Uncertainty quantification for wheeled locomotion machine learning predictions on soft soil

Abstract

Surrogate modeling with machine learning (ML) techniques is becoming increasingly popular in engineering and physical fields. Models based on statistical inference often lack uncertainty measures, which are crucial for comprehensive predictions. Uncertainty quantification (UQ) addresses these challenges, especially in tasks lacking analytical solutions or extensive experimental data, such as modeling wheel locomotion on soft soils. High-fidelity data from real experiments or precise particlelevel simulations are scarce, adding inherent uncertainty to statistical models. In our paper we analyzed the UQ aspect of the terramechanical surrogate modeling. Our surrogate model leverages the probabilistic nature of Gaussian processes to facilitate uncertainty calculation and make further analysis easier. We extend UQ analysis into a new multi-fidelity model for wheel locomotion. Our work aims to improve the model’s interpretability and optimization through uncertainty propagation, sensitivity analysis and uncertainty decomposition.