Machine learning methods for multidimensional analysis of gas gun experiments

Kurzfassung

Accurate calibration of launch parameters is critical in high-velocity impact testing, where the initial pressure determines the resulting impact velocity. In current practice, the inverse relationship between velocity and pressure is estimated using deterministic regression models fitted to small datasets. The approach neglects uncertainty and often leads to large deviations between predicted and observed velocities, reducing reproducibility and increasing experimental costs due to repeated calibration procedures. This work addresses the inverse estimation of initial pressure from a target velocity using a probabilistic framework based on Gaussian Process Regression. The methodology incorporates uncertain inputs and a log-normal assumption on the target variable to reflect physical constraints and heteroscedastic behavior. A systematic ablation study is conducted to evaluate the influence of key modeling choices, including data standardization, mean and kernel functions, objective functions, and training strategies. The results show that the log-normal model with standardized variables provides the best trade-off between predictive accuracy, robustness, and uncertainty calibration. Compared to the conventional regression approach, the proposed model significantly reduces prediction errors while providing calibrated predictive distributions. A limitation of the approach is the small and sparsely distributed dataset, which restricts the characterization of aleatoric uncertainty and limits extrapolation capabilities. Nevertheless, the findings demonstrate that probabilistic regression offers a more reliable and informative framework for inverse parameter estimation in experimental settings.