Initial Development of a Physics-Aware Machine Learning Framework for Soot Mass Prediction in Gasoline Direct Injection Engines

Abstract

<div class="section abstract"><div class="htmlview paragraph">Calibration of automotive engines to ensure compliance with emission regulations is a critical phase in product development. Control of engine-out particulate emissions, which directly impact the environment and public health, is particularly important. Detailed physics-based models are typically used to gain a rich understanding of the complex physical phenomena that drive the soot particle formation in an engine cylinder. However, such models often fail to correctly represent the highly dynamic nature of the underlying mechanisms under transient combustion conditions. Moreover, most physics-based models were initially developed for diesel engine applications and their applicability to gasoline engines remains questionable due to differences in flame structure and fuel-wall interactions. Black-box models have been previously proposed to predict engine-out soot emissions, but their lack of physical interpretability is an unsolved drawback. To address these limitations, we present a physics-aware twin-model machine learning framework to predict and analyze engine-out soot mass from a gasoline direct injection (GDI) engine. The framework combines a physics-based model with a bagging-type ensemble learning model that both maintains high accuracy and allows physical interpretation of results without using computationally intensive high-fidelity models. This work shows why a one-model-fits-all approach fails in the case of predicting soot emissions due to clustered co-occurrences of operating conditions that cause non-compliant behavior. We compare the performance of the proposed framework with that of the standalone baseline model and a feed-forward deep neural network. Using WLTP data from a 2.0L naturally aspirated GDI engine, the proposed framework predicts engine-out soot mass with an improvement of 29% in the R² value and 21% in the root mean squared error from the baseline physics-based model, without compromising physical interpretability. These improvements are significant enough to warrant further framework development with additional engine datasets.</div></div>