Gómez-Bombarelli Feature

RAFAEL GOMEZ-BOMBARELLI: So I'm Rafa Gomez-Bombarelli. I'm an associate professor in materials science and engineering. And my group works at the interface of machine learning and simulations for materials design.

My group works in a number of materials classes. We are purely computational. We don't do physical experiments, but we do collaborate obviously with people that do. So we're pretty agnostic about the materials classes and the applications that we pursue.

For instance, we're particularly excited about a class of heterogeneous catalysts that are called zeolites. They are these nanoporous materials that have a regular porous structure that can accommodate different substrates. They're used today in cracking, in industrial cracking of oil. And they're also used for cleaning exhaust emissions from diesel engines industrially at the billion dollar scale.

And they are a nice combinatorial problem that is very tractable for machine learning. There is many possible materials one could make with many different compositions. And each of these will result in different potential applications. So we're utilizing our tools to connect the synthesis recipe to the material that emerges to the properties that it would have and optimize them all at the same time to achieve new catalysts for diverse applications.

So a zeolite is a nanoporous ceramic material. So it's made of the same fundamental composition as sand or glass. It's mostly silica. But it can contain other elements that fine-tune its properties for either extracting, sieving molecules or for catalysis by creating customized, catalytic sites that carry out a reaction.

Some of the future potential applications of zeolites for energy and sustainability would be for instance, in biomass conversion, to take biomass products after agricultural applications and upgrade them to useful chemicals. That's a possibility. Another would be in CO2 upcycling. Zeolites are very selective catalysts. And they may be made in a way that allows for very selective transformations of CO2 and that would allow valorising the CO2 that gets captured from combustion engines or from other combustion sources and put it into valuable chemicals, such as aviation fuel.

Another possible application is for separation of metals. In particular, there is a lot of appetite for extracting lithium out of a possible sources. And because of their selective nature, zeolites can be applied to extract these high value metals from solution.

The oil and gas industry already utilizes zeolite at scale for cracking. As the chemical industry needs to get decarbonized, both the suppliers of raw feedstocks out of refineries and the end users that upgrade them to more value chemical can benefit from more selective and cheap catalysts. As we decarbonize the production of energy and chemicals, 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.

[MUSIC PLAYING]

[MUSIC PLAYING]

RAFAEL GOMEZ-BOMBARELLI: Another project we are actively working on is designing electrolyte materials for batteries. Today, lithium ion batteries are the dominant high-energy density energy storage technology. However, higher energy densities could be achieved by replacing some of the materials in them, and also, the electrolyte material that helps shuttle lithium from the anode to the cathode today is typically a flammable liquid, which results in large risks.

There is big appetite to replace this liquid with a solid that won't leak out if the pouch is perforated. We're utilizing high-throughput simulations, machine learning on literature data, and robotic experimentation, in collaboration with others at MIT, to design new polymer or liquid electrolytes that address these challenges. Another advantage of solid electrolytes is that they are themselves mechanically stable and allow to utilize their own stability and physical stability to construct simpler cells.

Furthermore, more efficient electrolytes would allow exploring other battery chemistries, such as sodium, which is more abundant and cheaper than lithium, or even so-called di-cations, such as magnesium, which is again, very abundant but hard to utilize in batteries. So discovering new electrolytes could power completely new battery technologies.

Today, two large challenges to current lithium ion battery technologies are cost and energy density. Lithium ion batteries are still challenging to utilize as grid scale storage and renewable sources produce energy intermittently, and ideally, we would like to store that energy for when it's needed. However, lithium ion batteries are proven hard to scale cost-wise and technology-wise to grid scale. So on the one hand, better electrolytes could allow cheaper chemistries-- perhaps sodium or possibly magnesium-- and make grid energy grid scale storage more accessible.

So on the one hand, there is the chances of enabling new chemistries that are just a lot less costly for grid scale. In this case, maximizing the amount of charge carriers is also critical. For example, lithium ion batteries, this is batteries that utilize lithium as the anode, would be very interesting for aviation because they can achieve very high charge densities. They can be very compact. However, they struggle because, among other reasons, electrolytes are not stable enough. So the electrolytes in contact with the lithium metal surface degrade.

So finding a new electrolyte would enable, potentially, lithium air batteries, which would themselves be dense enough for aviation applications. Energy density is defined as the amount of charge or total energy that a battery can store by either unit mass or volume mass, and depending by the application, volume or mass may be more important. But in general, the amount of capacity a battery has is related to this energy density.

In this project, we have demonstrated machine learning innovations to deal with chemical data and formulation data at the same time, such that we can optimize the formulation of additives that make an electrolyte as well as the individual chemistries of the battery components. In addition, we have utilized simulations to drive discovery in a quicker way than an experimental trial and error, and in collaboration with experimental partners here at MIT, we have demonstrated what's called closing the loop between computer-driven innovation and physical experiments. This platform is demonstrating discovery of new electrolytes which now must be subject to techno-economic analysis to see whether they can be scaled up and be compatible with current devices.

[MUSIC PLAYING]

[MUSIC PLAYING]

RAFAEL GOMEZ-BOMBARELLI: A general approach we're developing that is applicable through multiple classes of materials is so-called inverse design. In inverse design, we flip over the way traditional material science works. And instead of utilizing machine learning or simulations or experiments to evaluate a candidate that a human suggests, we flip those functions and we create generative functions that, given a desired property, can automatically dream up or generate new materials. Just like ChatGPT can write a poem when you ask it to, we are developing tools that can produce the right ceramic for your catalyst or the right insulator for your polymer.

So on the one hand, this has followed the development of generative AI in general. Over the last four or five years, generative AI has demonstrated that machine learning models can dream up or so-called hallucinate answers to very complex questions posed by humans. We were one of the first proposers of generative AI for chemistry back in 2016. And over the years, the tools have gotten increasingly and increasingly better to the point that now computers can dream up molecules that will satisfy criteria like being the right color or the right degree of stability or the right reactivity.

The key challenge that remains-- and this relates to a white paper we're writing under the umbrella of MIT's generative AI enterprise-- is the execution gap between what machine learning can dream and what can be realized physically and scaled up. Just like ChatGPT lives only in your computer and only needs to interact with you via the computer screen, materials only succeed when they are made at scale and transformed into products. So there is this extra layer of complexity and challenges that we need to infuse into the machine learning. And we are developing algorithms that not only can suggest some material that is good for some application but are aware of the stability of that material or its cost or its supply chain or how to manufacture it.

[MUSIC PLAYING]

[MUSIC PLAYING]

RAFAEL GOMEZ-BOMBARELLI: Some material design problems have to do with optimization in high dimensions. So for instance, so-called high entropy alloys are alloys made for a-- from a number of individual elements, maybe five, six, seven elements. So choosing which elements in which concentration becomes this sort of intractable combinatorial problem. The same is true for applications like drug discovery.

There are more molecules one could make than atoms in the universe. So there is not literally enough mass to even try to make every molecule. Humans are good at designing and moving in a smaller design spaces. But these very large number of options make hard for human intuition to navigate design spaces.

Over the last year and a half or two, for instance, building up on the success of alphafold, we have seen generative AI starting to transform protein engineering, where again, computers can dream up proteins, these very long sequences of amino acids, that were outside the ability of human design.

As everything related to machine learning, training data is a critical aspect, whereas generative models for images, or texts, are trained over essentially all of the internet, the data sets in material science or chemistry are much more limited. They're siloed, they're oftentimes private, and they're individual companies, paywalls, or internal assets.

So training data is a clear disadvantage compared with other domains. In our particular case, we believe that simulations, and the rules of physics can make up for that because we can-- we get the ability to generalize or extrapolate beyond the training data through our understanding of physics-based laws.

So I would say data is a big challenge. And perhaps simulations and physics can make up for it. And this is an active area of research. And then the other question that is still ongoing is, sure. The machine learning models can dream up potential materials that might be useful. But the model's imagination may not be aware of synthetic routes, supply chain, manufacturing, or cost.

Materials can only be deployed in the market if they are cost competitive. So it is on us, and this is current work, to connect the computational imagination to the very real constraints of making and productizing materials.

[MUSIC PLAYING]