Prediction of Lunar Regolith Excavation Forces Under Percussive and Non-percussive Conditions Using a Data-driven Machine Learning Approach
Year: 2026
Type: dissertation
Abstract: Accurate prediction of excavation forces in lunar regolith is critical for designing efficient excavation systems for future lunar surface operations. Excavation forces in lunar regolith are governed by multiple interacting parameters, such as soil properties, tool geometry, and operational conditions, which collectively influence resistance acting on a cutting tool. Traditional analytical models, however, struggle to represent the complex, nonlinear interactions among these multiple excavation parameters and show limited predictive capability. Moreover, when percussive energy is introduced as a method to reduce cutting resistance, these analytical models are unable to incorporate the associated dynamic effects, further reducing their applicability. While machine learning (ML) techniques have shown promise in terrestrial excavation and geotechnical applications, their use in predicting excavation forces in lunar regolith, particularly under percussive cutting conditions, remains unexplored. To address this gap, four ML regression models, multiple linear regression (MLR), support vector regression (SVR), random forest regression (RFR), and extreme gradient boosting-regression (XGBR) were developed in this study to predict horizontal, vertical and total excavation forces using experimental data from tests on lunar regolith simulants with and without percussion energy. The performance of each model was evaluated using standard regression metrics, and feature importance analysis was conducted based on the best-performing model. Among the models, XGBR achieved the highest prediction accuracy, with R2 values of 0.973, 0.879 and 0.953 for horizontal, vertical and total excavation forces, respectively, surpassing the best analytical model, Lockheed-Martin (R2 = 0.791 for horizontal and 0.723 for total force prediction), by 23% in horizontal force prediction and by 32% in total force prediction, and showing a substantial gain in vertical force prediction where all analytical models yielded negative R2 values, performing worse than a simple mean prediction. Further evaluation using only the non-percussive subset of the dataset showed that, although analytical models performed comparatively better under steady cutting conditions, XGBR still delivered the highest predictive accuracy, reaffirming its robustness across both percussive and non-percussive excavation scenarios. A graphical user interface (GUI) was also developed as an end product to provide accessible force predictions based on user-defined input conditions. This study contributes a novel ML-based framework for predicting excavation forces in lunar regolith across both percussive and non-percussive excavation conditions, offering an advancement to traditional analytical methods to support lunar excavation planning.
DOI: 10.7939/83965
Author: Srinivasan, Shiva Dhahini
Language: English
Subfield: Civil and Structural Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7138319071