Machine learning potentials (MLPs) are increasingly used to simulate functional materials with quantum mechanical accuracy at a fraction of the computational cost of traditional ab initio methods. By learning surrogate models of interatomic forces from high-level electronic structure calculations, MLPs enable large-scale molecular dynamics simulations that support advances in areas such as organic photovoltaics, battery electrolytes, and hydrogen storage. However, MLPs face significant criticism for their limited interpretability, lack of built-in uncertainty quantification, and failure to incorporate experimental data during training. Indeed, most current models rely solely on theoretical force predictions from density functional theory, neglecting valuable experimental data such as scattering and spectroscopic measurements that contain critical insights into interatomic interactions. To address these limitations, we propose a Bayesian framework that integrates experimental data directly into MLP training using a Gaussian process (GP) potential committee. On the experimental side, a GP potential is learned from scattering data using structure-optimized potential refinement, a probabilistic iterative Boltzmann inversion technique. Subsequently, a GP potential trained on ab initio data is combined with the previous model within a Bayesian committee machine, yielding a unified prediction with principled uncertainty quantification informed by both quantum theory and experimental scattering measurements. By combining ab initio calculations with experimental data in a transparent and probabilistic manner, this approach provides interpretable, uncertainty-aware MLPs aimed at improving modeling for functional materials.
