Credibility Assessment of Machine Learning in a Manufacturing Process Application

Abstract

We present a framework for establishing credibility of a machine learning model used to predict a key process control variable setting to maximize product quality in a nuclear component manufacturing application. Our model coupled a purely data-based machine learning model with a physics-based adjustment that encoded subject matter expertise of the physical process. Establishing credibility of the resulting model provided the basis for eliminating a costly intermediate testing process that was previously used to determine the control variable setting.