Predicting Engine Parameters with Cost Efficient AI Models – An Experimental Method Validation

Abstract

Artificial intelligence (AI) methods have rapidly become a best practice in various industrial applications due to their exceptional predictive abilities. Compared to traditional physical and chemical models, AI-driven approaches typically offer faster computation times without sacrificing accuracy. This makes them well-suited for enhancing or replacing conventional methods in engine technology. Building on this potential, our previous work focused on developing AI-based tools to accelerate the development of novel e-fuels for internal combustion engines. These AI tools require representative training datasets created through expensive engine dyno measurements. To address this challenge, we developed a general methodology to improve the cost-efficiency of dataset creation. This paper presents the experimental validation of this methodology by assessing its performance on seven different predictive tasks. Following the proposed framework, we designed and conducted an engine dyno experiment, then developed AI models to predict seven critical engine performance and emissions parameters: center of heat release, ignition delay, peak combustion temperature, peak pressure rise rate, brake thermal efficiency, exhaust opacity and NO_x emissions. The results demonstrated the effectiveness of the methodology, with five out of seven models achieving excellent predictive performance on unseen test data (R2 > 0.97). Peak pressure rise rate and opacity models had slightly lower performance (R2 > 0.94), however, given the well-known challenges associated with predicting these parameters, the results are acceptable.