Model-Based Calibration Optimization Using Machine Learning Algorithms

Calibration optimization of any production vehicle requires hardware prototypes, which could cost up to millions of dollars to be built, demand a lot of engineering time, and add a substantial cost to the vehicle powertrain (PT) design validation (DV) process. Electrified powertrains with their sophisticated supervisory control strategies and thousands of tunable calibration parameters are particularly challenging and time consuming to calibrate. High-fidelity computer models with embedded production software code could be used for initial calibration efforts to reduce the number of prototypes and engineering time required for powertrain calibration. This talk explains how MATLAB^® and Simulink^® vehicle models, machine learning algorithms, MATLAB Parallel Server™ capabilities, and high-performance computing were used for model-based calibration optimization. Over 10,000 different designs worth of 4000–6000 hours’ worth of simulations was completed in less than 15 hours to optimize fuel economy and other attributes.