Application of Machine Learning Techniques for Inverse Prediction in Manufacturing Process Chains

The use of physics-based simulations of manufacturing processes for the prediction of material properties and defects is increasingly widespread in industry. Such simulation tools help answer “what-if” questions, and a materials engineer may have to conduct a number of simulations for decision making. In practice the engineer is often seeking a solution to an “inverse problem”, i.e., prediction of inputs/process parameters for a desired outcome. Such an inverse problem is often solved by formulating it as a constrained-optimization problem. Extensive simulation in the input-parameter space when performing the optimization is avoided by approximate response surfaces iteratively constructed using simulations executed while traversing the design space. In this paper, we present a case study on the application of machine learning techniques to address such inverse problems. Specifically, using data from physics-based simulations, we explore the use of two different kinds of models constructed by machine learning. The first approach constructs a “generative” model (a Bayesian network), from which input values can be obtained directly from output values, without the need of an optimization step. It does however need additional knowledge in the form of conditional (in)-dependences between process parameters, intermediate state variables, and outputs. The second is a purely predictive machine learning model capturing complex non-linearity followed by the use of optimization methods (simulated annealing) for inverse prediction. We present results for modelling of a heat treatment process chain involving carburization, quenching and tempering. Our findings are as follows: For the range of output-values we examined, the predictive model performs better than the generative model. However the generative model has the ability to discover multiple solutions to the inverse problem, unlike in the traditional response-surface-based design of experiments. Thus the generative approach may prove more useful for exploratory industrial practice in the long run.