Max Welling tells Sam Charrington how CuspAI uses generative models, simulation and experimental partners to search for materials with useful properties, including materials for carbon capture and semiconductors. He distinguishes digital modelling from experimental validation and commercial deployment.
Welling describes machine-learning force fields, molecular representations and a staged process for filtering candidates before laboratory experiments. He discusses connecting that process to faster automated laboratories and reports that several projects are still underway.
The conversation turns to connections between generative models and non-equilibrium thermodynamics, then to experimental neural-network designs inspired by waves and symmetry breaking. Welling presents these as research directions and explains why they may help preserve information across long sequences or deep networks.
Watch on YouTube
