Comments on "Toward prediction and insight of porosity formation in laser welding: A physics-informed deep learning framework"

Abstract

Meng et al. (2025) introduce a physics-informed deep learning (PIDL) framework for predicting porosity in aluminum alloy laser welding. Their PIDL model, assessed via SHAP, exhibited superior predictive performance over conventional deep learning models, demonstrated by a 41% reduction in mean square error (MSE). However, feature importances derived from SHAP may be biased, potentially misrepresenting the genuine physical influences on porosity formation. High predictive accuracy does not automatically ensure the reliability of feature importance metrics. This letter underscores the critical need for rigorous statistical validation for reliable feature importance assessments. Integrating robust statistical methods like Spearman's rho, Goodman-Kruskal's gamma, Kendall's tau, and Somers' delta with machine learning enhances the credibility of insights in materials science and manufacturing. Future research should focus on combining ML with robust statistical analysis to improve feature importance reliability and deepen understanding of underlying physical mechanisms.