Application of Machine Learning to the Development of Ti Interatomic Potentials

Within computational physics and material science an important area of research is that of identifying potential function parameters that tie electronic scale bonding phenomena to atomic scale nanostructures. Force-matching makes use of quantum mechanical ab initio electronic structure calculations to produce computationally efficient analytic interatomic potentials (IP) for use in classical molecular dynamics (CMD) simulations of nanoscale defects. Although the principle of force-matching is straightforward, the practical application of the method and others like it is labor intensive and computationally prohibitive. For this reason the creation of useful and accurate IPs is a cumbersome process. The more sophisticated the IP, the larger the associated computational cost, as more fitted parameters are required. Furthermore, even sophisticated IPs tend to have a limited scope of usefulness.

In recent years, machine learning (ML) has emerged as a versatile and powerful tool within the scope of computational physics. Here, the viability of ML to more efficiently improve the accuracy of less sophisticated IPs for a titanium system is explored. The approach uses ML to optimize less sophisticated IPs through the use of an artificial neural network (ANN) which seeks to identify and apply more fitted parameters which create a “delta”. This delta, when applied to an IP, creates a hybridized IP. The intent is for the hybridized IP to have increased usefulness, with a lower computational cost as compared to traditional methods. This work ties into a larger effort to create a streamlined and automated process for making and fitting analytic IPs to facilitate high-throughput calculations.