Machine Learning-accelerated Molecular Design of Innovative Polymers: Shifting from Thomas Edison to Iron Man

Polymeric materials are key enablers in aerospace, mechanical, civil, and environmental engineering, such as reverse osmosis membranes for water treatment and desalination, coatings for building skins, and antifouling materials, etc. Nevertheless, the design and development of innovative polymers have been an experimental-driven and trial-and-error process guided by experience, intuition, and conceptual insights. This Edisonian approach is often costly, slow, biased towards certain chemical space domains, and limited to relatively small-scale studies, which may easily miss promising compounds. A grand challenge in designing these polymeric materials is the vast design space on the order of 10100, defined by the almost infinite combinations of chemical elements, molecular structures, and synthesis conditions. To tackle this challenge, I will present our recent works on developing a data-driven molecular simulation strategy that can efficiently discover and design novel polymers with unprecedented yet predictable combinations of properties. Specifically, we use machine-learning techniques to build a meaningful chemistry-property relation for polymeric materials. Then, we utilize generative adversarial networks, combined with Reinforcement Learning models, for the inverse molecular design of innovative polymers. Eventually, we apply the experimentally validated molecular dynamics simulations to verify these molecular designs. We expect this work can address a wide range of scientific questions in computational materials design and synthesis-structure-property relationships for polymeric materials. It will also benefit the broader scientific community and industry, which are interested in developing new types of polymers for medical, automotive, packaging, building and construction applications.