Hybrid AI-Physics Model Predicts Highly Transient Engine Emissions

Hybrid cars could become significantly cleaner in the stop-start conditions of everyday driving thanks to a new emissions-prediction system that combines artificial intelligence with conventional engine physics. In tests on a turbocharged gasoline engine, the hybrid model predicted nitrogen oxides, carbon monoxide, carbon dioxide and unburned hydrocarbons more accurately than either a machine-learning model or a physics-based simulation working alone. The approach reduced overall prediction error by about 25 percent on an independent real-driving test cycle, suggesting that it could help vehicle computers choose when to use an engine, an electric motor or both. The system is designed for rapid operation on relatively low-power hardware, potentially allowing emissions to become a real-time control target rather than something assessed only after laboratory testing. The study, conducted by researchers at RPTU University Kaiserslautern-Landau in Germany, addresses a central challenge for hybrid vehicles: the sharp, short-lived emission spikes produced when an engine rapidly changes speed and load.

Hybrid-electric vehicles can reduce fuel use by allowing an electric motor to assist during inefficient engine operation, recovering energy during braking and keeping the combustion engine closer to its most efficient operating points. But optimizing a hybrid drivetrain is far more complicated than simply minimizing fuel consumption. A controller must account for engine speed, requested torque, air–fuel ratio, battery charge, traffic, road grade, weather, route and driver behavior. It must also balance carbon dioxide against pollutants such as nitrogen oxides and carbon monoxide, which do not always rise and fall in precisely the same way. Emissions can change dramatically during acceleration from a standstill, rapid load changes or other transient events, when combustion, exhaust temperature and gas transport are evolving simultaneously. The European WLTP and real-driving-emissions regulations were introduced to make testing more representative of road use, but they also make accurate, fast simulations more demanding. A digital twin capable of reproducing these changing conditions could allow engineers to optimize powertrain strategies in software before conducting lengthy physical tests.

The researchers compared three approaches using a BMW B48B20M0 engine, a 2.0-liter, four-cylinder gasoline unit with direct injection, twin-scroll turbocharging and BMW’s VALVETRONIC variable valve system. The engine was operated on a dynamic test bench, where 146 measurement variables—including temperatures, pressures, fuel and air measurements, exhaust composition and signals from the production control unit—were recorded at 5 hertz. The data were synchronized to account for delays as exhaust gases traveled from the cylinders to the measurement equipment. The campaign included 66 steady-state operating points covering the engine map from idle to full load, as well as more than 10 hours of dynamic cycles derived from actual road-driving data. A separate driving profile, preserved with real-world disturbances such as changing driver demand and traffic-related variations, was held back for the final test. That separation was crucial: a model can appear highly accurate if it is evaluated on trajectories it has already seen, but such performance may collapse when conditions change.

The first model represented the engine using physical and phenomenological equations in GT-Suite, a commercial engineering simulation platform. Its combustion description used a spark-ignition turbulent-flame model, two combustion zones and a WoschniGT heat-transfer model. Separate emission submodels represented nitrogen oxides through an extended Zeldovich mechanism, carbon monoxide through a kinetic model and hydrocarbons through a two-plate quenching approach. These models calculate emissions from quantities such as pressure, temperature, mixture composition and residual gas, including variables that are difficult to measure directly inside a running cylinder. The researchers first used three-pressure analysis to tune simulated in-cylinder pressure against measurements, then calibrated combustion and emissions behavior using a genetic algorithm across the 66 steady-state points. At those calibration conditions, average deviations were approximately 10 percent for nitrogen oxides, 35 percent for carbon monoxide and 5 percent for hydrocarbons. Yet the model was much slower than a neural network, achieving an average real-time factor of about three, meaning a simulation could take roughly three times longer than the physical event being reproduced.

The second approach was a long short-term memory, or LSTM, neural network. LSTMs are recurrent networks designed to retain information from earlier points in a sequence, making them useful for systems with memory. In an engine, the current emissions state can depend on the recent history of temperature, fueling, residual exhaust gas and catalyst or manifold conditions, not merely on instantaneous speed and torque. The network received only three readily available inputs—engine speed, torque and air–fuel ratio—but processed them in fixed-length sequences of 41 samples. Its architecture contained eight layers with 20 neurons per layer, and its hyperparameters were selected using a Bayesian search called a tree-structured Parzen estimator. The network was trained on 80 percent of the available data, with 20 percent reserved for validation, while the independent road-derived cycle remained untouched. On that unseen test cycle, the LSTM achieved a normalized root-mean-square error, or RMSE, of 0.0401. Its individual errors were 0.0496 for nitrogen oxides, 0.0586 for carbon monoxide, 0.0154 for carbon dioxide and 0.0172 for total unburned hydrocarbons.

The physics-based model was not discarded simply because it performed poorly under rapidly changing conditions. Instead, its behavior was converted into a fast neural surrogate, called an imitation model. The GT-Suite engine simulation was driven with the same dynamic load profile used on the test bench, and its calculated emissions became the training targets for a second LSTM. This distinction matters: the imitation network was not trained to reproduce measured emissions, which would have made it another ordinary data-driven model. It was trained to reproduce the behavior of the physical simulation, including its physically structured but transiently imperfect predictions. The resulting surrogate could run at a real-time factor below 0.05, compared with about three for the original physical model, and could be implemented in the same PyTorch environment as the main machine-learning predictor. On the independent test cycle, however, the surrogate’s average RMSE was 0.1455, substantially worse than the direct LSTM. Its weakness reflected the limitations of the underlying combustion and emission model, which had been calibrated mainly at steady-state conditions rather than during highly dynamic operation.

The decisive step was to place both fast predictors in parallel and give their outputs to a third, lightweight fusion network. This combination model received 11 inputs: the four predicted species from the data-driven branch, the four predictions from the physics-derived branch, plus torque, engine speed and air–fuel ratio. It then learned how to weight or transform the two forecasts according to the engine’s operating state. The network was not given a hand-written rule saying when physics should dominate; instead, it inferred from the training data when each source was useful. On the held-out driving cycle, the hybrid model achieved an overall RMSE of 0.0295, compared with 0.0401 for the LSTM alone and 0.1455 for the physics-derived surrogate. Nitrogen-oxide error fell to 0.0366, carbon-monoxide error to 0.0410, carbon-dioxide error remained at 0.0154, and hydrocarbon error declined to 0.0148. Relative to the LSTM, the largest improvements were a 27 percent reduction for nitrogen oxides and a 30 percent reduction for carbon monoxide.

An analysis of the fusion network offered a clue to why imperfect physics could still improve the result. Permuting the network’s inputs showed that the data-driven nitrogen-oxide prediction and engine torque had the greatest influence on overall error, followed by air–fuel ratio and engine speed. Physics-derived and data-driven carbon-monoxide predictions also contributed, while the carbon-dioxide channels had little influence in this test. The result suggests that the physical branch supplied useful patterns even when its absolute values were inaccurate. For nitrogen oxides and carbon monoxide, its predictions appeared to retain information about how emissions responded to operating changes; for carbon dioxide and hydrocarbons, the physics-derived trajectories were less informative, so the fusion system relied more heavily on the direct LSTM. This is the central insight of the work: physical information does not need to be perfectly correct to increase the information available to a neural network. A model can use the shape and operating dependence of an imperfect simulation without accepting its forecast literally.

The researchers describe the system as a proof of concept rather than a finished automotive product. The measurement database covered one engine, one fuel and a limited set of environmental and operating conditions. A production system would need testing across cold starts, high altitudes, hot and cold climates, different fuels, engine aging, sensor drift and multiple engine designs. The team also proposes exploring gating networks, physics-based terms in the neural-network loss function and newer time-series transformer architectures. Validation on automotive microcontrollers and closed-loop vehicle testing will be needed to determine whether the model can reliably influence real control decisions. Even with those caveats, the compact input set and fast inference are attractive for hybrid powertrains, where control units must make decisions in milliseconds. By forecasting emissions alongside energy use, future controllers could avoid operating conditions that create pollutant spikes while preserving battery charge and drivability. The study points toward a broader role for hybrid artificial intelligence in engineering: rather than choosing between equations and data, designers may achieve the best results by allowing each to compensate for the other’s blind spots.