Adaptive Design of Experiments for Automotive Engine Applications Using Concurrent Bayesian Optimization

Abstract

Most practical automotive problems require the design of experiments (DoEs) over a number of different operating conditions to deliver optimal calibration parameters. DoE is especially crucial for automotive engine calibration problems due to its increasing complexity and nonlinearity. As the complexity of the system increases, the DoE applications require a significant amount of expensive testing. However, only a limited number of testings are available and desired. The current work addresses this issue by presenting an adaptive DoE method based on Bayesian optimization to find optimal parameter settings with a significantly reduced number of physical testings (or function evaluations). To further improve optimization efficiency, this work presents a new approach: concurrent Bayesian optimization, which searches for optimal DoE under multiple operating conditions simultaneously. The method utilizes a surrogate model and a novel concurrent evolutionary multi-objective optimization method: concurrent non-dominated sorting genetic algorithm-II, to solve adaptive DoE in multiple operating conditions with a limited number of function evaluations. The experimental study is carried out on a gasoline engine calibration problem using a high-fidelity GT-SUITE™ engine model. The experimental results demonstrate the effectiveness of the algorithm by optimizing engine performance with a significantly reduced number of expensive testings to achieve accurate optimal solutions. The method simultaneously performs engine calibration at eight different operating conditions using only 500–600 testings, compared to the traditional approach, where each operating condition requires 300–500 testings independently to achieve optimal results.