Uncertainty-aware multi-objective optimisation of a variable-compression-ratio hydrogen-diesel engine using physics-informed machine learning surrogates

Hydrogen substitution raises both the efficiency of a compression ignition engine and its oxides of nitrogen, and compression ratio moves the two together, so the favourable operating region must be located rather than reasoned to. Surrogate models are used for this, but where one is reported to improve because physical knowledge was embedded, the physics and the model carrying it are varied together and cannot be separated. Applying the same physical treatment to nine regressor families under one protocol separates them. On a variable-compression-ratio hydrogen–diesel engine at 32 conditions, spanning compression ratios of 12–18, loads of 22–90% of rated and hydrogen energy fractions of 0.261–0.401, the treatment improved 22 of 27 model–target combinations, and gains were largest where the model had least capacity of its own: random forest prediction of oxides of nitrogen rose from 0.821 to 0.949, and ridge regression with four parameters matched a neural network carrying 2881. Adding a second fuel mode improved prediction where the mechanism is shared and degraded it where the apparent activation temperature differs by 25%, so the benefit follows the mechanism rather than the quantity of data. Efficiency per unit of oxides of nitrogen fell 27% between compression ratios of 12 and 18, placing the compromise near a ratio of 13 at 70% load, where each gram of hydrogen displaces 4.01 g of diesel. Predictive intervals covered 62–81% against a nominal 95%; calibrated, they span 35% of the measured efficiency range, so the recommendation is a region rather than a point.