Physics-informed deep learning of gas flow-melt pool multi-physical dynamics during powder bed fusion

Abstract

The effect of inert gas on melt pool dynamics has been largely overlooked but is crucial for laser powder bed fusion (LPBF). Physics-based simulation models are computationally expensive while data-driven models lack transparency and need massive training data. This work presents a physics-informed deep learning (PIDL) model to accurately predict the temperature and velocity fields in the melting domain using only a small training data. The PIDL model can also learn unknown model constants (e.g., Reynolds number and Peclet number) of the governing equations. Furthermore, the robust PIDL algorithm converges very fast by enforcing physics via soft penalty constraints.