AI Summary & Key Takeaways
AI InsightThe fixed geometry, nominal fuel and simplified single-zone generator, together with the absence of experimental validation, preclude interpreting the reported gains as demonstrated improvements on real engines.
📌 3 Key Takeaways
Abstract
Accurate prediction of the in-cylinder pressure trace does not necessarily preserve pressure-derived combustion descriptors. This study uses a synthetic spark-ignition benchmark to examine whether explicit supervision of these descriptors changes surrogate accuracy. Six neural-network variants share a 5382-parameter architecture and training budget, with cumulative supervision of pressure derivatives, apparent heat release, peak descriptors, combustion phasing and gross indicated work. The benchmark comprises 960 operating points divided into training, validation, interpolation and simultaneous speed–load extrapolation sets, plus a closure-shift test with altered combustion and wall-heat-transfer relations. At the specified 1° output resolution, validation-selected A4 reduces interpolation CA50 mean absolute error from 0.0938° to 0.0688° and maximum pressure-rise-rate (MPRR) error from 0.0506 to 0.0303 bar/°, corresponding to improvements of 26.6% and 40.1%. Pressure RMSE increases from 0.0980 to 0.1031 bar, and gross IMEP error increases from 0.0296 to 0.0529 bar. A4 retains a lower MPRR error during speed–load extrapolation, with higher pressure and work errors, while all models deteriorate under the closure shift. The 1° MPRR discretisation discrepancy of 0.01702 bar/° is comparable in magnitude to the 0.0203 bar/° model improvement. These results establish a redistribution of approximation error within the prescribed benchmark. The fixed geometry, nominal fuel and simplified single-zone generator, together with the absence of experimental validation, preclude interpreting the reported gains as demonstrated improvements on real engines.