Explainable and generalizable AI-driven informatics for modeling an ultra-precision machining temperature conditioning system

Abstract

Abstract Certain ultra-precision machining requires high-precision control of its operating temperature, which requires high-fidelity modeling that reflects variations in the operating conditions. Machine learning based data-driven models and models derived from physical principles are currently inadequate in this regard. This paper develops a modeling method based on heterogenous informatics towards explainable and generalizable artificial intelligence (AI). The method integrates first principles of a white-box model with machine learning black boxes, resulting in a “gray-box model”. The physical principles play the role of an explainable global meta-structure of the overall system, while the black boxes play the role for generalizable local fitting. The gray-box model thus aggregates implicit variables and relationships between variables that cannot be captured otherwise in a white-box model due to ignored or unmeasurable nonlinearities, including nonlinear trends in the operating conditions. Experiments on an industrial clean-room high-precision temperature control system verify that the output of the gray-box model is closer to the actual system response compared with conventional models under varying operating conditions.