CASUS Institute Seminar, Dr. Maximilian Amsler, Cornell University
Machine learning (ML) has emerged as a powerful tool to improve the performance of various tasks involved in computational materials science. In my talk, I will cover two topics involving ML and AI to accelerate materials discovery.
First, I will highlight the use of ML, in particular active learning, to guide experimental high-throughput (HT) thin-film processing. Commonly, an exhaustive sampling of possible processing conditions demands a prohibitively large allocation of resources for any realistic system. Sophisticated schemes are thus required to rapidly learn the complex features of this high-dimensional processing space to accelerate such explorative efforts, honing in on the critical points of processing phase diagrams. I will present the scientific autonomous reasoning agent, SARA, which allows autonomous HT materials discovery by integrating robotic materials synthesis, materials characterization, and ML to iteratively explore synthesis phase maps of metastable materials.
Second, I will discuss our recent advances in using ML to accelerate atomistic simulations. In particular, I will present the charge equilibration via neural network technique (CENT) to train accurate interatomic potentials, as implemented in the FLAME atomistic simulation library. Trained on density functional theory (DFT) reference data, CENT provides highly transferable interatomic potentials that accurately reproduce complex potential energy surfaces. Based on examples, I will show how CENT can accelerate the prediction of novel materials and their properties by orders of magnitude compared to DFT.