Innovating ways to explore the materials phase space accelerates functional materials discovery. For breakthrough materials, faster exploration of larger phase spaces is a key goal. High-throughput computational screening (HTCS) is widely used to rapidly search for materials with the desired functional property. This article redefines the HTCS methods to combine multiple deep learning models and physics-based simulation to explore much larger chemical spaces than possible by pure physics-driven HTCS. Deep generative models are used to autonomously create materials libraries with a high likelihood of desired properties, inverting the standard design paradigm. Additionally, machine-learned surrogates enable the next layer of screening to prune the set further so that high-quality quantum-mechanical simulations can be performed. With organic photovoltaic (OPV) molecules as a test bench, the power of this redesigned HTCS approach is shown in the inverse design of OPV molecules with very limited computational expense using only ∼1% of the original physics-based screening dataset.
Raul Ortega Ochoa