Evaluating the Effectiveness and Accuracy of Multi-Table Potentials (MTPs) in Active Learning for Simulation and Prediction

This study assesses the effectiveness of Multi-Table Potentials (MTPs) across various levels and evaluates the validity of a bottom-up training scheme using active learning. We performed active learning runs with MTP levels ranging from 6 to 14 and analyzed their performance in terms of training efficiency and predictive accuracy. The results indicate that the bottom-up approach is effective, as it achieves a comprehensive repre-sentation without requiring preselected configurations. Higher-level MTPs generally required fewer iterations but incurred higher computational costs. MTPs trained at later stages exhibited reduced training iteration counts with no improvement in validation performance compared to lower-level MTPs. Further analysis revealed that MTPs with higher levels while providing more accurate force and energy predictions, showed variability due to initial parameters. Specifically, lower-level MTPs demon-strated greater variance, especially in force predictions, which was minimized in higher-level models. The performance of MTPs in predicting elastic properties and radial distribution functions for solid and liquid potassium revealed that general-purpose MTPs struggled with precise property predictions, necessitating tailored training schemes or higher MTP levels for accuracy. Validation against practical datasets confirmed that MTPs of level 10 or higher generally provided better predictive accuracy, although challenges remain in ensuring stability and accuracy for specific configurations. The study recommends using MTP levels of at least 10 for general-purpose simulations and exploring higher levels for more precise applications.