Physics-Informed ML Excels with Small Data in Manufacturing

In machining processes, experimental datasets are often small, costly to acquire, and specific to particular materials. This research investigates how physics-informed machine learning (PIML) can be effectively applied under these small-data constraints, using an abrasive waterjet milling dataset for Inconel 718 as a case study. The study makes several methodological contributions. Firstly, it distinguishes between physics-based data cleaning and statistical curation, treating the latter as competing modeling hypotheses. Secondly, it demonstrates that model rankings from small hold-out sets can be unstable, advocating for robust evaluation methods like 10-fold cross-validation, which showed Gaussian Process (GP) variants consistently at the top. Thirdly, the research explores different levels of physics integration, finding that residual learning on a compact physics baseline is competitive for GP models, offering lower variance and interpretability, though it can degrade tree-based models. Bayesian hyperparameter tuning improved some baselines but harmed multi-stage hybrid pipelines at this sample size. The overall picture emphasizes that for small, expensive process datasets, reliable model comparison requires explicit curation hypotheses, robust evaluation, and thoughtful integration of physics into the model.