The automotive industry is facing increasingly stringent regulatory constraints, driving the need for faster and more efficient powertrain development. This results in higher systems complexity, making internal combustion engine calibration progressively more challenging to meet performance and emissions targets. This, combined with the manual nature of traditional calibration workflows, leads to a time‑consuming process that heavily relies on human expertise. Although virtualization can reduce development time and costs, the overall workflow remains largely dependent on manual decision‑making and iterative refinement.
In this context, this work presents a virtual calibration framework based on a genetic algorithm, aimed at the automated optimization of engine calibration maps to satisfy performance and emissions constraints, while reducing manual effort. Each calibration map is represented through a polynomial parameterization. Specifically, a generic three‑dimensional polynomial with map‑specific order encodes the shape of each map, ensuring smoothness which directly impact on drivability. Accordingly, the calibration problem is reformulated as the optimization of a compact set of polynomial parameters that uniquely define the full set of calibration maps, rather than individual set-points. Each candidate solution is assessed by generating the corresponding calibration maps and simulating the engine behavior through a neural‑network‑based digital twin, providing predictions of operating conditions, hardware limits, performance metrics, and emissions.
The proposed framework was validated on a passenger‑car diesel engine, considering a reduced yet representative set of calibration maps, including Main injection SOI, air mass, boost pressure, and injection rail pressure. The objective of optimization was the minimization of BSFC, subject to an upper bound constraint on NOx emissions. The global optimization process explored approximately 10^6 different calibration candidates within about 36 hours, leveraging parallel computation on a standard laptop. The results indicate that the procedure can deliver multiple near‑optimal preliminary calibration solutions, providing an effective starting point for subsequent manual finetuning.