RLChina 2025 KEYNOTE

Chemistry underpins the quality of life and the life of industry. Compounded challenges of resource constraints, carbon-neutrality and increasing demand call for re-thinking of current practices and urgent exploration of untapped chem-space. Fundamentally, chemistry is about the harnessing of interactions, or bonds, between atoms and molecules for the manufacturing of desired products with minimal atom and energy costs. Despite centuries of development and many Nobel Laurates in the area, the fundamental challenge of bond breaking and forming is still not in our full grasp. This is limited by our poor understanding of the mechanistic processes and inadequate mastery of the entwined dynamics at multi- spatial and temporal scales.   Unprecedented development in machine learning (ML) provides great opportunities to augment the inadequacies of theory and experiment in chemistry, to accelerate the discovery of new structures and their manufacturing processes. From recent examples, the presentation will offer the following perspectives: 1) Integration of “deep-research”, entwining theory, simulations, machine-learning (ML) and experiment, for the discovery of molecules and materials for wide applications; 2) Development of electronic structural descriptors, valence band-alignment criteria, and machine-learning potentials to extend deep insight into charge-carrier dynamics and chemical reaction pathways; 3) Embed theoretical principles in machine-learning for interpretable reasoning, avoiding “black-box” findings ; 4) Delivery of tailored functional structures and modular devices (e.g. batteries / electrolysers / reactors / refinery systems) with digitally entwined software packages to demonstrate an intelligent manufacturing platform; 5) Establishment of self-driving Robo-Chem Synthesis Laboratory for ML-informed formulation, synthesis and rapid screening of chemical candidates; and, ultimately, 6) Reverse engineering and forward exploration, from performance demands to in silico ML optimisation of synthesis routes and chemical preform formulations. Overall, the future will witness a new paradigm in AI entwined chemistry, ChemAIstry, for accurate and efficient innovation, training and education in Chemistry.

Over the past few decades, we have witnessed AI achieving superhuman performance in specific tasks, primarily in gaming. In just the last few years, we have observed foundation models making breakthroughs toward general human intelligence—matching or even surpassing human performance across a wide range of tasks. What comes next? Clearly, the next milestone will be general superintelligence. In this talk, I will share my thoughts on the path to general superintelligence, with a focus on reinforcement learning's crucial role. I will also explore how embodied AI and scientific AI (AI4Science) contribute to this journey toward GSI and their connections to reinforcement learning.