How Paul Leu is Reinventing Glass with Machine Learning: Experiment Exchange Episode 7

Nick Payton
Bayesian Optimization, Machine Learning, Materials Science, Simulations & Backtests

How do you design a better glass?

In this week’s episode of Experiment Exchange, “How Paul Leu is Reinventing Glass with Advanced Machine Learning,” join Michael McCourt as he interviews Paul Leu, an Associate Professor of Industrial Engineering at the University of Pittsburgh. His group, the Laboratory for Advanced Materials at Pitt, works to design and fabricate new materials with unique properties. They discuss the collaboration between Pitt and SigOpt to design new additive manufacturing strategies to improve the performance of glass with a focus on reducing reflection, increasing transparency, and reducing the ability of water and oil to attach itself to the glass.

Below is a transcript from the interview. Subscribe to listen to the full podcast episode, or watch the video interview.

Q: Tell us about yourself.

I’m an Associate Professor at the University of Pittsburgh. I joined Pitt back in 2010. I did my Ph.D. at Stanford in Mechanical Engineering and then after that did a short postdoc at UC Berkeley before joining the Industrial Engineering Department. 

Q: How are you liking Pittsburgh? 

I really love the city. It’s a great hub for technology and health research. We have both Carnegie Mellon University, which is very strong in data science and machine learning, and the University of Pittsburgh, which is very strong in biomedical research. Pittsburgh is consistently ranked very high in livability surveys and it has most of the things that you would want out of a city—sports, arts, entertainment, and parks. It’s also very affordable—you can buy a house here for $200-300K, which is something that, in a lot of cities as a professor, you get priced out of. 

Q: What’s happening in your interdisciplinary department these days?

What’s interesting is that my background has been more focused on mechanical engineering. I got my postdoc in electrical engineering, and I was co-advised by a professor in Material Science when I was a Ph.D. student. Right now at Pitt, my main department is Industrial Engineering, but I do have joint appointments in Mechanical Engineering and Chemical Engineering.

What I think is really interesting is that a lot of research now is very interdisciplinary. A lot of these traditional boundaries are breaking down and I think that’s really where you get a lot of interesting new work being done. That’s a general trend across universities as well. I see a lot of research groups where the Ph.D. students inside that research group span a wide variety of backgrounds and disciplines and are getting their PhDs in different areas. 

Q: Tell us more about your research these days.

A lot of my research has been in functional materials and nanomaterials, so developing novel materials and trying to create new functionalities from these materials. 

Nanomaterials are very, very small. Nano means ten to the minus nine. To give you an idea of the length scale of that, if you have a meter and you go a thousand times smaller, you have millimeters, which are the size of wires and charging cables. You go another thousand times smaller, those are microns: red blood cells are just a few microns in diameter. Another thousand times smaller than that would be nanometers: individual atoms are about a 10th of a nanometer in diameter.

A good instructive example would be if you take an apple and you enlarge that apple to the size of the earth, then the atoms in that apple are now about the size of that original apple. So apple is to earth, as atom is to apple. So that just kind of gives you a rough idea of how incredibly small atoms are. Those are the types of length scales we’re working at. Because of how small things are, you have a lot of interesting new properties and functionalities that you can take advantage of.

Q: What are people building at that scale?

What some researchers are doing at that scale is building transistors—trying to make the little switches that go into your computer processing unit smaller, packing more devices into a smaller area, having more memory capability, or trying to create new sensors that can sense individual molecules or molecules at very small concentrations, things like that. 

There are also a lot of interesting possibilities with visible light. The light that humans see is around 400 to 700 nanometers in wavelength. A lot of these nanomaterials are smaller than that or on the same length scale as that, and you can design for a lot of interesting ways of interacting with light.

Q: What research have you been doing specifically to think about these questions around light and how light can be manipulated or interacted with? 

One of the big areas that we’ve worked on is solar. For example, can we create new structures that convert sunlight into electricity with higher conversion efficiencies? Can we create new materials or new solar cells that use less material so they could potentially be less expensive overall?

There are other types of light sensors that capture light within specific parts of the spectrum. You may want to capture, for example, ultraviolet light or infrared light and then filter out other parts of the spectrum. Those are all types of applications that we’ve worked on.

One of the big applications, which is what I talked about at the SigOpt Summit, is also related to light and nanostructures—creating new types of glass which allow light to pass through. We want to create glass that is very anti-reflective across a wide variety of wavelengths, as well as a wide variety of incidence angles. We’ve also looked at other types of functionalities that we can implement into that glass, things that could be beneficial depending on the application, like: self-cleaning, anti-soiling, stain resistance, or anti-fogging. 

Q: In this search for designing glass that has these properties, how does SigOpt play a role? 

SigOpt has been extremely useful in our materials work. One of the big challenges for most experimentalists is that they rely on physical intuition a lot. They may be working with 10-20 various possible parameters that they can adjust or tune, but in the end, they’ll probably fix most of those parameters and systematically vary one or two of them. A lot of experimentalists end up doing experiments in a very limited part of the entire design parameter space. SigOpt has enabled us to search the very high dimensional design or design parameter space more thoroughly, in a very fast and efficient manner, to look for material designs that are optimized for anti-reflection, self-cleaning, and low haze, for example.

Q: Can you tell us more about how your iterative process is going, in particular on the computational side of things?

One of the things that we’ve worked on is electrodynamic simulations, doing simulations to see what types of structures provide for the best anti-reflection properties. This could be across the visible spectrum (which is important for things like displays), or across the solar spectrum (which is more important for solar panels).

We’ve also looked at minimizing reflection, having as much anti-reflection as possible for high incidence angles. This is important because, for things like lighting or even for displays, you don’t want the light to be so directional. For something like a TV screen, you generally want to have a profile that is more uniform in all directions. For solar, this is also important because most solar installations are fixed at some angle on your roof. The sun moves through the sky during the day, and scattered light hits your solar panel from all different directions. Being able to capture that light efficiently when the sun’s not necessarily directly normal to the solar panel is helpful for improving its overall efficiency.

We’ve done a lot of work together with some of our tools, like MATLAB and Mathematica, and are trying to integrate them with SigOpt. Being able to work with SigOpt’s multiobjective optimization has also been very helpful. In a lot of materials design work, there are tradeoffs between the various objectives that you’re trying to optimize. Understanding where those tradeoffs are and what the limits are in terms of a Pareto frontier have been useful. A lot of SigOpt’s visualization tools have been also very helpful in understanding our data, understanding how the Bayesian optimization process is helping us move closer to a better design.

Q: What else are you working on this year?

You may have heard that 2022 is the Year of Glass, as designated by the United Nations. A lot of universities, museums, and companies are promoting all the different types of glass and their different applications. Wherever you are, be on the lookout for that, there are probably opportunities to visit your local museum and see different glass displays. Glass is ubiquitous and if you’re watching this on your computer, you’re probably looking through a display that has glass on the front of it. Your internet signal is being transmitted through fiber optic cable, which is glass. Glass has really transformed the way we live.

The other thing that has been big is this new National Science Foundation Center that both the University of Pittsburgh and Case Western Reserve University have started up—MDS-Rely. It’s an IUCRC, an Industry University Cooperative Research Center. We’ve been cooperating with various companies that have joined the center government labs, and are basically in the center trying to bring data science methods towards materials manufacturing, understanding how materials properties and functionalities change over time, understanding what’s causing these properties and functionalities to degrade, and trying to see if there are ways in which we can improve the materials so that they last longer.

For example, a solar panel commercially right now is expected to last about 25 years. If you put it up on your roof, after about ten years it’s not going to be able to convert sunlight as efficiently as when you initially installed it, perhaps at around 90% of its power conversion efficiency—and then around 80% after 20 years. After about 25 years, it may be time to get a new solar panel. 

We’re using these new types of data science and machine learning methodologies to study how the solar panel is changing over time. Obviously, you don’t want to do studies over 20 or 25 years, so there are a lot of accelerated protocols in laboratories where we subject the solar panel to things like temperature cycling. We heat it up and then cool it down, heat it up, cool it down. There are other things to cycle quickly as well, like exposure to humidity, exposure to ultraviolet light, and salt fog. There are all these types of protocols for accelerated testing, and we want to understand how the results of these accelerated protocols correlate to real world performance. We also want to understand the degradation mechanisms, what’s causing these panels to lose that ability to convert sunlight into electricity—whether that’s moisture getting in, oxygen, material reaction, cracking, or something else. The objective is to extend the lifetime of solar panels, so we can make solar panels last 30 or 40 years.

From SigOpt, Experiment Exchange is a podcast where we find out how the latest developments in AI are transforming our world. Host and SigOpt Head of Engineering Michael McCourt interviews researchers, industry leaders, and other technical experts about their work in AI, ML, and HPC — asking the hard questions we all want answers to. Subscribe through your favorite podcast platform including Spotify, Google Podcasts, Apple Podcasts, and more!

Nick Payton
Nick Payton Head of Marketing & Partnerships

Want more content from SigOpt? Sign up now.