Small Data Challenges in AI for Materials Science

There is great excitement in materials science about accelerating materials development and chemical synthesis via AI and ML. Traditionally, materials science has evaluated proposed material designs using time-consuming physical experiments and compute-intensive calculations, resulting in a slow, expensive design loop. This, and the lack of a programming standard for matter, hinders efforts to combat climate change, fight disease, and improve the human condition.

Research at the interface of AI/ML and materials science has begun to accelerate this process. Supervised learning can screen out materials that are likely to lack critical properties; Bayesian optimization and active learning can efficiently search a materials design space; Computer vision can improve the efficiency and the reproducibility of materials characterization.

Yet, this effort faces a major challenge: available data are much fewer than in traditional AI. We must learn smarter, making better use of heterogeneous and high-dimensional experimental measures and computational predictions, and assimilating multimodal structured data. In contrast to many other application areas of AI, there is abundant domain knowledge in the form of physical laws; incorporating this knowledge into the learning process is crucial to its success.

To advance AI-supported materials, this workshop will bring researchers from materials science together with those working in AI/ML, focusing on the small data challenges.  Jointly, we will identify common problems and develop plans for tackling them.