Domain-informed Adaptive Sampling for Generalizable PINNs in Metal Additive Manufacturing via Conditional Flow Matching
The work introduces a novel adaptive collocation sampling strategy for Physics-Informed Neural Networks (PINNs) used in metal Additive Manufacturing (AM). It proposes a two-stage framework using Conditional Flow Matching to dynamically select where to place training points based on process conditions. This method significantly improves generalization compared to traditional static sampling by focusing computational effort where thermal gradients are most complex.