Predictive analytics with an advanced Bayesian modeling framework

One of the main limitations in predictive analytics is the acquisition cost of engineering data due to slow-running computer code or expensive experiments. Also, data is often multi-dimensional and highly non-linear in nature, causing problems for standard statistical predictive models. Once data is collected and models are built, many applications require accurate and scalable uncertainty quantification (UQ) solutions to enable robust design and reliable decision making. A modeling framework that addresses these common problems is presented here. The advanced Bayesian modeling framework called “GE Bayesian Hybrid Modeling” (GEBHM) combines simulation and experimental data sources using machine learning techniques and Bayesian statistics to perform UQ, provides detailed sensitivity reports for design engineers, visualizes the problem and its solution succinctly, and aids in developing nextstep plans for decision makers. GEBHM works well for cases with sparse data situations but also handles large-data problems. Fundamentally, GEBHM relies on Markov Chain Monte Carlo techniques for accurately learning the Bayesian model parameters and employs state-of-the-art techniques for self-validation. The GEBHM framework applied to an example engineering design problem requiring FE and/or CFD data is presented. It is demonstrated how the surrogate model is built, how uncertainty sources are quantified and propagated, how the impact of design decisions on performance is quantified, and how the engineering design is optimized. It is shown that the framework is capable of handling problems with limited data, accurately capture the uncertainty, and provide a high level of detail and accuracy for the designers and decision makers.