Resources for Computer Vision Teams:

LinkedIn – Connect with Heather.
Computer Vision Insights Newsletter – A biweekly newsletter to help bring the latest machine learning and computer vision research to applications in people and planetary health.
Computer Vision Strategy Session – Not sure how to advance your computer vision project? Get unstuck with a clear set of next steps. Schedule a 1 hour strategy session now to advance your project.

Transcript:

[INTRODUCTION]

[00:00:04] HC: Welcome to Impact AI. Brought to you by Pixel Scientia Labs. I’m your host, Heather Couture. On this podcast, I interview innovators and entrepreneurs about building a mission-driven, machine-learning-powered company. If you like what you hear, please subscribe to my newsletter to be notified about new episodes. Plus, follow the latest research and computer vision for people and planetary health. You can sign up at pixelscientia.com/newsletter.

[INTERVIEW]

[00:00:34] HC: Today I’m joined by guest, Mathieu Bauchy, CTO and Co-Founder of Concrete.ai, to talk about optimizing concrete mixes. Mathieu, welcome to the show.

[00:00:44] MB: Yeah. Thanks for having me.

[00:00:45] HC: Mathieu, could you share a bit about your background and how that led you to create Concrete.ai?

[00:00:50] MB: Yeah. My background is in computational physics. I got a PhD from University in Paris. And ever since that, I basically never left school. I became a professor in civil engineering at UCLA and my group has conducted some research on concrete and construction materials for more than 10 years already. And so, we really realized that concrete production is inefficient. Basically, we realized that it’s more expensive that it needs to be. It’s also more CO2 intensive that it needs to be.

And so, as part of, this we conducted some earlier research at the university to explore how AI can really address those inefficiencies. And it seemed really like a promising approach. And I love teaching. I love academic research. But we started to wonder if we really could do something bigger at a larger scale. We realized that the only way that we would be able to really have a large impact, more like a short-term impact, would be to spin our research out of the university and really start exploring the startup world. This is when I co-founded Concrete.ai with my colleague, Gaurav Sant. [00:01:58] HC: What does Concrete.ai do? And why is it important for reducing emissions?

[00:02:02] MB: Yeah. Concrete.ai is a software as a service for concrete producers. We call this software Concrete Copilot. And, basically, it’s using generative AI to reduce both the cost but also the carbon footprint of concrete. And so, that’s a pretty important thing because concrete is the most produced material in the world. It’s actually the second most used material. Only second to water. The scale that we are talking about is really, really large.

And because of this very large scale, because we mix so much concrete every year, concrete is responsible for about 8% of the total CO2 emissions in the world. And just to give you some context, that’s about three times more emissions than the entire aviation industry. Concrete is kind of like hidden. It’s not really in the spotlight. But it’s really a big contributor to climate change.

And because of this scale, every small reduction, for every percentage reduction that we can make in terms of the carbon footprint of concrete, it can really have a big impact. And the CO2 that we emit due to concrete and construction in general, it’s really like a depth. Every ton of CO2 that we don’t emit now is a ton of CO2 that we won’t have to capture or immobilize later.

And so, this is really what we’re trying to do. This software, Concrete Copilots, can really reduce the carbon footprint of concrete. And the way it does that is by changing the recipe of concrete formulation. And, concrete, it’s really like making a cake in the sense that you have some ingredients. You mix them together. Probably, you mix some cement, some stones, some sand, some water. And depending on the quantities that you use for the ingredients and the types of ingredient that you use, it’s going to impact the performance of concrete. But it’s also going to impact its cost and its carbon intensity.

What we are trying to do as part of Concrete.ai is to basically solve an optimization problem in the sense that, for each concrete formulation – for different application in a building, you have different types of concrete that are used. And each type of concrete will have certain types of performance targets. But, also, some constraints, for example, based on the locally available materials.

And our goal is to give some tools to producers to quickly find the most optimal concrete recipe for every single job that’s going to minimize both the cost for them. But, also, the carbon footprint of concrete. While always making sure that their concrete is always going to meet or exceed all the performance targets.

[00:04:38] HC: And what role does machine learning, in particular, generative AI, like you mentioned, what role does it play in this technology?

[00:04:45] MB: Yeah. The issue is that it’s very hard to predict concrete performance. If you have a given recipe, some ingredients, like figuring out how this concrete is going to perform, there is really no good model or no simulation that allows you to do that. And to some extent, it’s kind of crazy to think that we can really very accurately simulate new exotic high-tech materials. We can also simulate drugs, proteins, things like that. But we really can’t do that for the most produced material in the world, which is concrete. This is where machine learning can help.

We have different types of models. Some of our models are I would say more traditional. Those are some forward-prediction models. For example, one of the models that we have is to be able to predict the performance of a given concrete. The idea is that if you know the ingredients of a concrete, the quantities of those ingredients, then we are able to predict how this concrete is going to perform on the long-term.

And so, basically those models, we can use them as surrogate to avoid to have to systematically batch and test this concrete in the lab. We also have some models to predict the manufacturability of concrete. And concrete is really like a magic material in the sense that when you start to mix your material with water, initially, it behaves like a liquid slurry. You can make it into any shape you want. And really on its own, it gradually hardens and turns into stone.

And so, when it comes to concrete, there are really two things that matter. The one thing is the long-term performance. How strong is it going to be on the long term? What also matters is to make sure it is liquid enough during the early age of the concrete life so that you can use it to make construction. We also want to make sure that the concrete doesn’t turn into a stone in the concrete truck while it is delivered. We always have to make sure that there’s always a balance between the long-term performance. But, also, making sure it has the right manufacturability properties right after it’s made. We have some models to also predict those manufacturing properties.

And then the last model that we have, which is more in line with generative AI. We call it like a backward model. That’s the model that allows us to do some inverse design. Where now the goal is different. Rather than predicting performance based on a given concrete recipe, the goal of those models is really to do the opposite. That is now we start with the targets, the performance targets that we need, the constraints, and based on that, the model is going to prescribe the most optimal concrete formulation. When I say most optimal, it means lower cost and lower carbon emissions. This is really where we use generative AI to generate the most economical and the less carbon recipe for resets of performance targets. It’s a pretty challenging task in the sense that, when we make concrete, there is really millions of possible recipes.

And so, in the case of a typical generative AI application, the goal of the application is typically to fool humans in the sense that if you take a common image, or video generator, or ChatGPT, for example, the goal is really to generate a photo, or a video, or a text that humans won’t be able to distinguish from reality. But in our case, our goal is not to generate just any concrete. Our goal is to generate like the most single, most optimal one concrete formulation. And so, this is really like a needle in the haystack problem, which is, in my mind, a little bit more challenging than typical generative AI applications.

[00:08:24] HC: Is this like a generative adversarial network where you have two different objectives competing? One, to generate the optimal mixture, and another to better tell whether this is a good mixture?

[00:08:36] MB: Yeah. This is basically kind of how it works in the sense that – we have a generator that is going to suggest some formulations. And then we have our other models that are going to tell us, is this formulation going to perform in terms of the strength that it’s going to have and in terms of any other requirement that the producers, or the contractors, or the owners may have? And, also, to make sure that it’s actually going to be manufacturable.

Yeah. It works in spirit like a typical adversarial model. But, really, in the sense that, in this case, the objective is not to try to fool humans to generate something that would be indistinguishable from reality. But in this case, the objective is to have the concrete that would be the most optimized in the sense of lowest cost and lowest carbon emissions.

[00:09:26] HC: One of the well-known challenges with machine learning is extrapolation. Models can perform well on data that is similar to what they’ve seen in training but are unpredictable outside that regime. But I understand from looking over your website that extrapolating to new concrete formulations is an important feature of what you’re doing. How do you develop machine learning models that are reliably able to extrapolate?

[00:09:50] MB: Yes. That’s an important challenge for us. Because our goal is really to generate like brand-new concrete recipes. Those recipes can be very different from the existing ones and, typically, those are going to be recipes that producers have not considered on their own. It’s typically going to be some formulation that really don’t exist yet and really that are not going to be in the training set.

Again, for us, that’s a pretty important difference as compared to other types of generative AI applications. I mean, most of existing application, usually they are based on really regurgitating the training set or really interpolating between existing data points but not really extrapolating. For us, it’s very important to have some models that can extrapolate outside of the range of data they are trained on because this is where the opportunity is. If we just give producers some new recipes that are just in between the one they already using, then we are unlikely to really significantly reduce the carbon footprint of concrete, for example.

This is really what we have been spending a lot of effort to make sure our model can really reliably extrapolate outside the range of the training set. And that’s obviously, as you said, a challenge. Because machine learning models are known to easily hallucinate when extrapolating. We spend a lot of effort on ensuring that our model can offer robust extrapolations. Actually, we got a patent granted on that very aspect of our model. And our strategy is really based on two key ideas. The first thing is that, unlike some other fields of machine learning, we are working with real-life materials for which we have some chemical and some physical knowledge. And so, one aspect of what we do is to really merge data and knowledge. That’s to say that our model are not trained simply based on data. But they are also trained based on the knowledge that we have.

We don’t have enough knowledge to directly predict concrete performance simply based on the knowledge that we have, simply based on the underlying physics or the underlying chemistry that is governing concrete. But we still have enough knowledge that we can use it to augment the assets that we have. And so, that’s really what we do. We found ways to really seamlessly merge within the same models, large assets and knowledge, and merging those two together. Like, really help in making sure that those models can extrapolate. It’s much easier to extrapolate when you have some underlying physical lows. Because they can make sure that your models is never going to break the laws of physics or the laws of chemistry, which is something that can easily happen when you just extrapolate.

And then the second aspect that we have is that we realize that the common best practices when you train a machine learning model, which is typically to do some cross-validation with some random split where we use a portion of the data set to train the model and another portion of the data set to validate the model. Those types of approaches, they are actually really testing the ability of the model to interpolate much more than to extrapolate.

We developed some new approach that are equivalent in spirit to cross-validation but that are really designed to test and validate how well the model extrapolates rather than how well it interpolates. And so, by combining those two things, these new types of cross-validation and the combination of physics and chemistry, this is how we can ensure that our models are able to extrapolate fairly robustly outside the range of the training data.

[00:13:31] HC: Another characteristic I read about when I was learning about what you’re doing at Concrete.ai is that you’re also estimating uncertainty. This is certainly possible with deep learning models. But most of us aren’t really concerned with it. For your application, why is estimating uncertainty important? And how do you approach this?

[00:13:48] MB: Yes. I mean, uncertainty for us is extremely important, especially since we are trying to extrapolate. And when we extrapolate, we always need to be aware of the underlying uncertainty. We’d be aware of how robust and how reliable this extrapolation is. I mean, more generally, uncertainty for us, it’s very critical in everything that we do. Because we are dealing with concrete. Concrete is used in construction and uncertainty is really used at every step of construction.

For example, typical building codes are based on probabilities of failure uncertainties. And, for example, typically, they are designed such that, for a given structure, there’s going to be less than 1% probability of failure over the lifetime of the structure. And it’s really the same for concrete. We always need to make sure that not only we predict that the performance will be higher than the target. But that we have a large confidence in doing so.

Because when we deal with construction, concrete is used in many infrastructure, like bridges or high-rise buildings. If there is any kind of failure, then lives can be at stake. Dealing with uncertainty is very important for us, and so all our models are uncertainty-aware. They are always aware of their own confidence. And so, whenever we prescribe a new concrete recipe to the producers, we always make sure that there is more than a 99.9% probability that this concrete is going to indeed meet the performance target or exceed the performance target.

And so, the only way to do that is to have some models that not only make some prediction but systematically also make an estimation of the confidence interval attached to every prediction. And so, the way we achieve that in practice from a more technical viewpoint is that we rely on assembling approach. Where, for everything we do, we never have a single model. We always have a collection of models. And it’s like every model is casting a vote. Then we take into account all the votes from all the models. If the votes are all unanimous, then basically that means that the uncertainty is going to be pretty low, and on the other hand, when there starts to be some disagreement between different models, then that means that the uncertainty is increasing.

Based on that, based on taking into account the votes of all the models, then we’re able to calculate the confidence interval of every prediction that we are making. And, basically, the idea is the further we extrapolate far from the existing data that we have in the training set, the more the uncertainty increases. And the more we need to be conservative in our prediction to make sure we always have this 99.9 probability that the performance of the concrete will be higher than the target. And then once we have those two information, once we have the knowledge of what the model predicts, but also the uncertainty of the prediction, then our models are able to find the best balance between exploration and exploit in the sense that we exploit both the knowledge of what our model confidently predict. And we also explore potential opportunities offered by new formulation that maybe have never been made in a lab that are more high-risk and high-reward, I would say.

And so, by finding the optimal balance between exploiting and exploring, this is how we are able to find some new concrete formulation that at the same time different from existing one with a significantly lower carbon footprint, for example, that are also a domain for which the uncertainty is not too high, so that we can always make sure that the performance is going to be met by the concrete formulations.

[00:17:42] HC: How do you validate your models? You mentioned cross-validation as one way to test extrapolation. Are there ways that you can validate your confidence intervals and further tests the ability of your models to extrapolate?

[00:17:56] MB: Yeah. I mean, the first thing is that our software is really built like a copilot. That’s why we call it Concrete Copilot. And whenever we make a prediction, it’s really made for humans to interact with rather than to replace them. And so, in practice, when a producer is optimizing a concrete and trying to reduce for a given application to reduce the cost of a given concrete, they can really interact with the prediction from the AI. And they can use their experience and their intuition to add/remove constraints, things like that, and to really see the results in real-time and interact seamlessly with the AI.

For us, that’s really – maybe to some extent, the most important validation is that there is always an immediate human validation in what the AI is outputting. We do have some internal like – as I mentioned, some kind of refined cross-validation to make sure that the model can extrapolate. But really, for us, what is the real validation, the one that matters is to actually make and test the concrete on the fields. And so, that’s one thing we have to systematically do.

We go to the concrete plant. We batch the new recipe that our software is recommending. We put that in the concrete truck. And then we take some few samples of concrete from the truck. We make some small cylinders out of it. And then we wait. We wait typically for 28 days. That’s typically the time that it takes to test a given concrete. That’s the time that it takes for the concrete to turn into a stone. And then we test the concrete. We check if it’s strong enough for the targeted application. And once we ensure that, then, basically, the concrete can be used on the field for constructions.

This validation process has to be done for every new concrete formulation irrespectively of whether this concrete was created by a human, just like it’s traditionally done, or by an AI. This type of validation always needs to be performed for any new concrete anyway. That’s what we use as the validation of the AI. Basically, we use the existing certification method for concrete to validate our models.

[00:20:03] HC: Why is now the right time to build this technology?

[00:20:06] MB: Yeah. I mean, it’s really pretty critical timing in the concrete world now because concrete producers are under a lot of pressure for three different reason. First, the supply chains in the world of concretes, they keep changing. There is some materials that used to be available that are not anymore. For example, we’re starting to really run out of sand in the world. There is some big shortage of sand.

On the other hand, there is also some new materials, new types of cements, new types of green cement that are being introduced in the concrete world. There is also the fact that the price of the material keeps changing. All of those are kind of degrees of freedom or parameters that really tend to change from month to month. For producers, it becomes very challenging because what used to be optimal yesterday or what is optimal today might not be optimal tomorrow. They are under some pressure to constantly change and adapt, to changing their supply and always adapt to that too, and make sure that their concrete is always going to meet all the right performance targets, and to do this at the lowest possible carbon emissions. That’s the first challenge.

The second challenge is really the fact that the cost of the materials that are used in concrete has really increased a lot over the past years. There’s some shortage. There is some disruption in the supply of those materials. Basically, all of those costs are really increasing a lot year to year. For producers, their margin is really shrinking. And they are already in the single digits. It’s a very low-margin business. It’s very challenging for concrete producers to still manage to make some profits.

Then the last thing is that concrete producers are also under a lot of pressure because of the fact that they are such a big contributor to climate change. Especially, in places of the world like Europe, for example, you start to have some carbon taxes where the producers have to pay for every ton of CO2 that they release in the atmosphere. So, there is also the public pressure that people don’t want those heavy industries, like the concrete industry, or the steel industry, or the glass industries. Those are industries that really contribute a lot to climate change. And so, they are facing a lot of pressure to really reduce this impact.

When we think about all of those challenges, they kind of conflict with each other. It’s hard to resolve all of them together. It’s like farming. We complain when they use a lot of water on the field. But we’re also happy to get the food that they produce at the lowest possible cost. And it’s the same for concrete. We blame producers for the CO2 they emit. But as a society, we always want new and improved infrastructures. And we don’t want to spend too much tax dollar on it. We want this to be as economical as possible.

This is especially true for developing countries. It’s unfair to really ask those countries to stop building new infrastructures or to improve the one they have while developed countries already have this infrastructure and they have already emitted some large amounts of CO2 emissions during their history, right? We think that it’s the right time based on all of those change for the concrete industry to benefit from what AI can offer to really avoid any kind of waste during the production of concrete.

And the idea is that if we really adopt those new technologies, then we can continue to improve our quality of life. We can continue to build those new infrastructures. We can continue to repair the existing ones and really make sure that we always have this new infrastructure that we need as a society. But we can also do so while making sure that we are minimizing the impact of those new infrastructures on climate change.

[00:23:51] HC: How do you measure the impacts that your technology is having? [00:23:55] MB: Yeah. It has really like a pretty short-term impact. And that’s what is also pretty exciting, especially coming from the academic world where sometimes what we do can only have an impact after a pretty long time. Or it’s pretty long typically to translate from the academic work to the real world. And, for us, typically, the impact happens like about 28 days after we recommend a new recipe. That’s the time that it takes to test this concrete and making sure it’s performed as expected.

Typically, we recommend a new formulation 28 days after producers can start. Making this new recipe and replacing the old recipe by the new recipe. And so, in terms of impact, the metric that we care about is really how many dollars have been saved for the producer. For every concrete truck that leave the plants, how much more economical this concrete was as compared to the one that they were using before?

Ultimately, the concrete that for us is really the most important is how much CO2 did we prevent from entering the atmosphere? And, typically, our objective is to not only be able to double a profit margin of producers by allowing them to save money on the materials that they use, but ultimately, we’ve been able to help them reduce carbon emission by about a third. That’s what I think is the most exciting, is that, again, we are just using AI to refine the recipe that they are using. It’s not like we are changing the way concrete is made. It’s still made in the same plant. It’s still made using the same material. We are just changing the recipe. And just that is able to save about a third of the emission of concrete.

I think that’s a good example of what can be achieved when you arrive to an industry that has not been really touched by AI yet. And AI is able to find those inefficiencies and really address them. And so, there is a pretty low-hanging fruit in terms of the opportunity to really reduce the CO2 that we are emitting now due to concrete productions.

[00:25:54] HC: Is there any advice you could offer to other leaders of AI-powered startups?

[00:25:59] MB: Yeah. I mean, I’m sure everybody has this in mind. But what I would really say is I would really invite everybody to really think about how can we use AI for really the good of humanity? And AI can be scary and in some cases, it can have a pretty negative image in the sense that there are some concerns, legitimate concerns that it may take away people’s jobs. And so, I think we really need to think collectively. How do we use AI, for example, to empower workers? Not to replace them.

In terms of the carbon footprint, AI typically also have its own carbon footprint. For example, we need some concrete to build data centers that are used to power AI. And those data centers are pretty energy-intensive to operate. AI also comes with its own carbon footprint and to some extent, also contributes to climate change. In my mind, we should think about how we use AI to really solve climate change and not further contribute to it. And so, in that respect, for example, one of the first client that we had is a data center provider. And so, it was fun to think that, in this case, we used AI to build the concrete that was used on the very data center that then was used to create some AI application. It’s kind of an interesting loop there.

But in my mind, AI clearly poses some challenges both to humans in general, society, and to the climate to some extent. But it can also really be used for good. To really have a positive impact on the planet on society. And so, I think it’s good that we always keep that in mind. To make sure that we use AI for good and not to impact climate change more than it already is.

[00:27:42] HC: And, finally, where do you see the impact of Concrete.ai in three to five years?

[00:27:47] MB: Yeah. For us, we are really trying at this stage to expand internationally beyond North America. Ultimately, I mean a very small fraction of the concrete that is made in the world is actually made in North America. There is really like an opportunity in developing countries. India, for example, is really making a lot of concrete. There is this number that in my mind is always very impressive, is that, in China, every two years, they make as much concrete as we did in the US for the entire 20th century. The scale of concrete production in India or China is very, very large. And so, ultimately, that’s where we think that we can have an impact.

Ultimately, we have a mission as our startup. And this mission is to expand as much as possible to – ultimately, our goal is to prevent 500 megatons of CO2 from entering the atmosphere annually. And we think we can achieve that. And so, that’s an ambitious objective. But, again, we think that that’s what can be done at a very – with a short-term objective. Simply by, again, using AI to address those inefficiencies without changing fundamentally the way concrete is made. This number that we have in mind this 500 megatons of CO2. That would be the equivalent of removing the annual emissions of a country like both France and Argentina combined. That would be really like a significant impact. We really think that due to the fact that we make so much concrete, there is really – even if we can have any kind of iterative improvement in the way concrete is made, there’s really a big opportunity to have a positive impact on the planet.

[00:29:29] HC: This has been great, Mathieu. I appreciate your insights today. I think this will be valuable to many listeners. Where can people find out more about you online?

[00:29:36] MB: Yeah. I mean, if people want to get in touch, I’m actually mostly active on YouTube. I have a channel there. If people are interested, I actually have – as part of my life as a teacher, I have a course on machine learning for engineers. If anybody actually wants to get started, I think that that could be a useful resource. I’m always happy to connect on LinkedIn. And if people want to know more about Concrete.ai, then they can go on our website, which is concrete.ai, and learn more about what we do.

[00:30:03] HC: Perfect. I’ll link to all of those in the show notes. Thanks for joining me today.

[00:30:06] MB: Thank you.

[00:30:08] HC: All right, everyone. Thanks for listening. I’m Heather Couture. And I hope you join me again next time for Impact AI.

[OUTRO]

[00:30:17] HC: Thank you for listening to Impact AI. If you enjoyed this episode, please subscribe and share with a friend. And if you’d like to learn more about computer vision applications for people in planetary health, you can sign up for my newsletter at pixelscientia.com/newsletter.

[END]