Designing Machine Learning Surrogates using Outputs of Molecular Dynamics Simulations as Soft Labels

Many outputs of molecular dynamics simulations of soft materials are associated with statistical uncertainties. We show that these uncertainties can be utilized to informate the training of deep neural networks for designing machine learning surrogates aimed at predicting the relationship between input variables and simulation outputs. The approach is illustrated with the design of a surrogate for molecular dynamics simulations of confined electrolytes to predict the complex relationship between the input electrolyte attributes and the output ionic structure. We demonstrate that the prediction error for samples in the unseen test data can be significantly reduced by utilizing a modified loss function that leverages the uncertainties in the output ionic distributions during training. Using such soft labels for the ground truth facilitates a sampling mechanism that implicitly expands the dataset with more samples as the model undergoes training over many epochs, yielding a surrogate with higher generalizability. The surrogate predictions for the ionic density profiles are found to be in excellent agreement with the ground truth results produced using molecular dynamics simulations.