(377e) Accelerating Electronic Structure Calculations with Machine Learning

Electronic structure calculations have produced mechanistic insights and have led to systematic design principles for catalysis. However, a typical atomistic calculation involves much computational "waste": often, thousands of single-point electronic structure calculations are performed to produce each one or two configurations which are ultimately published. Recently, methods have been developed which can "learn" from the output of electronic structure calculations, allowing for the prediction of the energies and forces of atomic configurations that they have not been exposed to. In this talk, we will discuss new algorithms that can use such machine learning techniques to strategically accelerate and extend electronic structure calculations, through such routines as transition-state searches and hybrid quantum-mechanics/molecular-mechanics approaches. These methods will be introduced in terms of our open-source machine-learning package, "Amp".