| Abstract Scope |
The rapid adoption of artificial intelligence (AI) in manufacturing has created new opportunities for process modeling, optimization, and quality control. However, many current AI applications rely heavily on data-driven approaches that function largely as black boxes, providing limited physical interpretability and often requiring extensive retraining when process conditions change. For manufacturing processes governed by well-understood physical mechanisms, a fundamental question arises: should machine learning replace physics, or should it complement it?
This presentation introduces a physics-dominant machine-learning framework developed for welding operating-window determination. The central premise is that the primary structure of a manufacturing process should be described by governing physics whenever possible, while machine learning should be used selectively to account for secondary effects and residual errors. Physics-informed analytical models are first derived from the dominant heat-generation mechanisms of individual welding processes, establishing physically meaningful operating-window boundaries. For processes exhibiting stochastic transition behavior, the framework is extended to probabilistic boundary descriptions that enable risk-based process design. Machine-learning methods are then incorporated only as residual-correction tools, preserving physical interpretability while improving predictive accuracy.
The methodology is demonstrated using multiple case studies involving high-frequency induction welding and resistance spot welding. Results show that relatively simple physics-based models capture the dominant operating-window structure, while residual learning provides targeted improvements near process-transition boundaries and under conditions exhibiting increased variability. Independent validation studies further demonstrate the adaptability of the framework to new applications and operating conditions.
Beyond welding, the proposed approach illustrates a broader paradigm for industrial AI in which physics, domain knowledge, uncertainty quantification, and machine learning are integrated into a unified modeling strategy. The framework provides a practical pathway toward interpretable, transferable, and industrially deployable AI systems for intelligent manufacturing. |