Accounting for Prediction Uncertainty from Machine Learning for Probabilistic Design

Abstract

Computational or simulation models derived from domain physics principles are commonly used in engineering design. But the models are often expensive. Machine learning is increasingly used to generate surrogate models to replace the computational models. Almost all machine learning models have errors, and the errors are unknown at a new design point. The model error can be estimated by quantifying the model uncertainty, and the estimated model uncertainty is now available in many machine learning techniques. This study uses a shaft design to investigate the effects of model uncertainty of surrogate models on the design result when the design is also subject to data (aleatory) uncertainty, which comes from the randomness in the model input. Gaussian process regression is used for the design. This study also discusses the unique features of model training of the surrogate for a computational model, such as noise-free training points and uncertainty free in prediction at a training point. The study indicates that different treatments of model and data uncertainties can result in quite different designs and that new ways for predicting design performance and optimizing design need to be developed.