Deep chemistry: AI and the future of chemistry and materials

As a chemical engineer by background I've always been fascinated by the intersection of technology and the physical world. I decided to write up some brief thoughts on recent progress in AI and its application to catalysis and materials.

Background

Catalysts are vital materials that are responsible for facilitating chemical reactions across a wide array of industries from fertilizers to gasoline refining to renewable energy storage. The ability to screen potential catalysts to identify those with commercial viability has been limited to more traditional means of computational chemistry. Due to the complex nature of the molecules and the number of degrees of freedom involved it can be both time and cost prohibitive to use the standard DFT methods to calculate relaxation energies which can be used to help infer the effectiveness of a given catalyst for a given adsorbate. Machine learning methods and specifically deep learning have shown early signs of being able to create models that can accurately determine these relaxation energies and decrease the compute from days to minutes.

Sufficiently accurate DL models could allow us to brute force combinations of atoms for potential catalysts which would drastically reduce the amount of time and energy spent on research and instead allow companies to focus on the synthesis and scale up of the catalytic materials for industrial consumption. Where AlphaFold 2 was able to solve protein folding and unlock massive potential for drug discovery and the biotech world, I aim to identify a team that can do the same with catalysts and unlock a new world of computational chemistry.

The Bet

Initiatives like Meta’s Open Catalyst project are slowly making progress towards using deep learning models to accurately predict the adsorption energy of common adsorbates for climate applications. However, most of the teams that are participating are part of larger organizations and primarily research focused rather than aiming for commercial success.

It is going to require a dedicated company that can strike the right balance of research and application to accelerate the commercialization of deep learning models for catalyst discovery and manufacturing. I would love to invest in a team that has a strong combination of deep learning technical prowess as well as commercial chemical background (process engineering, catalyst sales leadership, etc).

This is an area where the primary risk is technical and execution rather than on the market side. If a team is able to programmatically screen for high efficiency catalysts in a reasonable timeframe and either partner with manufacturers or manufacture them themselves, there is both a large financial victory ($25B+ market) to be had as well as a huge win for humanity in terms of more efficient and lower cost means of energy storage, fertilizer production, and materials production.

I’m convinced that this is a matter of when, not if. Advances in deep learning and more powerful GPUs will continue to unlock opportunities that were previously thought to be too computationally complex to tackle.

Potential Applications

As the world continues to transition to lower carbon sources of energy such as wind and solar we are continually faced with the problem of how to effectively store and utilize energy when the weather does not allow wind and solar to meet the needs of the grid. One solution is using hydrogen as an energy storage mechanism. In this system water is converted to hydrogen and oxygen through electrolysis and the hydrogen is stored until it is needed. The benefit of hydrogen is that we can use existing infrastructure to store and transmit the gas to where it is needed. The downside of this system is that it is energy intensive to split water into hydrogen and oxygen and also uses expensive platinum catalysts. In order for Hydrogen Energy Storage to be a viable system, we must identify catalysts that both improve the efficiency of the electrolysis reaction and eliminate the reliance on rare earth metals.

Our first step to developing catalysts that can improve the efficiency of these reactions is to create a deep learning model that accurately predicts the adsorption energies of various adsorbates on surfaces. This will allow us to correlate to activation energies and reactivity of catalysts and begin a screening process for potential candidates that can be synthesized for in-lab testing to determine efficacy.

Discovering an efficient catalyst for hydrogen energy storage applications would be a huge step in providing sustainable energy storage solutions that don’t rely on extensive mining programs.