Episode 7 “Machine Learning in Building Design and Operations”
Featuring Gary Chang
Kirsten
Hello, I'm Kirsten Nielsen. In this episode of Expanding the Possible, we're talking about machine learning. Our guest today is Gary Chang, and Gary is a data scientist on HH Angus’ digital services team. He's responsible for data analysis, visualization and applying machine learning algorithms to generate predictive models. I'll ask him to explain that in layman's terms in just a moment. Gary also has a background working in structural engineering and energy modeling before he moved into analytics, so he brings a really diverse and interesting set of skills to his work in digital services. Welcome, Gary!
As I just mentioned, your role at HH Angus involves data analysis and machine learning algorithms. Explain to us what machine learning means in the context of your work at HH Angus.
Gary
Machine learning is a very broad term. It's a term used to describe the sector of artificial learning. So, in a nutshell, it's a way of turning data into predictions. So, you're essentially telling the computer, you're trying to feed it knowledge and have it output results in human-readable language. Since we're in the AEC industry, it's turning building data, like from your controls, your set points, the temperatures of your chillers and boilers, and turning those into data points which you feed the machine learning model and it's able to output optimization strategies predictions to better optimize the systems to be more energy efficient or better meet the needs of your occupants, or whatever your problem statement is, and in a less building-conscious way, it's also a way of optimizing our workflow.
Kirsten
So how is HH Angus using machine learning? Appreciating that some of the work you do for our clients is confidential, tell us a little bit about some of the tools and the research that you're doing and using.
Gary
We are in the middle of creating a tool to help our engineering workflow. In a nutshell, the tool’s purpose is to automate some of the processes we're doing. So in this instance it's automating the space-naming process. When we're doing engineering designs, we apply CSA and ECB and ASHRAE standards based on the room types in a healthcare facility, and then that determines what kind of airflow requirements or duct work needs to be in place to meet the needs of that space. But it's a very time-intensive process. It takes a designer a full week of work hours to successfully do this manually. But by applying machine learning we're able to cut down the amount of time to do this process significantly, which also saves us man hours, but also frees up time for them to apply to other aspects to the project, making it more efficient workflow.
Kirsten
So, on the client side, you're looking at some computer vision model to detect when you have issues.
Gary
We have that computer vision pilot that we're testing out, which is, in this instance, feeding images to see when the signage in a parking lot is damaged or the lights are off, so essentially telling the user when there is something wrong with the sign. It's based on the pictures we feed this model, so it's very useful because in this context it's a commercial client and they don't necessarily have the facilities management team to constantly monitor these types of situations. So, it's a system to help this client automate the process and be able to manage all their assets in a more efficient manner. That's kind of where machine learning really shines because it's able to bridge the gap in resources when there aren't enough human resources essentially.
Kirsten
Does that work with like closed-circuit images?
Gary
The image quality really changes the performance of the model. We are using a closed-circuit camera for this model in particular; the image isn't as grainy as some of the other images you might get from closed circuit cameras. If you have a blurry picture, us as people, we wouldn't be able to tell what it is and this machine certainly wouldn't be able to tell either. So, there is like a certain threshold you need to pass to make a good model.
Kirsten
Tell me about the stack effect work you're doing with another client.
Gary
This project goes way back. It's actually what got me into HH Angus initially because I was working on the research side at the time when I was studying at Toronto Metropolitan University. And that's kind of what introduced me into the world of data analytics and machine learning. But this project in particular is looking at Stack effect. Stack effect, in very broad terms, is just the uncontrolled movement of air in very high buildings. Because in buildings, you have very large pressure differential gradients from your bottom to your top floor, and this causes a massive amount of air to just move uncontrollably from the bottom to the top or the top to the bottom. And it was particularly of concern because, when this project happened, it was during COVID, as we know now, COVID is an airborne virus. So having a massive air that's uncontrolled moving from space to space is not only a health risk, but it's also energy efficiency issue - you're potentially moving unconditioned air to conditioned spaces, so it causes higher load in your boilers and chillers or HVAC equipment. But the purpose of this project was to use the machine learning model to classify based on the differential pressures at various points of this building - when is stack effect detrimental to the performance of the building, health wise and energy wise, and right now that's kind of where the project’s at. The future scope of the project is to move forward with a recommender algorithm based on these such conditions. What can we do to fix the problem so it's like intelligently feeding back action items the facilities management team can use to mitigate the stack effect issues.
Kirsten
An early warning system.
Gary
Exactly, it's an early warning system.
Kirsten
You're involved in a research project with a Toronto University?
Gary
Yeah, so recently we've successfully gotten the grant to do a research project with Toronto Metropolitan University again. This one specifically involved looking at digital twinning of healthcare facilities. This is a very large multiyear research project and the goal of this project is to essentially be able to create a digital twin of a healthcare facility for the purposes of ongoing commissioning, energy optimization, just monitoring occupancy essentially like the name says, making a digital version of that healthcare facility. And as we know, healthcare facilities are very different than let's say commercial facilities or educational institutes. They have very different requirements. So this research is very novel in that aspect as not many researchers have been looking into this space because healthcare standards-wise, it's much more intensive. The standards are higher, so it's a very innovative piece of research to looking into this facility and seeing where we can go with that.
Kirsten
Sounds like that would be an interesting template for other healthcare facilities.
Gary
Exactly.
Kirsten
You are an Amazon Web Services cloud practitioner and a solutions architect. How do you use AWS to support machine learning?
Gary
AWS is very useful, especially when it comes to small teams because in let's say bigger tech companies like the Facebooks of this world or I guess they're called Meta now, but in companies of those scales, they have an entire division of machine learning engineers who are constantly controlling and reworking the back end. But the advantage of leveraging Cloud and AWS in this situation is we don't necessarily need our entire team of machine learning engineers to gain the same level of output because a lot of the resources are already managed by AWS. So we just need a few individuals on their keyboard and mouse typing, working away and making those same models at pretty much the same level they can do with the entire division of machine learning engineers. And that's kind of the advantage of cloud computing infrastructure when it comes to machine learning.
Kirsten
And moving along at rapid pace. Accepting that technology is advancing as you just said at a blistering pace these days, how do you think machine learning is going to change the AEC industry (and for those who don't know, that's architecture, engineering and construction), what do you see ahead for the industry in the next few years?
Gary
Machine learning is a very hot topic, especially in the research world. In the architecture and engineering construction industry. A lot of the research right now is focused on leveraging methods of applying machine learning to various problems within an industry. One common trend that a lot of even controller companies have been getting into, is looking at fault detection and energy optimization, like leveraging machine learning models to, like you were saying early warning system detection for when things could go wrong in your building. Energy optimization in the sense that you're able to take your inputs that you feed into the model and then it's able to output more energy efficient outputs, so controlling what set points you want to set your chillers and boilers at; so not only is the COP higher, but it still meets occupant standards better in your building. And moving further down the future, I could see generative AI, so chat bots, as we all know ChatGPT is really taking the world by storm, so leveraging tools like that I could really see there being uses for that when we want, say, rapid acquisition of information like for a lot of HH Angus’ digital services. The projects we're working on right now, we're looking primarily on dashboarding and showing our clients what information we can get them. But generative AI is in my opinion an extension of this. Because then we can intelligently answer questions the user might have; for instance, we work a lot in the field of occupancy. Let's say a commercial client is looking at maximizing their use of space for office space. They want to maximize how efficiently they're using it based on where the staff are. So, answering like Team A needs 45 desks, but Team B needs 30 desks. Where should these teams be allocated in this office plan? And then, based on the information you're feeding the chatbot (because we already have that dashboard and the information there), it's able to figure out and output decisions to the user. I think that's an incredible tool down the road and it would be very useful, especially because in this day and age we kind of want answers right away. So this is kind of an extension of that.
Kirsten
Do you think that property managers, our clients and other folks out there who are making these decisions have a good understanding of what machine learning can do for them? Or is it still a bit of the Wild West in terms of how they understand what the potential is?
Gary
I think it's still very wild Wild West. I feel a lot of people don't understand what machine learning really is. And when most people hear like ‘artificial intelligence’, we just think of probably pop culture references, because that's kind of where we were exposed to it early on, but that's not what really machine learning is. And it's not as smart as people think it is, but it's also more useful than some people think it is. It's a good way, like I was saying, to bridge the gap when you don't have enough resources to do the work you need to do, for instance, our internal project. It frees up our work hours for the engineering designer to work on other aspects of the design so we can bring clients a better product down the road instead of spending time on tedious products, processes that we can automate. So I think the AEC industry, as we get more exposure to machine learning, I think the opportunities will be more evident to potential clients down the road. Because there's a lot of things we could do, like energy optimization, fault detection, generative AI, predictive scheduling, the sky is the limit with machine learning, and what we can actually do with this tool because the data is there, we have a lot of data and that is the core of what machine learning needs to learn successfully.
Kirsten
And allowing engineers and designers to work on more high value work. So, let's just change direction here for a moment, Gary, and talk about your career path and how someone gets into data analytics and machine learning. How did you go from structural engineering to the role you have today and is yours a typical path, or is there a typical path to this kind of work?
Gary
I would definitely say it's not the most typical path. I kind of like stumbled upon it by accident if I'm being honest. So for my undergrad it was primarily in structural engineering, but I was exposed to building science later in the year. So that's kind of what got me into studying at Toronto Metropolitan University in the building science program. So that's primarily again more structural engineering building enclosures. But the bulk of my thesis work was looking at that stack effect project and that's what exposed me to a lot of the data analytics and machine learning methods I had to pick up on my own. I was just studying very serious classes during grad school, so just kind of picked up more information and more knowledge going along. And once I came to HH Angus, just being exposed to the projects, following the AWS courses, studying from a certification and getting hands-on experience is what kind of brought me to this point. I would say definitely it's not a very typical path, but I would say down the road it's probably going to be more of the norm eventually, in the future, as the AEC industry matures with its adoption of technology. Even in my previous graduate program, there's already a more of a focus on looking at data analytics and seeing the value that this provides because the AEC industry and engineering as a whole. What we've been doing for the last 100 years is not necessarily going to be what we're going to be doing 100 years from now, and we need to adopt to it and constantly evolve our skill set and our knowledge base to meet the demands of the future essentially.
Kirsten
Well, your whole group at HH Angus didn't exist a few years back, so it's pretty exciting to see how much how much growth and how many service offerings are coming out of this kind of work. And just for fun, let me ask you, should people be worried about machine learning? You know, it's been around for a long time. I was doing a bit of research and came across an article about an IBM computer that had beaten a human opponent at checkers back in 1962. We've come a long way since then and the capabilities of machine learning are vastly superior now. The apprehension around AI replacing human input - is that something that's well founded, or would you say that's unrealistic, or is it just inevitable?
Gary
I know exactly about the article you're talking about, and I think it's a bit of both, I think AI is a very exciting field, but we also need to be cautious of how we're applying it. It's not going to become Skynet if we don't make it into Skynet. Like the Terminator movies.
Machine learning is only as good as the data we feed it, I personally see it as a way of freeing up hours for more tedious and mundane tasks. So the designers, the engineers, can use their time for more high priority work and deliver a better version of the product down the line, instead of spending X amount of hours trying to figure out how to name each space type when a machine could just do that for you.
Machine learning is only as good as the data you're feeding it; let's say you tell the machine learning model how to play soccer, but now suddenly you're like, OK, switch sports. How do you play basketball? The AI's not going to know right away how to do it. But as you feed it more data and in a more robust way, the model is capable of, eventually down the line, predicting more generative results. So, you're able to actually expand upon the domain, but that gets into a very different side of machine learning because machine learning as a whole is only going to do what it you trained it to do. That's why the AI machine learning models aren't going to become Skynet because we're not trying to make it do bad things, we're trying to make it predict on the data we're giving it. So in the AEC industry, we're only telling it, OK, look at the chillers and boilers of this building and then it's going to be very good at looking at chillers and boilers variety of buildings. I think there's a lot of potential on that, but we also need to be cognizant of the cons of AI and the ethical considerations when we are talking about AI, just because it is a very powerful tool but with the wrong hands it could, turn into…
Kirsten
Skynet. Gary, thank you for your insights on machine learning and how it supports the work we do here at HH Angus, and a little look at what the future may hold there. Our guest today has been Gary Chang, data scientist at HH Angus’ digital services group. And to our listeners, thank you for joining us for this episode of Expanding the Possible. We'll see you next time.
Hello, I'm Kirsten Nielsen. In this episode of Expanding the Possible, we're talking about machine learning. Our guest today is Gary Chang, and Gary is a data scientist on HH Angus’ digital services team. He's responsible for data analysis, visualization and applying machine learning algorithms to generate predictive models. I'll ask him to explain that in layman's terms in just a moment. Gary also has a background working in structural engineering and energy modeling before he moved into analytics, so he brings a really diverse and interesting set of skills to his work in digital services. Welcome, Gary!
As I just mentioned, your role at HH Angus involves data analysis and machine learning algorithms. Explain to us what machine learning means in the context of your work at HH Angus.
Gary
Machine learning is a very broad term. It's a term used to describe the sector of artificial learning. So, in a nutshell, it's a way of turning data into predictions. So, you're essentially telling the computer, you're trying to feed it knowledge and have it output results in human-readable language. Since we're in the AEC industry, it's turning building data, like from your controls, your set points, the temperatures of your chillers and boilers, and turning those into data points which you feed the machine learning model and it's able to output optimization strategies predictions to better optimize the systems to be more energy efficient or better meet the needs of your occupants, or whatever your problem statement is, and in a less building-conscious way, it's also a way of optimizing our workflow.
Kirsten
So how is HH Angus using machine learning? Appreciating that some of the work you do for our clients is confidential, tell us a little bit about some of the tools and the research that you're doing and using.
Gary
We are in the middle of creating a tool to help our engineering workflow. In a nutshell, the tool’s purpose is to automate some of the processes we're doing. So in this instance it's automating the space-naming process. When we're doing engineering designs, we apply CSA and ECB and ASHRAE standards based on the room types in a healthcare facility, and then that determines what kind of airflow requirements or duct work needs to be in place to meet the needs of that space. But it's a very time-intensive process. It takes a designer a full week of work hours to successfully do this manually. But by applying machine learning we're able to cut down the amount of time to do this process significantly, which also saves us man hours, but also frees up time for them to apply to other aspects to the project, making it more efficient workflow.
Kirsten
So, on the client side, you're looking at some computer vision model to detect when you have issues.
Gary
We have that computer vision pilot that we're testing out, which is, in this instance, feeding images to see when the signage in a parking lot is damaged or the lights are off, so essentially telling the user when there is something wrong with the sign. It's based on the pictures we feed this model, so it's very useful because in this context it's a commercial client and they don't necessarily have the facilities management team to constantly monitor these types of situations. So, it's a system to help this client automate the process and be able to manage all their assets in a more efficient manner. That's kind of where machine learning really shines because it's able to bridge the gap in resources when there aren't enough human resources essentially.
Kirsten
Does that work with like closed-circuit images?
Gary
The image quality really changes the performance of the model. We are using a closed-circuit camera for this model in particular; the image isn't as grainy as some of the other images you might get from closed circuit cameras. If you have a blurry picture, us as people, we wouldn't be able to tell what it is and this machine certainly wouldn't be able to tell either. So, there is like a certain threshold you need to pass to make a good model.
Kirsten
Tell me about the stack effect work you're doing with another client.
Gary
This project goes way back. It's actually what got me into HH Angus initially because I was working on the research side at the time when I was studying at Toronto Metropolitan University. And that's kind of what introduced me into the world of data analytics and machine learning. But this project in particular is looking at Stack effect. Stack effect, in very broad terms, is just the uncontrolled movement of air in very high buildings. Because in buildings, you have very large pressure differential gradients from your bottom to your top floor, and this causes a massive amount of air to just move uncontrollably from the bottom to the top or the top to the bottom. And it was particularly of concern because, when this project happened, it was during COVID, as we know now, COVID is an airborne virus. So having a massive air that's uncontrolled moving from space to space is not only a health risk, but it's also energy efficiency issue - you're potentially moving unconditioned air to conditioned spaces, so it causes higher load in your boilers and chillers or HVAC equipment. But the purpose of this project was to use the machine learning model to classify based on the differential pressures at various points of this building - when is stack effect detrimental to the performance of the building, health wise and energy wise, and right now that's kind of where the project’s at. The future scope of the project is to move forward with a recommender algorithm based on these such conditions. What can we do to fix the problem so it's like intelligently feeding back action items the facilities management team can use to mitigate the stack effect issues.
Kirsten
An early warning system.
Gary
Exactly, it's an early warning system.
Kirsten
You're involved in a research project with a Toronto University?
Gary
Yeah, so recently we've successfully gotten the grant to do a research project with Toronto Metropolitan University again. This one specifically involved looking at digital twinning of healthcare facilities. This is a very large multiyear research project and the goal of this project is to essentially be able to create a digital twin of a healthcare facility for the purposes of ongoing commissioning, energy optimization, just monitoring occupancy essentially like the name says, making a digital version of that healthcare facility. And as we know, healthcare facilities are very different than let's say commercial facilities or educational institutes. They have very different requirements. So this research is very novel in that aspect as not many researchers have been looking into this space because healthcare standards-wise, it's much more intensive. The standards are higher, so it's a very innovative piece of research to looking into this facility and seeing where we can go with that.
Kirsten
Sounds like that would be an interesting template for other healthcare facilities.
Gary
Exactly.
Kirsten
You are an Amazon Web Services cloud practitioner and a solutions architect. How do you use AWS to support machine learning?
Gary
AWS is very useful, especially when it comes to small teams because in let's say bigger tech companies like the Facebooks of this world or I guess they're called Meta now, but in companies of those scales, they have an entire division of machine learning engineers who are constantly controlling and reworking the back end. But the advantage of leveraging Cloud and AWS in this situation is we don't necessarily need our entire team of machine learning engineers to gain the same level of output because a lot of the resources are already managed by AWS. So we just need a few individuals on their keyboard and mouse typing, working away and making those same models at pretty much the same level they can do with the entire division of machine learning engineers. And that's kind of the advantage of cloud computing infrastructure when it comes to machine learning.
Kirsten
And moving along at rapid pace. Accepting that technology is advancing as you just said at a blistering pace these days, how do you think machine learning is going to change the AEC industry (and for those who don't know, that's architecture, engineering and construction), what do you see ahead for the industry in the next few years?
Gary
Machine learning is a very hot topic, especially in the research world. In the architecture and engineering construction industry. A lot of the research right now is focused on leveraging methods of applying machine learning to various problems within an industry. One common trend that a lot of even controller companies have been getting into, is looking at fault detection and energy optimization, like leveraging machine learning models to, like you were saying early warning system detection for when things could go wrong in your building. Energy optimization in the sense that you're able to take your inputs that you feed into the model and then it's able to output more energy efficient outputs, so controlling what set points you want to set your chillers and boilers at; so not only is the COP higher, but it still meets occupant standards better in your building. And moving further down the future, I could see generative AI, so chat bots, as we all know ChatGPT is really taking the world by storm, so leveraging tools like that I could really see there being uses for that when we want, say, rapid acquisition of information like for a lot of HH Angus’ digital services. The projects we're working on right now, we're looking primarily on dashboarding and showing our clients what information we can get them. But generative AI is in my opinion an extension of this. Because then we can intelligently answer questions the user might have; for instance, we work a lot in the field of occupancy. Let's say a commercial client is looking at maximizing their use of space for office space. They want to maximize how efficiently they're using it based on where the staff are. So, answering like Team A needs 45 desks, but Team B needs 30 desks. Where should these teams be allocated in this office plan? And then, based on the information you're feeding the chatbot (because we already have that dashboard and the information there), it's able to figure out and output decisions to the user. I think that's an incredible tool down the road and it would be very useful, especially because in this day and age we kind of want answers right away. So this is kind of an extension of that.
Kirsten
Do you think that property managers, our clients and other folks out there who are making these decisions have a good understanding of what machine learning can do for them? Or is it still a bit of the Wild West in terms of how they understand what the potential is?
Gary
I think it's still very wild Wild West. I feel a lot of people don't understand what machine learning really is. And when most people hear like ‘artificial intelligence’, we just think of probably pop culture references, because that's kind of where we were exposed to it early on, but that's not what really machine learning is. And it's not as smart as people think it is, but it's also more useful than some people think it is. It's a good way, like I was saying, to bridge the gap when you don't have enough resources to do the work you need to do, for instance, our internal project. It frees up our work hours for the engineering designer to work on other aspects of the design so we can bring clients a better product down the road instead of spending time on tedious products, processes that we can automate. So I think the AEC industry, as we get more exposure to machine learning, I think the opportunities will be more evident to potential clients down the road. Because there's a lot of things we could do, like energy optimization, fault detection, generative AI, predictive scheduling, the sky is the limit with machine learning, and what we can actually do with this tool because the data is there, we have a lot of data and that is the core of what machine learning needs to learn successfully.
Kirsten
And allowing engineers and designers to work on more high value work. So, let's just change direction here for a moment, Gary, and talk about your career path and how someone gets into data analytics and machine learning. How did you go from structural engineering to the role you have today and is yours a typical path, or is there a typical path to this kind of work?
Gary
I would definitely say it's not the most typical path. I kind of like stumbled upon it by accident if I'm being honest. So for my undergrad it was primarily in structural engineering, but I was exposed to building science later in the year. So that's kind of what got me into studying at Toronto Metropolitan University in the building science program. So that's primarily again more structural engineering building enclosures. But the bulk of my thesis work was looking at that stack effect project and that's what exposed me to a lot of the data analytics and machine learning methods I had to pick up on my own. I was just studying very serious classes during grad school, so just kind of picked up more information and more knowledge going along. And once I came to HH Angus, just being exposed to the projects, following the AWS courses, studying from a certification and getting hands-on experience is what kind of brought me to this point. I would say definitely it's not a very typical path, but I would say down the road it's probably going to be more of the norm eventually, in the future, as the AEC industry matures with its adoption of technology. Even in my previous graduate program, there's already a more of a focus on looking at data analytics and seeing the value that this provides because the AEC industry and engineering as a whole. What we've been doing for the last 100 years is not necessarily going to be what we're going to be doing 100 years from now, and we need to adopt to it and constantly evolve our skill set and our knowledge base to meet the demands of the future essentially.
Kirsten
Well, your whole group at HH Angus didn't exist a few years back, so it's pretty exciting to see how much how much growth and how many service offerings are coming out of this kind of work. And just for fun, let me ask you, should people be worried about machine learning? You know, it's been around for a long time. I was doing a bit of research and came across an article about an IBM computer that had beaten a human opponent at checkers back in 1962. We've come a long way since then and the capabilities of machine learning are vastly superior now. The apprehension around AI replacing human input - is that something that's well founded, or would you say that's unrealistic, or is it just inevitable?
Gary
I know exactly about the article you're talking about, and I think it's a bit of both, I think AI is a very exciting field, but we also need to be cautious of how we're applying it. It's not going to become Skynet if we don't make it into Skynet. Like the Terminator movies.
Machine learning is only as good as the data we feed it, I personally see it as a way of freeing up hours for more tedious and mundane tasks. So the designers, the engineers, can use their time for more high priority work and deliver a better version of the product down the line, instead of spending X amount of hours trying to figure out how to name each space type when a machine could just do that for you.
Machine learning is only as good as the data you're feeding it; let's say you tell the machine learning model how to play soccer, but now suddenly you're like, OK, switch sports. How do you play basketball? The AI's not going to know right away how to do it. But as you feed it more data and in a more robust way, the model is capable of, eventually down the line, predicting more generative results. So, you're able to actually expand upon the domain, but that gets into a very different side of machine learning because machine learning as a whole is only going to do what it you trained it to do. That's why the AI machine learning models aren't going to become Skynet because we're not trying to make it do bad things, we're trying to make it predict on the data we're giving it. So in the AEC industry, we're only telling it, OK, look at the chillers and boilers of this building and then it's going to be very good at looking at chillers and boilers variety of buildings. I think there's a lot of potential on that, but we also need to be cognizant of the cons of AI and the ethical considerations when we are talking about AI, just because it is a very powerful tool but with the wrong hands it could, turn into…
Kirsten
Skynet. Gary, thank you for your insights on machine learning and how it supports the work we do here at HH Angus, and a little look at what the future may hold there. Our guest today has been Gary Chang, data scientist at HH Angus’ digital services group. And to our listeners, thank you for joining us for this episode of Expanding the Possible. We'll see you next time.