A Comparative Study of Metamodeling Techniques for Predictive Process Control of Welding Applications

Abstract

Metamodeling is investigated as a tool for predictive process control of welding applications. The motivation for predictive process control is to incorporate physics-based models in place of empirical models that are statistically derived from repeated physical tests. Predictive process control promises to be particularly useful for high flexibility applications, such as frequent modifications of material or geometry. One of the primary challenges is that accurate physics-based, thermal models of the welding process usually require computationally expensive software such as FLUENT [1], while faster models, such as the analytical models of Rosenthal [2], are typically less accurate. Metamodeling or surrogate modeling is investigated as an alternative modeling technique for combining the accuracy of the detailed models with the speed of the faster models, thereby enabling real-time control of the welding process. Four of the most promising metamodeling techniques—polynomial regression, multivariate adaptive regression splines, kriging, and support vector regression—are selected based on a set of preliminary criteria. Each technique is used to build surrogate models of a representative welding process, based on FLUENT data obtained with statistically designed experiments. The techniques are compared with respect to accuracy, speed of model construction, and speed of prediction. Implications for predictive process control of a welding process are also discussed.