Optimizing System Design Performance and Its Accuracy Using Feature Importance

With advancements in model accuracy and computational power, system simulation is increasingly integrated into development tools as a “virtual test bed” alongside experimental testing. However, virtual vehicle and powertrain thermal models still face challenges, particularly in ensuring accuracy across systems developed by various internal and external sources. These models, often built using different software platforms, are difficult to validate consistently, especially when integrated in a Co-simulation environment. This integration can degrade the overall accuracy of the Vehicle Simulation Platform, reducing the return on investment in model development. To address these limitations, this paper proposes the use of machine learning-based feature importance techniques at the vehicle-level simulation stage. Feature importance helps identify the most influential variables affecting system outputs. By focusing calibration and validation efforts on these key variables, the approach aims to improve correlation with physical test data across drive cycles, targeting an accuracy threshold within ±90%