Workshop III: Generative Models and the Challenge of Extrapolation

Generative machine learning is opening new possibilities for the inverse design of molecules, materials, and complex physical systems. Rather than predicting the properties of given structures, generative models aim to identify structures and compositions that satisfy desired functional requirements. Recent advances in diffusion and flow-based models have demonstrated considerable potential, particularly in molecular and drug design. However, extending these approaches to materials science raises fundamental questions about physical validity, generalization beyond known structures, and the efficient exploration of large and highly constrained design spaces.

The workshop will bring together machine learning researchers, materials scientists, chemists, physicists, and engineers to examine the methodological foundations and applications of generative and inverse design. Applications will include semiconductors, alloys, catalysts, and functional materials, alongside molecular and ligand design. Particular attention will be given to what materials design can learn from advances in drug discovery, where generative approaches are more mature in certain areas. The workshop will also explore how generative modeling can extend beyond atomistic representations to mesoscopic structures and engineering-scale design problems.

A central theme is how generative models can produce physically valid candidates with targeted properties, particularly in regions of design space that are poorly represented in training data. How should physical constraints and domain knowledge be incorporated into the generation process? To what extent can general-purpose generative architectures learn these relationships from data? How can models explore previously unknown structures and compositions while maintaining physical consistency? The workshop will also examine when generative modeling offers advantages over established optimization and inverse design methods, and how generated candidates should be evaluated and validated.

A further theme is the relationship between generative modeling and physical simulation. Classical sampling methods, including molecular and Langevin dynamics, have inspired modern generative approaches, while learned models offer new possibilities for accelerating the exploration of complex physical distributions. The workshop will examine how these connections can inform efficient sampling algorithms and physically grounded generative models. Building on the electronic structure methods and reduced-order models discussed in the preceding workshops, the aim is to identify the methodological advances needed to turn predictive models into reliable tools for scientific discovery and inverse design.