Quantifying uncertainty

Recent advances in machine-learned interatomic potentials (MLIPs) offer an efficient alternative to highly precise but computationally expensive all-atom simulations. Protein folding and other biologically relevant phenomena are of particular interest in this context, as their characteristic time and length scales extend well beyond the density functional theory (DFT) regime. Rigorous uncertainty quantification (UQ) is essential for evaluating the reliability of model predictions, with the latest methods ranging from ensembles approaches to last layer approximations. In this work, we apply a graph neural network–based MLIP to simulate the folding dynamics of a de novo mini-protein chignolin, with a primary focus on probabilistic uncertainty quantification and its propagation within the simulation. We thus look at the possible trade-offs between accuracy, precision, and efficiency of classical forcefields and MLIPs when scaled up to proteins.

Figure source: https://arxiv.org/abs/2310.18278