Emine Kucukbenli1,2
Boston University1,Harvard University2
Emine Kucukbenli1,2
Boston University1,Harvard University2
Neural network interatomic potentials (NNIPs) trained using first principles data have been shown to reach an accuracy close to the reference one at a fraction of the computational cost for relatively simple, well-known systems. Before adapting NNIPs as the workhorse for linear-scaling atomistic simulations however, we need an analysis of the true predictive power and the cost effectiveness of the NNIPs for materials with complex structural and chemical features.<br/> <br/>In this talk I will address this need by first reviewing recent methodological advances in NNIP generation for materials with long range interactions, and then by demonstrating the data-dependence of predictive power on systems ranging from layered materials with dispersion forces to ionic liquids and charged molecules. Finally, I will provide demonstrations of possible paths to increasing cost-effectiveness of NNIPs through techniques developed in the neural network community outside physical sciences, such as transfer learning and pruning.