Prof. Mario Boley
01/01/2026
David Wang Auditorium, 3rd Floor, Dalia Maydan Bldg.
13:30
AI technology holds great potential for the systematic exploration of complicated chemical design and synthesis spaces that currently rely on human trial-and-error experimentation. Specifically, epistemic agents for knowledge acquisition can manage limited experimental resources more efficiently and generate important chemical knowledge faster than trial-and-error processes. However, the design principles for such agents are often misunderstood, starting from their goal definitions to the requirements for their key components: their statistical belief model and the decision strategy that converts beliefs into actions. In this talk, I provide a taxonomy of three types of epistemic agents along with exemplary designs that led to successful applications in materials science. Firstly, I talk about traditional active learning agents, illustrated with the example of automatically generating phase diagrams for polymerisation-induced self-assemblies. Secondly, I discuss blackbox optimisation agents for property maximisation, illustrated with the example of navigating the design space of double perovskites to accelerate the discovery of theoretical materials with high bulk modulus. Finally, I present a new kind of “collector” agent for mapping out complicated synthesis spaces in terms of what materials can form—and if so under what specific conditions. In collaboration with Nobel Laureate Prof. Omar Yaghi, this type of agent was employed to discover a wide range of zeolitic imidazolate frameworks, a subclass of metal-organic frameworks. A key takeaway from these applications is that model interpretability, sound uncertainty quantification, and out-of-distribution generalisation in the small data regime tend to be equally if not more important than raw predictive performance.