0:00 Hello everyone, done, thanks, my name is Gabriel Garcia. My background is architecture, so I studied architecture, and then I did a master degree in geomatics for the build environment. Then I joined Royal Haskoning three, four years ago, and it’s a consulting engineering firm, so we are doing, at least in my department, we are doing buildings, or engineering buildings, mostly. And I’m going to talk to you about computational design and machine learning in me systems for large scale architecture, so very long thing, but we will go through the title along the, the presentation, but I would like to start with machine learning, so I would like to put, we talked about AI, we talked about machine learning, we talked about everything, but I would like to put a name on it right now.

0:51 These are the plugins that we investigated, we are very happy, so some from Microsoft, most of them from Food for Rhino, so we are using a lot Grasshopper. And we are very happy with AUM and Tony, so these are the tools that we are using for machine learning.

1:11 Then another concept, MEP, me stands for mechanical, electrical, and plumbing, and it’s the, the ugly part of buildings, right? So this is, is, is, is basically what you’re seeing, it’s even pixelated, so that’s why I left it there. And all these systems are behind the walls, behind the ceilings, are hidden, and, and, and, but actually these are vital elements for buildings, right, same as we have vital organs, all these elements are vital for a building. And I made, since I started working with airports, large scale architecture, I just realized that, or I found the similarity between the human body and those large scale systems, just to make an example, the Pomu Center in Paris, they put all the MEP in the facade, it’s just the one of the best examples that about the complexity of buildings and MEP systems.

2:05 This is another image that I tried, I made with AI, in this case on internet co-pilot, trying to change, or to switch the veins for cables, and try to make a comparison between the human body and, and these systems. But we’re going to talk about cables at the beginning, right, these are data cables, this is just an, an image I found on the internet, the servers are behind the curtain of cables, and just to give you an idea, this is what it should look like, right, or something like this. And this goes to my first topic of my presentation, I will present three topics, this is the first one, and it’s data cables, right, so all these data cables that are hidden behind the walls, behind the ceilings.

2:55 And just to give you a bit of background, a bit of context, this is just a floor plan of a random, actually a hospital, and we have a technical room. Then from this technical room we define the network, basically the corridors, because in all these buildings the cables are traveling through the corridors, and then we start defining the shortest path, we use the shortest spth algorithm, in this case AAR, to find the shortest path from the technical room to another room in the floor plan. Then we do it another time, then we do it another time, and then we need to find the shortest path from the technical room to any other room, or to any outlet, data outlet, in the building, could be for computers, cameras, any device that needs to be connected wirely to the internet.

3:41 What are the constraints about these data cables? The maximum length of a data cable should be 90 m, not more than that, otherwise the installation cannot be certified. There is 180 cables connected to a rack, and then a maximum of four racks per technical room, right, there are certain constraints. And just to make it a bit graphic, we have in the bottom, I, I can, I maybe I can use this, yes, so we have here technical room, and then you see that the, the red, no, sorry, the green and yellow is everything within 90 m, all the red dots are longer than 90 m. So we already know that in this floor plan we need two, three, maybe four technical rooms to have a full coverage, right?

4:36 What are the goals? So of course, in a large scale building like an airport, we need to minimize the number of technical rooms, minimize the total cable length, make it, make the calculations in 3D. What you saw before was a 2D floor plan, but now we need to make it in 3D, and we need to make, we need to provide interactive feedback to the architect, because at the end the architect is the one defining the spaces, and we need to tell them where to position these technical rooms, because a technical room is like 8×4 m, for example, so it is a room, yeah.

5:04 First use case, Foster and Partners, an airport in Poland, and just to give you the, the, to give you a scale of the building, 1.4 km from side to side, right. The step one, what we do at the beginning, is to define a few areas depending on the function, and then we translate this into points, we make simulations for data outlets, for example, one point every 5 square m, or one point every 10 square m. Next step is, on an early design stage, we run a simplified simulation in Tod, and this is just to understand, more or less, in a very early design stage, how many technical rooms we may need, right, and this is just cayman’s clustering, it’s just reorganizing the data and trying to find the clusters.

5:54 In a, in a later design stage, we use these clusters from a 2D calculation, and introduce shafts, shafts are these yellow elements that are used to, I will remove the point now, these yellow elements are the, the, the risers, so cables can travel vertically. So a technical room in the ground floor or in the basement can fit the first floor, second floor, and maybe third floor, and create like kind of a Christmas tree shape, right, of all those cables. And then we go for a short Spath algor, so we run again the shortest P algorithm, we use a bit of machine learning, and this is already full resolution, you see the amount of points that are in a single floor plan, and, and you see also how the, the areas are clustered and distributed by the algorithm.

6:49 But to show you how it works in 3D, I made this animation, because also I needed to show this to the engineers, the experts in the company, they, U, they told me, we need to make this tool, and then they click a button, and they were complaining, oh, why this takes 30 seconds, or why this takes a minute to calculate, and he’s like, because I’m doing this in the background, right, finding all the shortest path from all the technical rooms to all the outlets, which it’s a huge amount, amount of data. And finally we export the coverage areas, yeah, this is just to graphically show what is the coverage of every technical room, yeah, in the airport.

7:37 So in comparison with a traditional process, a traditional process was, maybe place the technical room, make a radius, that’s the coverage area. In comparison with that, we reduce the total cable length from 50 to 20%, that means that if there are 2,50 kilm of cable in the airport, and we reduce by 50%, which is somehow the minimum, 322 km were reduced of data cables. Technical rooms were reduced from 5 to 10%, so from 200 technical rooms in the airport, where we saved 10. And the response to the architectural teams is, yeah, it, we pass from two weeks to a couple of days, so we can respon, we can give feedback right away to the architects, and with a high level of accuracy. So yeah, I think we, we, we are happy with the results.

Second topic, this is a guy not making a selfie but searching for Wi-Fi signal, right, so the topic number two is the Wi-Fi signal attenuation by wall materials. This is a very simple graph, showing that if you have the router in this position, as soon as you go far away, 1 m, 2 m, 4 m, 8, whatever, and then if you find an obstacle like a wall in between, so the, the, the signal is decreasing, there is an attenuation, and then until you find 68, minus 68 dB, which is the minimum, right, after that you don’t have, you cannot connect to the internet. And to make it a bit visual, so this is just a Wi-Fi router traveling through different parts of a floor plan, different materials, glass, wooden, and brick walls, and you see that the signal changes depending on the position.

9:29 And this was a calculation that we did, we placed different technical rooms in the floor plan, and now you can see that we can calculate the overlapping areas, which we are overkilling it, you can see the coverage, and of course it’s just one way of representing all those signals in a floor plan, in 2D. Of course there is another way, in this case we place the same, this is just to show you the amount of calculations we are doing in the background, every point in those stars, or those Wi-Fi routers, has a value, a decibels, so with this information we can easily create heat maps, and heat maps are very well received by engineers, they like to see hit maps. But in this case, somehow, the, we, you cannot see clearly the overlapping, etc, it, this is more expensive in calculation. So fine, depending on what is the, the, the scope, we can use one method or the other one.

10:29 What are the goals here? So minimize the number of Wi-Fi routers, maximize the coverage area, of course, and minimize overlapping areas, because of course, once we run a simulation with machine learning, we just went for a lot of calculations, and we could have a, the best solution, or the most efficient solution, because you have a lot of routers, a lot of coverage, you have very little amount of routers, very little coverage. So, so we have a par front like this one, and somehow the best areas are here, so you need to balance between the amount of routers and the amount of coverage area you, you have.

11:10 What were the challenge in this use case? The machine learning process is not linear at all, this means that you have two routers, and then in the next iteration you just move it 2 m, and it’s on the other side of the wall, the outcome will be completely different. And then if you move the other one on the other side of another wall, for the computer it’s a completely random calculation, even if it’s just one meter aside or one meter next to the other, so it’s a multiobjective optimization, so that’s another level of complexity. It’s expensive in calculation, so expensive, I would say like 10 seconds each iteration, which is expensive for us, because of course we need to run a lot, and then the signal reflections are still to be included, right, if you have, if you put a router in a corridor, the signals will bounce, some of the, some of the signal will pass through the walls, and another will just bounce in the wall, so it’s overlapping right away. It’s work in progress, very complex, I would say, but interesting also for computational design.

12:24 And then I go to the third topic, I would like to attach to the previous presentation, I think we worked also in this building, were made by B, so you know already about the data, the, the timber structure of it, but in, in, in, in our case, we calculate sunlight reflection, and actually only for this tiny part of the control tower. So we, we, the question was, explore the glare disturbance in airplane landings, for those 72 panels, there were only 72 panels, simulate some reflections for every minute of the year, so you, we need to calculate a lot, data visualization, of course, is heavy when you have that amount of data. And the normative was saying that the calculation should be in 2 miles from 50 ft above the landing threshold, using the standard 3° G path, I will explain this in a minute.

13:18 I think it’s, so the terminal will be here, right, in this, in this area, the, these are the three runways that we need to calculate, or we need to investigate if are disturbed or not, but the actual calculation is in the yellow arrows, these are the yellow arrows, represented 2 miles away from the starting point of the runway. And to make it more graphical, we went Rino now and we just draw it, so these are the two m, every blue line or blue stripe, it’s two, 2 miles long, and it’s already at the right height, depending on the, on the, on the height of the terrain, etc. And the next step was to calculate the reflections, for this we use this method, so basically, if we have, if we have a triangular panel, we have a triangular reflection, right, if the angle is less than 180°, of course.

14:18 And this is how it looks, to make this animation it was too much to calculate, 10 million, to, to visualize 10 million reflections, so in this animation there is only 3% of the data, 3.2%, it was made not every minute, but it was made every 5 minutes, and just a single day, a week. And the colors, you see that is changing colors, it’s colorized by the, the, the panel number, the hour of the day, or theth month of the day or the week, we can colorize by, yeah, we can filter as we want.

15:00 Then, if you wonder how, so a s, what is the, what are the reflections of a single panel, this is how it looks, so this is a single panel with all the 3%, of course we still in the 3% through the whole year, and it’s colorized by the hour of the day, so the, the green side is the early in the morning, and then the red side is late in the afternoon. So this is the whole spectrum of reflections for a single panel.

15:29 And then I just made a very quick animation, so frame by frame, panel by panel, you can see that some panels reflect a bit in the morning, a bit in the afternoon, depending on the position, if, if there are south, north, it’s just more or less to understand what was in that big sphere of reflections. But let’s go back, and we go to the now, the, the landing paths, we intersect all those lines that you show in the big sphere into these stripes, right, and this is just a quick example, another example is this, so you see that the, the triangles are drawn through the stripe if they intersect. And then we F, we figured out another way to visualize this, so we made it 2D.

16:16 And you see these three, I don’t know if you can see the airplanes, but I will try to zoom in over there, so you see that these are the landing paths, so it’s when the airplane, it’s about, about to arrive to the r way, to the runway, and these are the two miles, so full stripes, and this is how it looks once we intersect all those triangles, it looks like this, and if you want to zoom in over there, in that blue rectangle, this is how it looks. But this is, of course, not realistic, because these are all, of course, it’s 3% of the data, and it’s all the reflections of a year into a single spot, right, so we need to find a way to visualize it differently and a bit more wisely.

17:04 So we made this kind of clouds, what you’re seeing right now is every color represents the reflection of a panel at a specific hour, or in a specific week, or in a specific month, so it’s just filtering the data in a way that we can understand, yeah, we can understand the data. Basically I zoom in a bit closer to, to one of those landing paths, and you see that here, in the right side, the triangles are very narrow, so it’s just a bit of a reflection on every triangle here, but here this is very dense, this may be a problem, but this doesn’t represent a big problem, right.

17:53 And this is just for comparison, if you want to run an early design analysis, you can use 3% of the data, which is on the top, and then you see that you have a few triangles, but the, the, the area is quite okay, and in the bottom side we have 100% of the panel, so it’s a lot of computational effort, but it can be used. And finally we use machine learning to, to find this, yeah, to analyze the data, there was a lot of information in these reflections, and with machine learning we can, we can figure out what is the worst moment of the day, and also what panel, what is the worst moment of the week or the month, what panel, at what time, or just define what is the worst panel.

18:38 And we can provide feedback to the architect, on telling them the panel number 32 has a lot of, a lot of reflections to the runway, or the landing path of the runway SE, and then maybe you can, or tilt it a bit, or just make it less reflective, right, so this is a very good, to the, to the architect, or to the designer, to understand what they can do.

18:59 An open question before closing, we didn’t calculate bounces in light, the normative is not that strict, so bounces are not yet there. We understand that it is like a diamond, or it will be like a diamond when, while the, the airplanes are landing, so they will see a lot of shine, but we don’t know how bad, and this is just an open question, because it’s a work in progress, we are understanding how to deal with that, because this changes the whole story completely. We are not calculating triangles anymore, we are calculating rays, points, and then every, the same panel may reflect to different directions in the second bounds, or third bounds, or fifth bounds, right, because there are also photovolatic panels in it, so it’s an open question for now.

My conclusions, so as per today, processes using machine learning algorithms need help from humans, it is not a plug and play process. So you can download a Posum, it works like Galapagos, just plug, it’s a multiobjective optimization tool, very nice, but if you don’t have a very good process behind, is not going to work. In the case of a Wi-Fi signals, we struggled a lot, because it’s just not optim, it’s not, it’s not learning, right, so the, it’s difficult, so I, I think at this point my conclusion is that it’s 50% humans, we need to tailor our model in a way that the machine learning algorithms can perform at its best.

20:41 Buildings are evolving, we are making buildings very efficient in energy consumption, embodied carbon, so materials, and the space optimization, finally, yeah, we are going to a data driven construction and engineering process, right. I think that’s it, 20 minutes, 21, that’s it.