A data-driven method for optimization of classical interatomic potentials

Training an interatomic potential (IP) to predict material properties requires appropriate experimental or first principles, e.g. density functional theory (DFT), ground truth values, along with an efficient optimization algorithm to select parameter values. Atomistic simulations are required to check each proposed parameter set, which can be costly depending on the desired property. We present an optimization algorithm that leverages existing model parameter data with a dual neural network approach to accelerate the fitting process. We extract model parameters from OpenKIM and identify correlations between them and select material properties. We then create a surrogate model and couple it with an optimization algorithm to determine the desired IP parameters. This information can be leveraged, along with DFT training data and additional atomistic simulations, to further optimize the parameters. We believe this framework can be used to expedite the optimization process and enable better models for large scale properties.