Applying Machine Learning for Calibrating Kinetic Monte Carlo Simulations of Additive Manufacturing

Abstract

The kinetic Monte Carlo (kMC) method is a powerful tool for simulating the evolution of processes with known transition rates between states i.e., curvature driven, which covers a wide range of problems and systems, including microstructure evolution in metallic additive manufacturing. Although modeling additive manufacturing with kMC has been shown to accurately reproduce many microstructural morphologies that are influenced by laser speed, the exact relationship between the simulation space and the experimental space remains ambiguous. For calibrating the simulation results, we apply machine learning and a function transformation-based technique. The nonlinear correlations between kMC simulation and experimental data of the effect of laser speed on grain size are discovered using a machine learning-based method. Linear regression is found sufficiently precise when trying to convert in both directions in the specified speed range, however it fails to account for the nonlinearity between the two measurements, adding nonlinear terms using automatic feature engineering (AFR) corrects the error and gives a R2 score of 0.99.