🔬Why There Is No "AlphaFold for Materials" — AI for Materials Discovery with Heather Kulik

The discussion begins with the question of why one should learn chemistry when AI like ChatGPT has a PhD-level understanding. Heather Kulik shares her experience using AI in accelerated discovery of new materials. Her group uses AI to make predictions faster than traditional computational models. She highlights a project where AI screened thousands of materials, uncovering an unexpected chemical phenomenon that made a polymer four times tougher. The AI design surprised experimentalists, who then made it in the lab and confirmed the results. This tougher plastic has applications in improving overall durability and use of plastics.

Around 2015-2016, Heather Kulik shifted from calling her work "cheminformatics" to embracing machine learning. A student, John-Paul Jeunet, adapted materials design concepts into training neural networks. Kulik sees promise in machine learning for solving multidimensional challenges, such as optimizing metal-organic frameworks for CO2 capture. This involves balancing multiple objectives like stability, CO2 absorption, mechanical stability, and thermal stability. Machine learning offers significant speedups in optimizing across multiple dimensions, enabling the search for materials with multiple desirable properties.

Metal-organic frameworks are useful for gas storage, sensing, separations, and CO2 capture. Their use in catalysis is limited by their stability. These frameworks can place precise chemical groups in specific orientations, allowing targeted interactions with guest molecules. Heather Kulik explains that metal-organic frameworks are like LEGOs for chemists, with building blocks that can be combined in infinite ways to create precise chemistry. Before AI, Kulik used quantum mechanical modeling to understand transition metal catalysis, which is computationally costly, taking hours to weeks for a single prediction.

Heather Kulik identifies gaps in machine learning for chemistry, particularly in reactivity predictions and diverse chemical bonding. There is a lack of data on complex phenomena involving multiple elements. Existing datasets are often focused on "boring chemistry" like organic molecules binding to proteins. She suggests the need for more data sets and leaderboards for areas like how matter behaves when light is shined on it. She notes that while there are repositories of DFT data on crystalline materials, the data often comes from low-fidelity density functional theory, and lacks experimental ground truth.

Heather Kulik discusses the need for a community challenge to break open problems in material science, similar to CASP in the protein world. She questions whether machine learning potentials can really replace conventional physics-based modeling. While new foundation models look good initially, they can exhibit "wacky things" like molecules falling apart. She calls for more rigor in evaluating whether these models can truly replace physics-based modeling, rather than just fitting data that may lack quality.

The discussion shifts to bridging the gap between computational design and experimental chemistry. Heather Kulik notes that high-throughput synthesis and experimentation are being explored, but some experiments are easier for humans than for autonomous systems, and vice versa. She also highlights the importance of the process in getting materials to the device scale, which is an area where machine learning is still at "ground zero."

Heather Kulik discusses integrating textual information from papers into AI models. They use natural language processing and graph digitization to extract datasets of properties from the literature. One challenge is that people interpret their results in different ways, which can bias the models. LLMs are sensitive to false positives, requiring significant time to check the accuracy of ingested data. To address the bias towards previously reported results, they train models on literature but apply them to new structures.

Heather Kulik suggests creating an initiative or multi-institutional funding resource to get data in a high-throughput automated way. She envisions user facilities where computational researchers can design experiments and have them executed, with data collected publicly. She emphasizes the need for systematization of how results get reported so that they can be machine learning ready from day one. She also notes the importance of materials for climate change.