Researchers at RWTH Aachen University and the University of Alberta tested an artificial-intelligence controller on a hydrogen-diesel engine. Compared with a diesel-only controller at the same speed and load, it cut the mass of exhaust particles by 61.1% but raised nitrogen-oxide emissions by 105.1%. The result comes from 4,900 engine cycles on one cylinder. The paper is an unreviewed preprint submitted on September 14.
A hydrogen-diesel engine replaces part of the diesel fuel with hydrogen. That can reduce soot and direct carbon-dioxide emissions, but hydrogen burns quickly and can raise combustion temperatures, which encourages nitrogen oxides to form. The balance changes as the driver changes load. A fixed engine map cannot readily adjust all of these effects from one combustion cycle to the next.
The researchers trained a gated recurrent unit on 99,800 measured cycles. This type of neural network uses recent cycles to predict what will happen next. It receives commands for diesel quantity and timing, the amount of hydrogen and the previous load. A model-predictive controller tests possible commands against those forecasts, sends the best first adjustment to the engine and repeats the calculation after the next measurement.
The test used a modified Cummins engine running at a constant 1,500 revolutions per minute. Hydrogen entered one of its four cylinders, and the researchers measured that cylinder's exhaust separately. For about six and a half minutes, the controller followed a load schedule it had not seen during training. A production diesel controller ran the same schedule as the comparison. The engine had no turbocharger or exhaust-gas recirculation during the trial.
At the setting reported in the title, hydrogen supplied an average 39.7% of the fuel energy. The controller kept the commanded load more closely than the diesel baseline while producing the large fall in particle mass and the equally clear rise in nitrogen oxides. Its computer solved each adjustment in 3 to 7 milliseconds, within the 18-millisecond window available between decisions. No calculation deadline was missed during the run.
The trial shows that the control loop can operate on engine hardware, but it covers one cylinder, one speed and a short laboratory schedule. The particle measurement also has substantial uncertainty. The authors are extending the work to a larger multicylinder engine. Variable-speed tests with turbocharging, exhaust treatment, fuel-use measurements and longer duty cycles will determine whether the controller can manage the emissions trade-off in a production vehicle.
Sources
- Alexander Winkler and colleagues, “Hydrogen–Diesel Dual-Fuel Engine Operation Using Real-Time GRU-Based Nonlinear Model Predictive Control,” arXiv, September 14, 2026. Abstract, authorship, submission date and principal results.
- Full paper. Engine configuration, controller sequence, test protocol, timing, results and limitations.
- Lead image: fuel-injection hardware on a Gardner 6LW diesel engine, photographed by Zenwort/Adi Meyerhofer and reused unmodified under CC BY-SA 4.0. The museum photograph is illustrative and does not show the reported experiment.
