<![CDATA[ABSTRACT Contemporary engine development programmes face an intensifying conflict between compressed time-to-market schedules, escalating computational demands of high-fidelity simulation, and the exponentially growing dimensionality of the electronic control unit (ECU) calibration space in modern internal combustion engines. This article presents a novel digital twin (DT) framework capable of predicting engine thermal behaviour and exhaust emission characteristics in real time across variable ECU calibration strategies, without requiring a physical engine to be operated at every calibration point of interest. The proposed framework integrates a physics-informed neural network (PINN) thermal model of the engine cooling circuit and combustion chamber wall with a long short-term memory (LSTM) recurrent network trained on high-resolution in-cylinder pressure traces and exhaust emission measurements acquired from a production four-cylinder turbocharged gasoline direct injection (TGDI) engine. The PINN layer enforces energy conservation constraints during network training, ensuring thermodynamic consistency of predictions even in interpolated regions of the calibration space that were not represented in the training dataset. Across a validation set spanning 420 distinct ECU calibration points—varying ignition advance, fuel injection pressure, injection split ratio, boost pressure, and exhaust gas recirculation (EGR) rate—the digital twin predicted coolant outlet temperature with a root-mean-square error (RMSE) of 1.43°C, indicated thermal efficiency with RMSE of 0.61%, brake-specific CO with RMSE of 3.82 g/kWh, brake-specific NOx with RMSE of 2.74 g/kWh, and total hydrocarbon (THC) emissions with RMSE of 1.96 g/kWh. Particulate number (PN) prediction achieved an RMSE of 6.3 × 10⁵ #/cm³ relative to measured values. Real-time inference latency averaged 4.7 ms per operating point on standard automotive-grade embedded hardware, enabling integration into hardware-in-the-loop (HIL) and model-in-the-loop (MIL) calibration workflows. The framework is shown to reduce the number of physical engine test cell hours required to populate a full-factorial calibration map by 73.6%, while maintaining prediction accuracy within the measurement uncertainty of the physical instrumentation. These results demonstrate that physics-informed digital twins represent a viable and transformative tool for next-generation ECU calibration methodology, with direct implications for achieving simultaneous fuel economy improvement and regulatory emission compliance. Keywords: digital twin, ecu calibration, emission prediction, engine thermal model, hardware-in-the-loop, lstm, physics-informed neural network, turbocharged gdi.]]>
Copyrights © 2026
