Physics-Informed Machine Learning for Predicting Thermal Distortion in Metal Additive Manufacturing

Abstract

Thermal distortion remains a critical challenge in metal additive manufacturing because repeated localized heating, rapid cooling, thermal expansion, plastic deformation, and residual-stress accumulation can produce warping, dimensional inaccuracies, component rejection, and costly post-processing. This study aimed to develop and quantitatively evaluate an integrated physics-informed machine-learning framework for improving thermal-distortion prediction across metal additive-manufacturing enterprise and research cases. A quantitative, cross-sectional, multiple-case-based design was employed using a structured five-point Likert-scale questionnaire. The cases included industrial manufacturing organizations, universities, research institutions, engineering laboratories, technical centers, and additive-manufacturing technology or software providers working with laser powder bed fusion, directed energy deposition, wire arc additive manufacturing, electron-beam processes, thermal simulation, process monitoring, and dimensional-quality evaluation. Of 350 questionnaires distributed, 318 were returned and 300 valid responses were retained, producing an 85.71 percent valid response rate. The key variables were Physics-Law Integration, Thermal-Process Data Quality and Integration, Thermomechanical Coupling Representation, Model Validation and Generalizability, Computational Efficiency and Real-Time Capability, and Thermal Distortion Prediction Effectiveness. Data were analyzed using descriptive statistics, Cronbach’s alpha, exploratory factor analysis, Pearson correlation, multiple regression, analysis of variance, and regression diagnostics in SPSS. The instrument demonstrated strong reliability, with construct alpha coefficients ranging from .83 to .90 and an overall alpha of .92. The KMO value was .913, Bartlett’s test was significant, χ²(435) = 4,286.37, p < .001, and the six-factor structure explained 69.80 percent of total variance. Thermal-Process Data Quality and Integration achieved the highest mean of 4.14, while Thermal Distortion Prediction Effectiveness recorded 4.09. The five predictors jointly explained 66.60 percent of the variance in prediction effectiveness, R² = .666, adjusted R² = .660, F(5, 294) = 117.25, p < .001. Thermomechanical Coupling Representation was the strongest predictor, β = .30, followed by data quality, β = .26, physics-law integration, β = .21, validation and generalizability, β = .17, and computational capability, β = .11. The findings imply that reliable distortion prediction requires coordinated physical constraints, integrated process data, coupled thermal-mechanical modeling, rigorous validation, and efficient computation rather than reliance on isolated algorithms.