Creating Materials with Machine Learning
Rafael Gómez-Bombarelli is an associate professor of materials science and engineering, and, as the title would indicate, he designs new materials. He just doesn’t do it with physical experiments. It’s all computational, as he and his team work at the interface of machine learning and simulations. The targets aren’t specific, he says. They range from capturing carbon dioxide to creating new drugs to making long-duration battery storage, to, well, pretty much anything else. He’s a self-proclaimed agnostic about what material classes to pursue. The intention is the same: to increase the materials available to keep pace with industry needs.
Computers provide an advantage. They can move faster, but that mostly applies to analyzing what’s been given. What Gómez-Bombarelli really wants is to have computers be able to generate ideas, not to replace people – they can’t – but to augment what’s possible.
As he says, “There is this big appetite to develop algorithms that help scientists do science and invent and develop technology.”
Increasing the Number of Catalysts One of the projects that gets him excited is creating new zeolites. They’re catalysts with a composition similar to sand or glass, containing mostly silica. But unlike sand or glass, they’re nanoporous, making them effective at interacting with different molecular substrates. Zeolites themselves aren’t new. They are about 250 and some are already being used for cracking – the process in oil refining in which large hydrocarbon molecules are broken down – and cleaning exhaust emissions from diesel engines.
But when it comes to how many zeolites there could be, “hundreds of thousands are theoretically possible,” and Gómez-Bombarelli says that they could be used to take biomass products from agriculture and turn them into useful chemicals. They may also valorize the CO2 that gets captured from combustion engines and put into aviation fuel. Yet another is the extraction of metals, one in particular being lithium.
The reason for the potential is that zeolites have specific pore sizes and can be selectively matched with a substrate. A recent example is separating ethane from ethylene. They’re similar molecules, he says, with only two hydrogen atoms of difference. But zeolites would allow capturing one while letting the other through.
The work in his lab has been promising. “We’ve been able to push the limits of what particular compositions, synthesis routes and applications were achievable for known materials,” Gómez-Bombarelli says. The next result would be the creation of altogether new ones. It’s a beneficial move, especially for something like the chemical industry, as it will need to decarbonize more and more along the supply chain, from refineries to suppliers to end users. Such a dynamic environment will require the availability of more and cheaper catalysts.
“The whole palette of catalysts that the industry needs to apply will need to change and that can be optimized and accelerated through machine learning,” he says.
Boosting Battery Power Gómez-Bombarelli is also looking to design electrolyte materials for batteries. Lithium-ion is the dominant high-density storage technology. One problem is that batteries with current liquid electrolytes can be flammable. Since that makes it a big risk, “There is a big appetite to replace these liquid with a solid that won’t leak or burn” he says.
That could come from a polymer and the advantage from having that stability is that simpler cells could be constructed. Novel electrolytes could also benefit other battery chemistries, such as using sodium, which is more abundant and less expensive than lithium, he says. Along with the appealing cost reduction, density can also be impacted. The challenge with lithium-ion that it’s difficult to store energy, and with renewables being intermittent, being able to call upon a reliable supply of electricity when it’s needed is key.
Pushing the density of batteries is a driving force. It goes for cars. It goes for airplanes. It means being able to store the same amount of energy in something that is smaller and lighter. It would allow cars run longer and it would help planes since there’d have less weight to carry.
“Discovering new electrolytes could power completely new battery technologies,” he says.
Taking a Load Off of People Defocusing from applications or material classes, Gómez-Bombarelli’s work is about inverse design. Rather than just scoring human suggestions, he wants machine learning to be able to generate specific ideas, be it the correct ceramic catalyst or best insulating polymer.
He’s building the tools for it, and the tools continue to improve. “Now computers can dream up molecules that would satisfy criteria like being the right color or right reactivity,” he says. It helps that they work faster. Although people are good in smaller spaces, they’re limited by time and energy, and when a high-entropy alloy has five, six, seven elements, it’s impossible for a person to decide the concentration of each one. The same holds for what’s required to formulate a new drug.
“We’re good at inferring laws from a few examples, but we struggle to leverage large datasets or learn multi-dimensional relationships,” he says.
Machine learning can step in and streamline the process. The challenge is closing the gap between what computers can dream up and what can be realized. Gómez-Bombarelli wants to embed more understanding about synthetic accessibility and synthesis routes and be able to show its broad applicability.
But computers have their limitations. They can extrapolate and come up with potential materials, but they can’t make decisions or take into account the structure of the supply chain or manufacturing costs, or unexpected pitfalls that may delay development or make it unfeasible.
That’s the irreplaceable human element. Computers can achieve objectives, but it is people who decide what’s important – it’s called domain expertise – and executives who respect and maintain that duality see better results.
“We found that working with partners that not only can evaluate the property of a material or make a material, but can also steer machine learning models away from known issues is critical for success,” Gómez-Bombarelli says.