Model-Based ECU calibration: leveraging Machine Learning to optimize engine test time

Engine ECU calibration is costly, time-consuming, and requires expert effort. Traditionally, for each operating point (fixed speed × load) engineers test many input–parameter combinations, measure the output, and iteratively retune an approach that becomes impractical as the number of parameters grows. This thesis investigates the potential of Machine Learning to develop new control-calibration methods, with the aim of drastically reducing calibration effort and improving robustness in real-world operations. Premise. With surrogate models, the primary objective is to learn a predictor with high accuracy across the entire admissible input domain while using as few training points as possible. Achieving this reduces physical engine tests and there- fore development time and cost. In an open-loop configuration, the only effective lever to cut the number of points is the shape and quality of the initial space-filling DoE—i.e., the workflow is fixed and non-adaptive: a DoE is generated a priori, the points are tested on the engine, outputs are collected, a model surrogate is trained, and calibration is performed on the surrogate alone; this hinges on the surrogate having high predictive capability over the reference domain. This thesis builds a Gaussian Process Regression (GPR) surrogate to enable calibration with fewer bench tests. All experiments refer to one operating point while exploring a nine-dimensional input space. The baseline follows an open-loop workflow: define a space-filling Design of Experiments, acquire data on the engine, train the GP, and calibrate on the learned response surface. Since DoE selection is the tunable stage, I systematically compare Latin Hypercube Sampling (LHS), Maximin LHS, and Sobol designs over increasing sample sizes, assessing predictive accuracy together with coverage metrics. In the open-loop setting, the three space- filling designs delivered broadly similar performance. Building on this analysis, I introduce a closed-loop (active) strategy that uses the GP’s predictive variance to select the next most informative point, adding tests where uncertainty is highest. For the same target accuracy at the chosen operating point, the variance-driven closed loop achieves calibration quality with fewer training points than extending the open loop, directly reducing bench time and cost. The thesis also provides practical guidance on initial Design of Experiments selection under tight budgets and on performing active sampling with a small number of test points.