finished yet. Orlando sitting here worked on a proof of concept in a smaller area of Singapore and now Qiming is leading the effort to do the model for all of Singapore to just show how much faster we can be. Obviously we lose some accuracy, but how much does that matter? Because obviously running the really detailed models is very detailed but you just can't do it in a reasonable time. So we've got to figure out a better way forward.

Zoom in then on number four as well. And that is looking at. And that's what you see here on the graph or the figure is really looking at.

Specific modeling the rollout of the electric charging stations in Singapore or electric charging infrastructure, taking into consideration all of the different uncertainties. I don't know how many of you actually. Well, some of you are definitely involved in planning things for the long term. But the thing is you're just confronted with massive amounts of certainties. Uncertainties, sorry, not certainties.

Like you don't quite know when you're going to lose the screen.

But it's okay, we're adaptive.

But you have to worry about the power grid, you have to worry about where you can connect things. So how much space is there? You have to worry about what is the demand going to be? Are people really going to be. Is everything really going to be all electric? And if it is, which technologies are going to be in the vehicles and which technologies are going to be used to charge them? And so we have all of these uncertainties. And then this was a special effort of Qi Ming here to figure out the very best model to do this and then to do it quickly. He built it on top of using reinforcement learning to help accelerate everything and came up with a very innovative way to do it. And obviously there's more things to do, but it looks quite good.

The last bit on the research front is really, well, it's great to have models. I think everyone in this room probably develops models. A lot of people do data analysis, A lot of people do models. But then how are you actually going to change decision making? And here we thought we want impact, so we really want to be involved with helping organizations change how they're making decisions. And here I've called out a few things that I would consider the main contributions. Like Pratik mentioned earlier, not all of our work's done until end of October next year, or maybe a little bit longer, depending how things go. But we started here. I'll zoom in on number nine for a minute. Is we actually mapped the whole planning process of Switzerland, which is the first time it's been put down on paper in a. A public forum anyway, to say, well, you know, who are you trying to help in this whole thing? You're not just developing a wonderful model and hoping somebody's going to pick it up. Who are you targeting? And that allowed us to identify some areas of potential improvement. We have research with Jin really looking at, well, how can we use the models to help look at path dependencies, like taking really into consideration the fact that what you do today, what you decide today, limits your choices in the future. Now, you can't tell definitely what's going to happen in the future, but you do have a good. You could have a good idea of. There are multiple possible futures that could happen. And if you know that you can make choices today that help put you in the space, that you're ready to adapt easier in different ways in the future, that might be something that's better or more useful to you as a decision maker. A last thing I call out there, I can't call out all of them, is this little graph right here, which we thought was actually, it's definitely a unique thing in the world at the moment is, we said, well, when people start thinking about where they can expand infrastructure, change infrastructure, it's normally a very qualitative process. So you get a few people together, I'm making it sound too simple. And they will do some brainstorming and figure out what are the possibilities and then they'll do some evaluations to see if it makes sense and then it'll go forward. Actually, it's a long iterative process, but that's kind of the thing we said, well, that's good, but maybe they're missing stuff that would be even better. So we worked on an algorithm that would allow scanning of an entire area to look at where new infrastructure could go in and then gave rough estimates of how beneficial that would be over time. And this is something that is very innovative, very new, and hopefully we can evolve it further in the future to make. Make it even more impactful.

Great. Thanks, Brian.

So we thought like this is kind of a formal final session, so it will be good to talk about.

The concrete output that module delivered after less than three years around. We started in November 2022. So I would like to highlight that there are various types of KPIs and we have planned and attract two columns there. So the first one I would like to highlight is the publications. We plan to do 14 publications in journals. We have 12 published already. Three are at the revise and resubmit stage and many are at the submission stage right now. Another thing I would like to highlight is the, the media engagements. I had some engagement at the talking point at cna, then Orlando had another media engagement and Arnor had with Icelandic News Channel. So we have been actively engaged with as much as possible outreach and disseminate our findings. And at the same time, one important session that we did was the AI for Speed Smart Mobility at URA and which was well received. We have guests from industry as well as from URA and lta and I think that's the direction we are moving forward in. So, yeah, that's it. And we are happy to provide more details as required.

So you'd be happy to notice here we massively beat all of the KPIs, except that this one, although I hope that one's not a big problem.

All right, so we think that after these, we started quite late. So we're only about two and a half years into our project. And so we're quite happy with how everything's turned out so far and we've had a wonderful team that's worked great together. We by no Means believe that work is done and we think we can do a lot more. So we are gearing up for FCL4 to build on what we've been doing. We had a, which Pratik didn't mention specifically, but we had a wonderful collaboration with LTA and URA the whole time. We had regular workshops with them and exchanged thoughts and it brought us, we're both pretty new to Singapore back then, so we brought everything a lot, lot closer together. So we really understand the needs of Singapore a lot better now and adaptiveness. Singapore keeps adapting. If you do a quick Google on their websites, you can see a future town here, the Ecotown of Tenggae with 42,000 new homes across the five residential districts. You can see they're really changing Singapore. They're adapting to the changing needs and we think we can. Obviously you can see one thing that's starkly absent here is cars. Singapore is doing a lot to make car light districts and that leads us right into our proposal for next time.

So as Brian said for the FCL4, we are thinking of exploring the different aspects of the carlite districts, thinking from health perspective, thinking from environmental perspective. So we are planning to conduct first like some longitudinal studies to understand the causal effect of different type of carlite districts. Then we are planning to build upon all the platforms, the digital twin platforms that we have developed for mobility. We are planning to build on that and bring the health related digital twin on top of that to quantify the island wide impact of different car light scenarios and then finally come up with an evaluation framework considering the health, environment and mobility impacts. So that's what we are gearing toward and looking forward to it.

And the reason we're way over is because Julio told us that they weren't so strict with the 10 minutes.

Thank you.

Well, we can adapt, right?

Speaking of that, what is adaptive planning? Maybe it's a terminological question or conceptual one, but how is it different from traditional planning? Just very quick, if that's possible.

Very good question. It's probably more about the explicitness of being adaptive. I think for many people they would think right now, okay, we have a specific problem. We, we want car light districts and then they would go implement, implement, implement, implement, implement. And they don't think anymore about the multiple possible futures.

Obviously they will continue to adapt. So five years later, if all of a sudden they need something completely different, they will change. That's how humans work in their planning environment. The key of calling out the adaptiveness in the beginning is to say, hey Maybe, maybe, maybe you make bit better decisions right now if you realize more explicitly that hey, the future might not turn out exactly how you're expecting it. And you know, if you can imagine a few possible scenarios and you can, you might tweak your infrastructure design a little bit so that you could accommodate them more easily than you otherwise would have done. That's it.

Thank you very much for the clarification. I would suggest we go directly to the next presentation. Pengyuan Li from C City Interface, the stage is yours.

All right, thank you very much and it's such a pleasure for me to represent our module CCT interface and my name is Penguan Liu and I worked as a module coordinator for this module and now I moved to uk so I fly all the way back from UK for this event. So two days ago. So if you say I'm a bit on and off, it's because I'm still experienced at jetpack. But anyway, the CCT interface, this project is basically trying to understand the climate change and trying to mitigate the impact of the climate change through the urban planning and design process. This goal brings up more than 20 researchers in our module from Singapore side, from Zurich side and from international side. And we're trying to translate our common shared understanding of the coastal cities into actionable urban planning strategies.

So very first question, why do we want to study coastal cities? So coastal cities, the low elevation coastal zones, usually defined as those kind of areas, is below 10 meters of the mean sea level. Despite the fact that it only occupies 2% of the world's land area, it actually serves as 13% of the global population. And 2/3 of our mega cities are actually distributed around the coastal areas. And unsurprisingly, our city is still growing. By 2050, it is estimated that the urban growth rate is going to reach to the 70%. So it brings our coastal cities more important than ever before. And this undoubtedly it's almost like everyone now thinks that the coastal cities are the ancients for the blue or the green economy. It's basically serving as the main point for, for the shipping and ports and its green energy hubs. And there are so many potentials of the coastal cities. However, coastal cities have shared stresses because we are so close to the sea and because our world is getting warmer and warmer. For example, UK this summer is experiencing the hottest summer ever. So all our worlds are experiencing this kind of. All the coastal cities are sharing these stresses. It suffers from the hydroclimatic vulnerability, it suffers from the microclimate vulnerability as well. For Example the urban heat, it also suffers from the economic vulnerability because there are just so many population centered in the coastal cities. And also it has some shoreline vulnerabilities such as the mean sea level change and everything. Right? So how as a module we answer this question, how do we actually try to propose something new for us to understand the growing risk that we are facing? As we're living in Singapore, we're living in such a coastal city. We're trying to answer these questions through three different kind of aspect, Research studio and impact.

In terms of the research, we answer one core question, how to prepare and design coastal cities for the future. This is not only the question from our module, this is also a question from a survey that actually like more than people from more than 199 cities worldwide. Trying to find the common answers to this, right? How to prepare and design our coastal cities for for the future. In our module, the Zurich side is focusing on the manned hazard and developing methods to quantify climate change climate change impact in a multi hazard context. And in Singapore we are developing and test urban system designs that can be more resistant to the urban climate change.

As such, our research has a global vision. We're trying to develop new methods, we're looking at fine grained spatial resolutions and also we have an inclusive focus. All in all, our ultimate goal is trying to develop these digital twin systems that can help us to develop new sustainable urban planning strategies and trying to mitigate the climate mitigation. We're trying to study the urban form, we're trying to understand human activities, we're trying to understand the impact of the climate change. And through all this we're trying to propose new urban design and planning strategies.

As for the urban global vision, this is mainly the effort from our Zurich team. They started over 22 coastal cities across the world and trying to identify the common shared risks. And what they find out is that actually most coastal cities all around the world are suffering from the risks of the urban heat. This is so true in our context as we are in Singapore because Singapore is just too hot and it's getting hotter and hotter and we identify that trying to say now we have the shared risks through our all coastal city, how it actually impact to our urban everyday activities. As such, we propose new methods to address these kind of questions. This is one of the research that I was leading in this module. We're trying to. Well, the topic for this module is called Sensing Climate justice and we use a new graph modeling strategies or the modeling of the urban areas and border strength from the trending new data source such as street view imagery, trying to understand and classify.

The urban heat and flood vulnerability, how it's actually distributed across the urban areas through new modeling technologies and GeoAI. In the meantime, traditional technologies are trying to understand the urban area or understanding urban form using pretty much the street networks. But other networks also matter in cities. In this research, this is another paper just published a few months ago, we're trying to combine the urban street systems together with other nature systems in Singapore. And by using that we prove that a holistic modeling of urban area is can help us to better understanding the urban environment. So the output of this particular model on the left, we don't need any additional data. We can already accurately predict the land surface temperature across Singapore.

Apart from the in addition to our urban form modeling kind of techniques, we're also trying to identify the flooding risks across the nation. In this particular research, our PhD student Abraham Wu was developing this kind of GAN based neural network feeding the information that is often missing in the official report. He used the whole United States as a case study because the official report when it's reporting the flooding risk often focus on big cities or some data rich regions. However those small cities are often missing out. So he developed this new method in order to fill this gap. Using the GAN network, he could be accurately predict where those kind of areas might have the high risks of the urban flooding.

And we're looking at the fine grained spatial resolution as well. One of our PhD students Cuiyi, who is also sitting in this room, developing this new method, trying to downscale the local climate zones from the traditional grid based spatial resolution to the building scale so that in our future understanding of the urban heat impact of the urban area we could, you know, instead of looking at, you know, aggregated area, we could now look at building scale so that we could probably in the future to adjust our urban building energy kind of planning or indicators more accurately.

And we have inclusive focus. One of our PhD student Jen Yu is doing this kind of very, you know, time consuming studies by recruiting aging populations in Singapore over 65 years old. She trying to let them to wear the Apple Watch which is the wearable smart devices and also combined with the indoor sensors trying to understand how the urban heat.

Disproportionately impact the aging population's everyday activities. And this is a very long process. This study is still going but our results actually the current preliminary results is actually interesting. We identify that especially for aging populations who are maybe suffering from a not so well economic Background they might not stay at home long enough because the electricity cost is just so expensive. They probably prefer to stay in the hawker center to avoid the heat impacts. But we're trying to understand this and maybe in the future better to help them and support them to have a better healthy life. And in addition to our research, we also have the studios. The design studios are basically a test bath for our master students trying to make use of our models and some of our technologies to help them to understand the rising risk in coastal cities. We have this kind of test what if scenarios for the pluvier flooding using the 3D models that we have for Singapore. And we also have this kind of. There's a very fancy kind of simulation going on is that we're trying to say if we can use the 3D models of the urban area to see how the wind flowing. For example, if we actually change the building angle to change the building morphology, would that be better support the wind corridor so that Singapore can be cooler? Since we have those kind of aim to actually around 5 degrees of temperature in Singapore, which I'm not sure if that's going to be achievable. But hopefully through our future planning and simulation strategies, we can propose better idea on how do we actually eventually achieve that. And we're trying to also using the 3D model to understand the urban heat, how it's actually distributed in the urban area. And students use this opportunity to explore where they might feel more comfortable when they're looking for accommodations. But through all this brings up to our impact. So one of the impacts that I think well as particular problem from myself, I'm pretty proud to say that two module coordinators from this module got faculty positions in prestigious universities. Now I am a lecturer assistant professor at the University of Glasgow and our previous module coordinator, Professor Zheng Xing now is assistant professor in.

Cityu Hong Kong. And this goes credit to our mojo. And this also goes credit.

For the FCL and the sec. And we very much appreciate that.

We co organized workshops with Tongji University this June using the method and also models we developed. We hosted this kind of eight days workshop at Tongji Yu and let our students to explore the new method trying to design new strategies to support a more sustainable city.

And we also have collaborations with the local agencies and the companies overseas. For example, in our design studios we have URE visits to visit us. And we also hosted a company visit from China Southwest Architecture Company last October. And they are very interesting in what we're doing and we are building this kind of Long term collaborations with them and well taken together I think there are just some points to summarize. We are identifying the common shared risks for coastal cities and we think the sea city phase is dynamic. It's constantly shaped by our human activities and how do we deal with the nature environment and this kind of climate change, our impact population in, in an uneven manner. So different groups might react to the climate change in a different way. And we believe that integrated perspectives are essential and there is always place for more innovations. And we believe in the future the collaboration is key and I want to collaborate with everyone here to probably support a more sustainable city development. And thank you very much. And this is our module and hope you enjoy it. Yeah, thank you.

I have a very quick question before we move on. And this would be. So you show very nicely the complexity of coastal cities around the globe. And then also a big aspect of your research is about climate change. What according to your research is the specificity of coastal cities being exposed to this. You speak about the interface and these dynamic interface. Maybe you could just quickly jump on that.

Okay, well, so I think, well first of all like different kind of team members of our team looking at different aspect of these kind of challenges, right. But I think the primary thing that we are dealing with is mostly focused on the urban heat and how could we actually understand a more fine grained urban heat at a more fine grained spatial resolutions and how do we actually react to that? And the reason why I was saying that maybe the CCTV is dynamic because if you actually look at Singapore, right? Singapore.

The beach site of Singapore is actually pushed a lot towards the sea. They states, because the new urban strategies are trying to develop new.

Shore areas and those are raising so many questions and also so many challenges, but also brings a lot of opportunities to bring people in. And then there are so many new things could be going on there. And we're trying to understand how human activities or the new sort of like urban form, how they actually shape the new those kind of boundary areas between the city and the sea probably. And I'm not going to say that our mojo is doing a perfect job or we done that 100%, but I think it's still exploring process.

Thank you very much.

Thank you.

Hello everyone. I'm Deng Kai, the coordinator of Comparative Ecology of Cities module. So I'm going to present our module's work on pattern function, on exploring pattern function relationships in urban ecosystems across cities.

So we are a large interdisciplinary team led by Professor Tam Pollock from AUS and Professor Paulo Blando from ATH Zurich. Our team spans around 8 country universities organizations and we have in total of 16 members. Team members.

We bring together urban ecology, urban climate, hydrology, public health and social economics. So these multidisciplinary disciplines because cities are complex systems and no single lens is enough.

Our project is grounded in the landscape ecology theory which provides the conceptual framework for examining urban pattern process function relationships in cities. Urban patterns are the spatial arrangement of green and built elements in the cities which can affect a lot of urban processes, which are flows and exchanges of materials, energy and information, which in turn will affect important urban factions or outcomes related to livability, sustainability and resilience.

Currently, the evidence about urban pattern function relationships is quite scattered due to the majority being case studies. Therefore, we aim to general to derive some generalized for pattern function relationships across cities. Specifically, we ask how urban patterns affect urban functions across different scales. What urban patterns, what emergent patterns can be generalized across many cities and what principles can urban practitioners use in the urban planning practice? Methodologically, we combine machine learning, simulation, curiosity, comparison and synthesis analysis to address all these research questions.

We analyze different urban elements including green space, street networks and buildings. We not only look at two dimension but also three dimension of green space and buildings. We use landscape metrics to quantify their spatial characteristics in terms of the size, shape, complexity, fragmentation level, aggregation level and connectedness level and so on.

And also we use spatial analytics to quantify the street structure, centrality, public transport accessibility and land use in density. So all these urban pattern metrics provide us a common language to compare very different cities.

On the function sides. We simulate pluvial flooding using physics based engineering model. We estimate the building energy demand by developing a novel coupled urban climate. And the building energy model.

We look at even species richness using around 27 million birds records in 60 cities. We use around 12 million small businesses in 80 cities to quantify the small business density as outcome of social economics. We investigate the mortality risk of humans in around 6.3 million adults in Switzerland and around 400,000 under five children in 57 countries.

These function outcomes were then linked to the urban pattern metrics that I mentioned earlier in the total of 100 cities in the world. So this multi city design led us to assess whether the pattern function relationships across all these cities are robust versus context specific.

This slide summarize our work packages and the workflow of our project. So we in parallel we analyze pattern function relationships in five individual domains including hydrology, even biodiversity, urban climate and energy, public health and social economics. And then in the Second part of the project, we conducted a synthesis study across these five different functions and the Design studio. In the following slides, I will briefly introduce the key findings of these work packages. We'll start with the individual domains.

In the hydrology work package, we found that impervious edge density is negatively associated with pluvial flooding, but green space aggregation can reduce flood volume. And we also found that the effects of green space aggregation actually depend on urban expression level. We found that when the urban coverage is about 50%, the effects of green space decline sharply and when it's over 75% it becomes negligible.

For the even species richness, we identified some very important urban pattern metrics affecting even species richness, including tree can't be height and its variants root length and the patch area and aggregation of tree canopies. So again, we found that effect size of all these important urban pattern metrics affecting even species richness varies across local climate forms and ecological contexts.

We found that trees can also reduce can also reduce building energy demand among cities. Among cities, we found that trees can provide the largest energy use intensity reduction in hot desert cities. Within cities, we found trees provide the largest energy use intensity reduction in open low rise neighborhoods.

Trees also matter in terms of improving public health. In the public health work package, we found tree canopy can improve the tree canopy configuration. Especially tree canopy aggregation can potentially reduce mortality both adult and children mortality risks. But the effect size of tree canopy a configuration also varies across local environments, tree canopy cover levels and urbanization levels. We also estimated how many under 5 mortality can be prevented if we improve the tree canopy aggregation globally. But we found that this figure also various cross countries depending on the current tree canopy situation.

And finally, in the small business work package, we found that compact and workable urban areas can support local small business development and we also found this effect. These effects of different urban pattern elements actually vary across different cities. We found urban densification is more effective in more developed regions, while public transport improvements such as improving public transport accessibility can be more effective in less developed regions in terms of improving local small business development.

So in all these individual domains we were able to identify some key urban pattern indicators affect affecting different urban functions in many cities. Based on this, we conducted the second part work of our project including a synthesis study and design studio.

In the synthesis study, we conducted a comparative analysis between Hong Kong and Zurich, two very different cities. So we aim to identify the two synergies and trade offs across these five very different functions. In terms of effects of urban patterns.

We found very consistent urban pattern function relationships between these very two different cities. In terms of the multifunctional performance, we found impervious services generally constrain this multifunctional performance by a water proximity to generally it enhances multifunctional performance. In terms of green space configuration, we found that it can lead to multifunctional trade offs.

At the final stage of this project.

We assessed.

How the urban how the research findings can better inform urban planning practice design practice focusing on four one by one sites comprising both brown fields and green fields designed for future towns along the Hong Kong and Shenzhen border.

So students.

Redesign these four sites through an intuitive planning process guided by two key instruments. The first one is the qualitative guidelines derived from the research findings of our individual domains and the second part is about quantitative dashboard developed based on our synthesis approach. So this dashboard approach is a machine learning based approach and it can predict the performance of different functions for the improve or for the proposed plans of the size.

According to the students. Our approach of the dance studio is very informative. It brings bridges the disciplinary gaps making multifunctionality legible to urban planners. It also supports the decision making when comparing very different schemes.

So to summarize, our project is able to identify some very important urban pattern metrics affecting different urban functions across many cities. Our comparative study across cities approach.

Is a very useful method to uncover these patterns even though the patterns are not always uniform because.

They can depend on a set of urban contextual factors. But I would say also I think the most important one of the most important part of our projects are that we demonstrate that the methodologies and the observed patterns in this project are applicable to design process.

So we after the two year and a half work hard work, we got some very important deliverables. Our project enhanced the understanding of complex urban systems. We demonstrate the proof of concepts implemented in practice. We generated global urban pattern data sets which will be coming soon and we developed a building energy demand model urban multifunctionality assessment tool and finally we got 14 publications and presentations.

Lastly, our project was able to help a very successful project symposium in June this year. Our symposium attracted more than 90 registrations and 50 of them attended on site. We also got representatives from URA and HDB who share their expertise and the perspectives in how research findings can inform better informed urban planning practice.

These are some publications and the presentations and I would also take this opportunity to thank the wonderful team, CEC team, all of the team members, their hard work making all these outcomes possible. Thank you. This all about our sharing.

Thank You. Thank you, Denkai. Thank you for showing us this impressive work. I have just a very quick clarification question. You show 120 study areas.

How did you choose them? Just out of curiosity, Randomly or was there a pattern behind.

Thank you for the question. Thank you. Very good question. So actually the main goal of our project was to develop some generalized or urban pattern function relationships across different functions. So the first step is to select as many as possible cities, especially representative cities. And we also need to consider that these cities shouldn't be uniform in terms of urban context factors or some socioeconomic status. We need to make all these factors diversity so that our generalizable urban plant functions relationships are reliable across the cities. So based on this, we conducted a very deep search for these.

Many cities in the world to see how their urban forms.

Vary between each other. And then we select a set of cities that share the same urban pattern factors, but also the diverse aspects of the urban patterns to make sure that the and also other factors were consider, such as climate, context or other socioeconomic conditions. So we make this uniform, both uniform and diverse, to make sure that our concurrence are reliable. Thank you.

And I invite all speakers to come to the front for the panel discussion, which is the second part of this session. So I will start with Adrian, who has a burning question.

So thanks for these interesting insights. I have two questions.

One question to the group working more.

On adaptation and what I see a little bit the direction you're working on is trying to assimilate data with sensors or whatever to get in real time some information. But looking at how people adapt this, for me it's more a reactive planning than an adaptive planning, because adaptive planning needs some future.

Shared vision to develop.

Plans which are really adaptive to these changes. So I was wondering, how do you define really what you're doing in terms of the time frame you're looking at? Is it really adaptation in terms of that people can adapt or is it a reactive effect then because they might.

Have more data and they know that flooding is coming, but.

They haven't adapted to it.

Because they hadn't had a plan of the future. So I'm a little bit, I don't really understand why.

Digital twins or these very fast feedback loops you're trying to do are really helping me to adapt.

If they're used in real time as.

A decision support system, it's a little bit. Going back to your question, I think, what's adaptiveness? So yes, we're using data to improve on the simulations that we can do. So part of it is just understand how do people change their behavior. And so with collecting the data that we see in different areas and improving that modeling a little bit, it allows us to say, well, there are these things that could happen. These things that could happen are these things that could happen. Because not all behavior is the same.

That's one bit. But that all goes more or less towards improving the models that we have. One thing, the second thing is we really try to.

We think models are quite useful to understand what's going to happen. And we have found.

Frustratingly also in our module even that we just. There's no way that a reasonable group of people can use very sophisticated models in some sort of joint decision making environment. It's not possible because you have, okay, let's have a little conversation about it. All right, we'll come back next week and then maybe we'll have some results. And we think that is relatively disaster. So it's great to have the detailed models, but we have to find a way that we can play with possibilities of what we could do. So the second thing is there speeding up the models greatly so that we can have something that we can use that gives us a reasonable idea of what would happen if we play with some of the parameters. Now you're right, we don't really know. But that is where that last bit really comes in and say, well, all right, we could pretend that we're just going to want this and that this one scenario is going to happen, but we all know intuitively that whatever we plan for exactly is not going to happen. Something else will happen. And so we wanted to push people through suggestions for improving the process of, well, how can we imagine more the different possible scenarios of happening? And then we can think back to, okay, what is it that we do today not just to enable that future, but to enable that future, but to make it maybe easier to change in a different way if the other future happens. Now part of that as well is the integration of really the temporal aspect from the planning thing, like any big thing that we plan. And a good connection with The C interface 1 is the rollout of the protection on the E East coast. Like you can plan it, but It'll be done 20 years later. I'm not sure the exact time frame on that one. And that's the same with all these big infrastructure changes. Think about when things might need to be done. So then you can also collect some more data over time to realize which of these multiple futures might be unfolding. So it's still a big messy process. I don't want to say that it's some magic bullet or anything, but that's the whole idea of trying to be more adaptive in our module.

So just to summarize, the model you're.

Developing is to understand the robustness of different solutions.

But.

The digital idea of having all the time improved situation is not to be in that sense at the moment more adaptive.

It's more to improve the model to think about different solutions.

Fair? Okay. Yeah. Okay, good.

As well. Just wanted to share some perspective and I thought that, no, the difference between being reactive versus being adaptive. There's a temporal dimension which was also mentioned that. No, if you look at the coastal protection efforts for the east coast or in fact throughout the course of Singapore, we are anticipating climate change to impact Singapore in a significant. And whatever plans are being developed now, do you call them as adaptive plans or do you call them as reactive plans? I think it's very difficult to put a definition to it. I suppose reactive means that we are reacting to drivers that are already present and affecting and impacting the way the city operates. And therefore reactive changes have to be done fairly quickly. Adaptive seems to me to indicate that something is coming on the horizon. We do need to change the way the built environment is being designed and modified. That's being adaptive. But the temporal dimension seems important in answering this question. I don't have an answer to this, but I think that's important consideration.

Can I have just a second and then I'll shut.

Can I just jump in on that just before.

So I think different people would have different opinions of it, but in a spontaneous reaction, I think I would say, well, reactive. If it was not really seen that it was coming. And I suppose depends on your time frame you're looking at. So if you look back 20, 30 years, probably no one was thinking about sea level rise and then all of a sudden they realized it was coming and then they said, okay, we have to do something about it. So in a way that's a bit more reactive. Obviously it's an adaptation. The other one is, well, hey, we know we have our scenarios of what's coming and we have planned out what we would do in different situations and that would be more adaptive, I would say.

And I think that infers that we need to have the ability to modify our plans and adapt to changes are coming so that the ability to have the space to modify. Because built structures are very resistant to change, that once it's built, it's very difficult to change until the next redevelopment cycle. So having the flexibility to change, I think is an important part of adaptation.

Can I just ask one question to you? I think with all these matrices, I.

Was wondering which one was surprising because.

I had the impression the results are.

Like, okay, well, more greenery.

More trees, larger patch. Was there anything you had the impression, like combination of factors where you were like, we hadn't thought about this or.

Was it the support of the classical things? We were expecting that more greenery, larger.

Patch, better connected patches will be better for ecosystem services in that sense.

Yeah, I think that's an important question.

I would say that the results.

Or the study led to some uncovering of things that we're not expecting now. One of the important metrics that appear in different functions that tend to be important is the age. Density and age conditions, for instance, that had a surprising effect on PULVIR flood volume. The age conditions also had important effect on which is related to aggregation and connectivity, on mortality related metrics. But beyond the metrics that emerge as surprising, we also were able to show that even for a single dimension domain like mortality and how green spaces can reduce mortality risk, the skill analysis.

Became so important. So one example that we are now also trying to put into the paper is that for instance, if we analyze mortality at the individual household scale, when we draw a buffer around that person or household compared to analyzing the metrics at aggregated, say municipality level, the differences are quite stark. Aggregation is important when we're analyzing at the individual person scale, but aggregation becomes a detrimental effect when we're analyzing at the bigger scale. And there are plausible reasons for this. So this kind of, I suppose new insights they were uncovering from a project tell us that nothing is simplistic. That really when you're applying research from applying results from research that we need to be quite cognizant that what skill are you talking about? How does it affect planning? Those are the insights that we're trying to put into a paper. So writing up in the Moon.

Yeah, super, thanks.

Okay. I think we are on a very interesting topic and it's actually something that cuts across all the presentations, which is modeling. So I would like to go a bit deeper on that, especially on the different approaches. I heard social network, I heard feedback loops, guai also different techniques that inform your modeling. And maybe Penguin, we can start with you.

Gui, I know you are one of the leading.

Figures in that emerging technique. Do you want to. And you'd only mentioned it one or twice, maybe you can tell us a bit more what is the difference to spatial analytics, what makes Guai different? What does it help in modeling other aspects than we've seen with more traditional approaches?

Sure, sure.

Well, first of all, I wouldn't consider myself as a leading figure in that, but I do have like few years experience in urban modeling with geo techniques. So I think the fundamental question is that why do we want to use AI to help us to study urban issues, right? I think many of our urban issues or the urban phenomena are present in nonlinear manner, right? For example, if you're looking at the traffic, despite the fact that I'm not really an expert in the traffic, but we know from the urban planning perspective is that the traffic volume in urban areas often present in a non linear manner and then traditional model, probably back in 30 years, 40 years ago, usually using the models that assumes the underlying assumptions is that these kind of phenomenas may be presenting in a nonlinear way. However, that's not the truth, right? So we want to understand this nonlinearity that existed in any sort of like urban phenomena. And that's why AI comes to help us to solve that. Because AI with this capability to incorporate the nonlinearity help us to address the urban issues better. And in terms of the urban modeling techniques that I am particularly interested in, I am more of like using the graph modeling techniques to model off urban phenomenas. Because urban area city is a complex system and we want something that to simplify our understanding of the urban systems that why the urban graphs comes to play the role. And with the urban graphs we can simplify many things. For example, we simplify the complex geometrics geometries of the buildings into single points. Or we can conceptualize the graph, connect the street network connections at nodes and graphs. And we saw that with the new trending technologies such as the graph neural network, we'd be able to make better predictions to capture those kind of underlying nonlinearities that incorporated into the, into the, into the urban phenomena to provide a better or accurate estimations or understanding of the urban area. Another core aspect is that despite we are planning the urban area, the urban planning process is a geospatial kind of process. I'm coming from a geography background per se, so I'm looking at the urban planning from a very like scale that is like very large from usually from urban scale. So when it's close to the urban scale, it's like a geospatial process. It's a geography problem per se, right? And then the we want and in geography we have a fundamental law which is the first law of geography. So that everything that close to each other will be presenting more similar kind of features. And we're using the model techniques while we're using the graph. We want to capture those kind of neighboring effects of if we want to say like we want to study a set but it might affect it by all the neighboring features. We want to use this kind of new graph modeling techniques that can capture those kind of impact from all the neighboring things and trying to, you know, just trying to present our urban system in a more formalized or even more systematic way, if that makes sense. Yeah.

So the AI is it like it helps to put sense in the non linearity of data or.

Well.

It'S definitely the nonlinearity of the data. But on the other hand, because different data will present it comes from different ways. Also another kind of new opportunities to do the multi dimensional data fusion so that you can integrate data from different kind of sources and also aspect put them together sending to an AI model and then see how it reacts. It's not always working. Right. The data quality matters, the data sources matters. But I think through the calf or sort of like engineering of the data and the modeling techniques we might eventually have something to build up.

Just the last point on that. What kind of AI model is it like we. I mean it's. I imagine it's not the familiar large language models that we are all used to. It's do you design it yourself? Do you or is it like GIS or something like that?

Right. So if I'm going to talk about myself, right. I am more familiar with a set of graph neural network called graph neural networks which I'm super enthusiastic about the graph modeling of the AI systems but obviously like AI is just more than. Is much more than the graph neural network. There are like LLM, you were just mentioning there are the GAN networks that one of our PhD student doing in the group trying to, you know the one with trying to map the flooding in the national scale. And there are a lot of kind of things but the targeting of different questions is really depending on the questions about like which kind of task you're doing. And in terms of the urban planning area, I feel like maybe especially when it goes to a trending topic of generative kind of design. I feel like the GAN AI is kind of standing out on that and there's just opportunities and I think it's for us is to explore basically.

Thank you.

Still on modeling, Pratik, you mentioned synthetic population.

I can't imagine anything.

I remember you talked about the social network and there's little data on that or an approach which is quite novel. But maybe could you explain what synthetic population helps us find out or you in your case?

Yeah, thanks for your question. First, I wanted to comment on the non linearity aspect of AI. I think it's true that AI can capture some of the non linearities, but we had non parametric models to capture that before AI. Right. That could still capture the non linearities if you put the constraints there. Right. So one of the major issues with using AI for anything nonlinear, is it a real nonlinearity or is it some obtained from some sparse part of the data which is not really a true nonlinearity? So for example, to give a concrete example, if you fit the income and education data on an AI model, you might see that as education increases, the income decreases because there might be some part of the data where you know, there are some people who didn't get a lot of education, but they are doing businesses and have a lot of income. So AI will pick those nonlinearities and you will see in some spectrum something like having weird phenomena. So I think that's where there is a need to bring the analytical approaches with AI if you want to do more interesting things. But obviously as soon as you bring analytical approaches, throwing in that black box is not trivial because you have to put some constraints on the functions that you are learning. So I think we have been doing a lot of work on, because I come from econometrics. So when AI comes we don't take it for granted. We have to make sure those theory driven constraints are satisfied whenever we are using AI. So bringing those theory driven constraints in the AI models is something I think is very important if you are generating any data or trying to fit any model. So that's on that part. Now regarding the synthetic population, if I broadly talk about the synthetic population, it's like let's say you have 6 million people in Singapore. Your objective is to generate the those 6 million agents in computer which has the exact same demographic distribution, homework location distribution as you observed in the population. So that's what actually synthetic population in principle means. But let's say if you collect the survey data from.

1% of the population or 0.5% of the population, you are not going to get all possible combinations of the demographic characteristics of the people. For example, it is possible that you might not have people who have very low income, but have cars in the population and maybe traveling a lot with the flights. So these kind of rare groups are difficult to sample and not just rare if you're just having the 1% population and you're looking at generating synthetic population which is consistent with the actual population across, let's say 50 attributes. You will miss many combinations which are not observed in the data. So it's a data problem that there is not enough combinations which you want to preserve in the population that you want to generate. So then people said, okay, how about we use generative AI for that so that we can go beyond the sample. The let's say we have the sample boundary and our population boundary is very big. But what happens when you try to generate using any AI based model? You will generate many agents with weird attribute combinations. For example, you will generate the daughter having higher age than your father. So these kind of weird combinations will be generated because there is no way that AI may know these kind of semantics. And that's the major problem with GANs. So you can increase the diversity of the population by extending to more attributes, but you will also generate these weird infeasible combinations. So now to improve this further, LLMs might be helpful because LLMs have some of the semantic relationships in them that, okay, I think if it's a daughter and father, then father should have higher age than and daughter, right? You can manually put these combinations, but it's too many if you're looking at 50 attributes. If you start putting these constraints, they explode and it's manually not possible to do that. So now the next step is to think about how do you combine your traditional generative models with LLMs to generate synthetic population which are diverse, which have all attribute possible combinations that you observed in the population as well as feasible agent attributes. So these are the directions in which the literature in synthetic population is evolving. Now coming back to the social network part in the social network, it becomes a graph problem that I just mentioned. So now the idea is again thinking about the. But there is no data, right? Maybe large language models have some data about the social connections that, okay, the female and male in this age group are likely to have these many ties or these many friends, for example. But this information is. Even if LLM throws this information to you, there is no real validation of this information. So you still need some data to first validate the information from LLM. And then maybe you can rely or fine tune LLMs to generate more social network data from from LLMs and use those attributes as a part of the synthetic population. So now the attributes that I was saying that this agent is age, gender and the income and various other characteristics. We will also tell that who is friend with whom and who can go for activities with whom. And we are not matching individual level, but we are matching the distributions for example, on an average male with this income category, with this age is likely to have these many friends in this category. So these distributions are matched when we generate those 6 million agents and we try to group them. So that's the main idea of generating synthetic population and how AI can help and what LLM can be helpful in.

Thank you very much. I think I learned something here. Here Tanpuyok and Denkai also you mentioned machine learning simulations.

Do you want to complement the picture? Is it. I mean for me who is.

A non modeling background, does it speak to.

You or do you use similar modes of reasoning?

Thank you, thank you for the very good comments before. So actually in our module we also conducted some very good work, both modeling and geoai. And I would like to pass these questions to our co investigator, Professor Simona Fratecki who supervises the work package of climate urban climate and also our research fellow Dr. Zhu Yue for doing the Geo AI work.

So in our work package I think we have been using machine learning AI mostly to understand connection which will not be clear in system like the effect of greening on health or aspect related to urban morphology and pluvial floods where.

There is not a clear theory or.

Conceptual model behind that we can exploit.

So we kind of outsource this type.

Of.

Modeling to the eye to figure it out the connection between these aspects.

Which in a certain way is giving us the capability to identify things we will not think about.

Like the hedges, the role of the hedges between impervious.

Impervious space that Professor Tampuajoco was mentioning before.

But the same time it's limiting us because we cannot really.

Explain too much.

Of why we see certain aspects which.

I think is one of the limitations. So in domains which are closer to.

What I'm doing, where we have a.

Mechanistic understanding like modeling of a herbal.

Microclimate or role of vegetation urban microclimate.

And we extend it to building energy.

Consumption, we try to go with more traditional mechanistic process based approaches because in that case even if the predictability may.

Be lower than AI approaches, the R.

Square is lower but it gives a.

Very clear mechanism, interpretation and understanding of what's going on. Which I think was one of my.

Questions to the overall audience.

So how you kind of push this AI? I mean it's very good in predicting.

Things but sometimes it's very difficult to.

Beyond the learning of these nonlinear relation.

To understand why they're in place and.

If it's just about making prediction, it's fine.

If I'm an authority that needs to.

Make pluvial flood predictions, I'm probably okay with having a certain AI.

But if I have to understand a little bit deeper to plan and modify.

Things or be prepared for situation which.

Are not in the current data, I.

Think that's where we have a challenge. And I'm probably turning back the question to the overall panel. But.

Interesting point does.

Yeah, yeah. I'm the research fellow in the group of cec. So because in our project that's the main idea is try to have a better understanding about the pattern function relationship to identify the most important indicators. So for us it's very important that we need to have explainable model. So for example that we also have researchers in our groups using causal machine learning to understand the relationships and we also use deep learning models, but mainly in terms of to solve data availability issues to improve the quality of the data. So I guess that there's also an area about one.

To decide which approach will benefit most to understand how to model the urban model to model the urban systems.

And so I also want to ask a question about the panels here about if you also counter this wondering about when the advanced technology can help us to to model the urban systems more accurately. But if this expendability is still unresolved, then would you also want to search any other approaches to complement the current advanced deep learning technologies?

I think there are at least two questions here and I want to give the word to the panel whoever wants.

To go first.

I can answer question on the interpretability part. So I think it's a fair question how do we handle these infeasible relationships that we obtain? So some of the work and that's one of the reasons AI has not been penetrated so much in economics community, not so much this technology rely on their fundamental models because they worry the relationships that they are learning might not be what they are expecting.

So one of the things that in economics the utility theory says that as you consume more then your utilities start decreasing that you gain from the extra apple, right? So if you eat first apple then it gives you more utilities second apple. Maybe your utility of eating apple decreases like basic microeconomics principle. So how do you account for these things when you are doing AI related models and building these behavioral models, for example? Because that's what I work on. So we started thinking of creating combining the existing parametric models with the AI. So for example we created a piecewise linear function and created the constraints on those pieces of this piecewise linear function that the.

Slope should be decreasing over time. Then you can put these many piecewise linear functions together as the layers of the neural network and train them using the traditional AI based approaches. Basically bringing your existing theory driven knowledge to design the architecture of the. The AI can help you put those constraints. When you train the AI models that, okay now the slope should decrease so that constraints in place. You use those constraints when you are doing your gradient descent algorithm. So you are basically able to handle those constraints and able to maintain the flexibility at the same time. So that's the way forward. But again that requires beyond going beyond just the packages.

You.

You have to understand your problem fundamentally. And that's why I still tell PhD students that AI will not solve all the problems. You have to learn those fundamentals before you can use AI in a more creative way.

I think regarding to the questions here, I think there are two aspects that I probably want to mention. First of all is that when we're talking about interpretability and go beyond predictions, right? There is a new set of technologies about expandable AI in the field. People using XAI trying to understand the feature importance, why its model makes certain predictions, in what ways. There was one of research that I was involved in is trying to make use this kind of feature importance beyond just saying which important which features makes more use of the for the prediction. But if we actually manipulate the feature importance, you know, manipulate the features by generating more counterfactuals within the AI models, can we for example like provide at least the preliminary kind of indications for the planning practice? Like for example, like we. If we looking at the say I don't know, just a simple example is that we saw the street view images, right? And then we want to say okay, well if we increase the number of trees here in the images, will these areas or at least our understanding of these areas change to another sort of like functions or whatever. So in that way we sort of like beyond a little bit to the just prediction self, but also provide more sort of like interactions with the model. And then it brings to. Actually the second part is that actually we need human to be in the loop. Is that we need human to. We should have the model that involves the human loop that human's feedback matters to the AI output. And meanwhile I would say that beyond predictions and also classifications, whatever the uncertainties, quantifications here is also very important when you are thinking about applying the models into the real world. Cases such as I believe When Professor Patek talking about synthetic populations, uncertainty estimation going to play a huge role because you want to quantify those kind of impact so that you don't simply make those kind of mistakes. So I think we are still in the phase of exploring and.

Maybe in the future some of us sitting here might develop something more concrete to address that. So yeah, it's quite looking forward to say that, to be honest.

Thank you. There's already another question.

Could I just respond to this set of comments and questions from the ground? I think it is not possible for a single project to be able to undertake all aspects of.

Research including uncovering mechanisms. I'll just say on behalf of project that the approach primarily we've taken is inductive in nature. We are at the moment trying to look at patterns that we believe can be generalized. And if there are patterns that can be generalized using the comparative approach that we have undertaken, then it may be possible to do further tests to uncover mechanisms. All the work or most of the work, not everything, most of the work we produce or share is really correlation. We don't make a claim that they are causative in nature. But because we're able to make a claim that some of the things we observe are generalizable, then there may be some interest later on on to begin to develop theories and mechanism to try to explain this. So hydrology is certainly looking at biophysical processes, human, the social economics is social, human behavior. And then for avian biodiversity we have looked at ecology to try to understand what about behavior of birds that lead to the kind of patterns you observe. So I think this is what the project is. I don't think we are able, at least for a project to get to the level except explaining how things happen. But I think it's really a big step to say this is what we have observed.

Thank you. There's one question there and one there.

I have a follow up question on the explainability of geoai. I think one of the most fascinating feature of GEOAI is that it can learn the latent features and of the complex real world features and convert them to embeddings. But explainable AI often relies on.

The features, the features that human can understand. So the geo AI, they normally have more predicting powers, but feature engineering will have more explainability powers. Is there a way that the two approach can converge? Like could we explain the embeddings learned from e.g. graph neural networks?

Well, I think. Well, it really comes to the AI part of this conversation is that first of all I want to Say that there is a set of GUI called perturbation kind of based. That means that you try basically the model do itself internally is that they're trying to modify the features that are already there and then to say how the model actually react to the predictions and then give the importance based on that. Right. So obviously, yes, you are correct that AI technology generally when this goes to the urban system, generating kind of like embeddings that can sort of like representing the urban areas into the high level. Right. But then in this case that we use the embeddings so for the downstream tasks such as classification predictions or whatever. Right. But then if we actually apply those kind of perturbation kind of based interventions to the AI systems, we'd be able to say which features are just more important to whatever. So in that case. Well, I don't know if I'm saying that correctly, maybe people may disagree with me, but.

I do not necessarily think that data engineering and the AI prediction are two separate kind of systems. They are essentially a coherent kind of thing that when you develop an AI system you probably want to consider sort of like two together, if that makes sense. Do you want to say something?

Yeah. So I think that's a fair point. I generally agree with him. But I think the key point when you mentioned about can be interpret embeddings. Right. In most of the applications that I have seen, people first get those latent spaces and then in the second stage they use them for prediction. But if you can do the jointly the creation of the latent space and the use of that for the predictions together, if you try to minimize both those functions together in your training, then your embeddings are more interpretable and you will have better chances of figuring out the latent features importance for the prediction. But that training is bit complicated. Jointly we have done some work on using aerial images to figure out the factors that affect the route choices and then estimated the model simultaneously the joint training and it helped us improve the understanding of the latent spaces better. So I would say the joint training helps in making the embeddings more interpretable.

Thanks. There's another question.

Thank you for the interesting discussion. I want to ask questions about the design process because I noticed all of the module have like a design studio integrated in the package. So when you created this kind of model and you implement it in the design studio, how do you oversee the way we designer, change the way we think and we operationalize the design process to accommodate for the adaptability in terms of longer time frame? And multidimensionality of the problems that you all cover.

Considering that, as Prof. Hwanyuk mentioned, our built environment is very resistant to change once we build it. So I just want to know from your design studio, is there any particular insight how we can integrate it better and change the way we think about our design process?

If the question is clear, yes, I.

Think I got it. Maybe I'll start first. We certainly thought of our work as providing a tool to help design. It is not a tool to dictate the design outcome. So the tool that we created enables the students in a safe environment of education to explore what are the impacts of different design features they have put in considering the five domains that we have covered. So in that way, through the iterative process, they can evaluate if I change certain things as part of my design, what are the likely outcome at the end? The whole end of this process. Also have to say that no design is largely a normative process. It's not meant to be fully dictated by.

Quantitative or measuring outcomes because the qualitative aspects of design is so important. Like for instance, it's so difficult to design for beauty. But when someone looks at beauty, they know that it is important. It touches the heart. You can't quote it at the moment, as far as we know, you can't quote it in any kind of template. So it is a combination of therefore being clear that what we provided, it is a tool that is guided by the normative design outcomes that's already described in the brief or in the future by your client. So the short of this is that think of it as a tool that we help students to design or think about outcomes, possible outcomes of what they're designing for. It's a tool. It's not meant to dictate outcome.

Anyone else.

If not, I think we've covered a lot the modeling or different approaches, or to use supernal's word, dispositions of modeling. There was another thread I wanted to touch upon, but maybe we are short of time, but there's the different domains. You know, we heard traffic, engineering, we heard biodiversity, health, socioeconomics. So I wanted to ask what is the benefit to look across those domains? But maybe if we have some time left, we can go on that.

But first I was really interested because the overall theme of these two days is science, design, policy. And I know that you all had also experiences with policymakers or there's been some knowledge transfer, some discussions with agencies. And I would really be curious if you could hear some testimonies, some experiences, what does it mean to exchange with stakeholders in government? I know Professor Tonpignoc has also a dual role in some degree. I know that Brian and Pratik had some vivid exchanges with LTA and ura. Maybe like we can start here and would be curious to hear your experiences because it's also about.

Translating research into other practices.

Sure. Thanks. I can start. Pratik can chime in.

So from my point of view, everything we were trying to do in our module AMIL was really about impact. So trying to help people improve their, their decision making and it's pretty hard to try to do that if you don't talk to the people who are making the decisions. It's relatively illusory to believe that you'll develop some analysis and some models and then at a final event they will come and then they will.

That's great. And then they'll take all your stuff and they'll spend all the time to try to figure it out when you're not even there. So we, in our module, we embarked, as I said at the beginning, we didn't have. We were not well connected. So we spent a lot of time.

With URA and lta, having them at workshops, explaining to them what we're doing and getting feedback from them. It's not always easy when you work together with stakeholders because you have some ideas of what would be helpful and they don't think it would be helpful. But that's part of the whole process of. I think we probably tweaked some of our research.

In midstream as we're moving on to try to be more helpful. And I think over that time we've really developed a good relationship and now our new proposal is a lot more targeted to their specific needs.

I'd add there though that another nice thing that's happened through that and that's a lot because of Pratik and his role at nus. It's kind of hard to do it from Switzerland. It spawned a lot of other.

Connections that led to very targeted research being done directly. It's outside of fcl, but it's strengthened. I don't know if it's instigated, but it's strengthened and everything through the fcl. So I think in that type of situation it's been quite a big success.

Very interesting. Thank you, Pratik.

Thanks, Brian. I would like to add one thing that we always care about is data and getting that from agencies, which is important for policy making. So I think when, as Brian said in the beginning, it was difficult to connect with agencies and asking for data because we were also not sure what they what would interest them. So initial dialogues helped and we started working this data fusion work in South Korea first because we have some data for Seoul, we had smart card data and the survey data and we started showing that how this data can be integrated together and how the digital twins can be created. And that's when they figured out that oh, this might be useful. And then they started interacting more and started giving access to Singapore data. Now we have significant amount of data for Singapore from LTA and we try to meet them every six months to talk about the research that we are building in and how this could be useful to them. I mean, I wouldn't say that it has been fully translated what we are doing, but I think the part of the reason is if you want to make something translatable, I think we need to make a lot of validation because the bar is very high to implement something or publishing something. These are two different things. Maybe you can publish something at the r square of 0.6 0.7 but if you want to, that's my experience. Maybe someone has different experience about it. But if you want to implement something, they want it to be more perfect in terms of if it's quantifiable to an extent. So the validation becomes very important. So now what we did is we got external data sets for both Singapore and Seoul which are not. Which we are not using and trying to validate those approaches in a hope that maybe that convinces LTA and URA to use our models and tool directly in house without worrying too much. So I think it's a process of.

Learning from each other and figuring out how you can convince them that your methods are still useful. But I had long sessions, I think that helps. I had three hour long sessions with them some time to go through the research in detail and they have shown interest in that which I was surprised. So I think it's about engaging and having those long conversations to see if there is a synergies and something that can be converted into the policy in that. And we are still doing back and forth and let's see how it goes.

In the future like breaking up the silos and the mutual learning. Very interesting, thank you. Prateek.

Pratik mentioned a little bit with respect to the data bit and also through the longer workshops and discussions. But I think one thing that's very important and that's also why it's very, very good that FCL4 is coming. I mean if you want to have impact and change. It's going to happen in probably two ways, but one of them is really going to be through the relationships like planning infrastructure takes an awful long time and building infrastructure takes a long time. But establishing the necessary relationships to help change how people are thinking and transferring stuff, that also takes a long time. Not as long as infrastructure, but a long time. So, so you can't show up one day to the next and then say, hey, look how great this is, and then expect everything to change. And they're not going to give you data if you just show up and say, can you give me a whole lot of data? That's not going to work either. It's something that builds up over time. And so yeah, I think that's worked out quite well. I said the other way. I said there were two ways. The other is really.

I say in a lot of cases, I think through the fcl, a lot of people are developing models. We have developed lots of models, but I don't expect or want. I mean there's the data parts of analysis, but a lot of the models are not going to be picked up because they have models and they have large teams of people developing models and they spend a lot of money to hire companies to develop models. But through these relationships they get new ideas of which way things can go and that influences their thinking. Sometimes short term, sometimes long, sometimes in two or three years down the road when the right time comes in their development process and where we help out there a lot as well, I think is the people that come through here and is working with them and through those relationships, I think some people have their own relationships when they go on to their other posts and some people even go on and work at organizations like the URA and lta. So I think that's what I would have to say.

No, very good. I think there's not one way to bridge this gap or have this translation. There's also different.

Ways of making that happen. Yeah, very interesting.

Penguin.

Well, I'm not really in a good position to answer this question. Well, given the fact that I'm an early career researcher and also I only joined the the project halfway but doing my one year experience with the module, we did have developer collaboration with Ares recently. We submitted a proposal together. But generally speaking, I think one thing, given my very limited experience, I could say that at least Singapore agencies do have this kind of soil that's seeking for innovations and collaborations with our researchers with which I think is great, but we are still exploring this kind of collaborations. Together. So we're going to say personally, I will see how that goes. And if there are some obstacles, apart from the data issues you just mentioned. Yeah, I just probably pass to Professor Tampuyo.

Yeah, thanks. I have to say that no, our project didn't set out to to do policy translation of the research as a big part of the project focus. Largely because the project is inductive in nature. We are exploring, trying to uncover new things. So things that we uncover need not be at a stage where they are immediately actionable and translated to results or policies. So this is our project. But I do agree that if you want to move into a space, the realm where your research becomes translated into to policies or adoption for use as models, then building relationships is the first thing you have to do.

I do also question in my mind sometimes when you go into that space, is it research or is it consultancy and research? I'm not very clear myself sometimes to get into. Because you're advising based on your technical expertise developing and co developing a solution with an agency working on. Is that consultancy in some form? Maybe it is, right, so many is for discussion.

No, I agree because I mean as researchers one thing is to do like fundamental research, like white research, questioning the fundamentals, posing new questions. And obviously most of the time agencies have very practical problems that need direct solutions.

And as researchers, I guess you're not there to provide a service, but it's about building capacities or instigating new problems to speak about or new or generate new views on aspects. But Denka, you want to.

I also want to share some my experience in this project because at the beginning of this project I asked our PI, Professor Tamboyak. So it's very important to look at urban patterns and functions. These relationships are very important. But in practice, how do inform urban planners like you should improve tree canopy aggregation into like 80%. How does this really look like in the real world? And it's not as strict straightforward as tree canopy coverage, because I work on trees. So I take this example. So who suggests that you should have 30% tree canopy coverage in your neighborhood? But I also. My research tells the urban planners you should also improve the tree canopy configuration. So this is very.

Since the beginning of this project we have been working a lot on this perspective. We try to build relationships and we also want to this translated into practice. And I think there are two main parts. I know this is very important, but also very difficult. But I think there are two main aspects. So we built based on our relationship analysis, we are able to raise Awareness of the importance of urban patterns to the urban planners. Something that they haven't probably they haven't thought a lot before. And the other thing is that since during the past two years and a half work we were able to quantify some to quantify the contribution of improving urban patterns to the improved urban functions. So. And we are on the way to move.

So I think in the future this will come true that we can really.

Our research findings can be really translated into research into urban planning practice.

Thank you very much. Are there maybe some questions from the audience? We have a few minutes left.

Derek.

Yes.

I thank you for all your presentations. I have one clarifying question to the sea city interface to Mr. Pengyon Liu in your presentation, I felt there was.

A sort of.

An accent to urban heat situation. And.

How is that, let's say, how.

Can you link that to the.

City sea confrontation? Because the sea in a way is.

Of course sort of mitigating elements in the urban heat.

So Madrid has a greater urban heat.

Island problem than Barcelona for instance.

How does it relate with the flooding.

Situation in these.

World cities you're exploring?

I didn't understand that completely.

Right.

So thank you very much. So in terms of the urban heat, the heat issue is actually if you're looking at the mega cities that distributed worldwide, I mean it might not actually be applied for all of them. But if you're looking at Singapore as a case study here, Singapore as the island city, despite we got the sea area surrounded by sea. But Singapore is hot and because of the intensified urbanization and also the tall buildings, people often saying that, at least what I heard, I'm not So sure if 100% true is that Singapore is almost like a city with no wind. Right. And then.

This kind of environment, built up environment actually intensify this kind of threat that comes from goes to the urban heat, despite we are surrounded by the sea. So our study basically, at least in Singapore side we're trying to understand this kind of impact, you know, but in terms of the risks, because I think this is more on our Zurich team's effort is that they were trying to say, for example when the urban areas expanding their land, especially when you let's again in simple cases is that when they expand the land areas they are causing some like the internal issues which I'm not really an expert in that.

That might leading to the careful design might leading to certain sort of.

Flooding issues that might happen. But you know, and also in Singapore, you know, one of the issues here, the intensified rain, especially in the monsoon seasons caused a lot of like, flooding issues as well. Yes. So one of the reasons why we want to, for example, downscale the local, the local climate zones from the grid sites to the building size is trying to sort of, sort of like, say if there is any potential that we can mitigate that at least. But I'm probably not really answering your question because.

Well, yeah, I was asking it because.

The midterm review, one of the comments.

For this module was where's the coherence.

In the whole program? There are many things that are being.

Covered, but where is the coherence and.

That that improve in the last two.

Years, let's say the last phase of.

The research?

Well.

I think practically we are collaborating with the coastal protection agencies, I've got its name. And we're also trying to build up. Well, who are also kind of trying to build up these kind of digital systems in the west coast area of the park and trying to say they build up this kind of large system, trying to say that. And we want to sort of like.

Play a part in that. And we are constantly developing our collaboration with that. But at least for the. Yes. Well, I think it's still an ongoing process. I would say so. But at least we are looking into this kind of issue and trying to make the whole package more coherent. Okay, thank you.

Thank you.