Mehnatkesh Ghadikolaei, Hossein (University of Alberta), Gordon, David (Univ. of Alberta), Koch, Charles Robert (University of Alberta)

Physics-Informed Neural Networks for In-Cylinder Pressure Prediction in Hydrogen/Diesel Dual-Fuel Engines

Scheduled for presentation during the Regular Session "Diagnostics, optimization and control for hydrogen combustion" (MoA1), Monday, June 16, 2025, 11:20−11:40, Kapel

AAC 2025 11th IFAC International Symposium on Advances in Automotive Control, June 15-18, 2025, Eindhoven, Netherlands

This information is tentative and subject to change. Compiled on May 31, 2025

Abstract

A heavy-duty diesel engine converted to a hydrogen/diesel dual-fuel (HDDF) engine can reduce fossil fuel usage and harmful emissions. To maximize the hydrogen energy share compared to diesel, it is important to monitor and control the combustion process to maintain engine durability. Predicting the combustion process in these retrofitted engines at different operating points using simple combustion models such as the Wiebe function often leads to significant mismatches. These simple models are insufficient for real-time diagnostics, which is essential at high hydrogen replacement ratios. One promising way to improve the accuracy of combustion models is using machine learning (ML) methods, which can potentially enhance the computational speed and predictive accuracy. Combining physics knowledge with ML methods like deep neural networks (DNN) or Kolmogorov-Arnold networks (KAN) is a useful hybrid method called a physics-informed neural network (PINN). This study compares the ML methods and PINN networks for predicting the in-cylinder pressure of the HDDF engine. The most accurate model tested is an integrated KAN and DNN model that incorporates the underlying physics of the system to predict in-cylinder pressure for unseen data. All the models tested utilize crank-angle data and injection timings as the inputs. The results showed that while the ML models have a high prediction error on the unseen data, adding a physics loss function, which penalizes deviations from physical laws, increases the generalization capability, making them a good choice for diagnostic models. The root mean square error for the novel PINN-KAN-DNN network on unseen cylinder pressure data is 15.1 bar, which represents a decrease of 14.8%, 50.3%, and 16.3% compared to DNN, KAN, and KAN-DNN, respectively.