Local-Transfer Gaussian Process (LTGP) Learning for Multi-fuel Capable Engines

Data-driven engine surrogate models have been widely used to emulate in-cylinder trends of pressure and heat release rate for a wide variety of applications. For example, engines using multi-fuels, e.g., varying fuel cetane number (CN) or different sustainable aviation fuel (SAF) blends, require optimization of input parameters related to fuel injection and ignition assistance to achieve maximum combustion efficiency. Such an optimization task requires building an accurate surrogate model for the engine. Gaussian processes (GPs) are a popular choice: they provide accurate predictions as well as efficient uncertainty quantification to guide decision-making. One challenge, however, is the costly nature of engine combustion experiments, which results in limited data for surrogate training with many input parameters, i.e., with significant variability in engine parameters and conditions. To address this, we present a new local transfer learning Gaussian process (LTGP) surrogate, which transfers knowledge from CFD simulations to learn the expensive combustion response surface, on which limited data is available. A key novelty of the LTGP is the use of a carefully-integrated classifier that regulates when learning should be transferred using ignition misfire data from CFD simulations. Compared to the standard GP surrogate, we show that the proposed model achieves superior prediction performance for engine combustion modeling.