Neural network (NN) potentials promise highly accurate molecular dynamics (MD) simulations within the computational complexity of classical MD force fields. However, when applied outside their training domain, NN potential predictions can be inaccurate, increasing the need for Uncertainty Quantification (UQ). Bayesian modeling provides the mathematical framework for UQ, but classical Bayesian methods based on Markov chain Monte Carlo (MCMC) are computationally intractable for NN potentials. By training graph NN potentials for coarse-grained systems of liquid water and alanine dipeptide, we demonstrate here that scalable Bayesian UQ via stochastic gradient MCMC (SG-MCMC) yields reliable uncertainty estimates for MD observables. We show that cold posteriors can reduce the required training data size and that for reliable UQ, multiple Markov chains are needed. Additionally, we find that SG-MCMC and the Deep Ensemble method achieve comparable results, despite shorter training and less hyperparameter tuning of the latter. We show that both methods can capture aleatoric and epistemic uncertainty reliably, but not systematic uncertainty, which needs to be minimized by adequate modeling to obtain accurate credible intervals for MD observables. Our results represent a step toward accurate UQ that is of vital importance for trustworthy NN potential-based MD simulations required for decision-making in practice.

In this work we present a parsimonious set of ordinary differential equations (ODEs) that describes with satisfactory precision the linear and non-linear dynamics of a typical laminar premixed flame in time and frequency domain. The proposed model is characterized by two ODEs of second-order that can be interpreted as two coupled mass-spring-damper oscillators with a symmetric, nonlinear damping term. This non-linear term is identified as function of the rate of displacement following [Formula: see text]. The model requires only four constants to be calibrated. This is achieved by carrying out an optimization procedure on one input and one output broadband signal obtained from high-fidelity numerical simulations (CFD). Note that the Transfer Function (FTF) or describing function (FDF) of the flame under investigation are not known a-priori, and therefore not used in the optimization procedure. Once the model is trained on CFD input and output time series, it is capable of recovering with quantitative accuracy the impulse response of the laminar flame under investigation and, hence, the corresponding frequency response (FTF). If fed with harmonic signals of different frequency and amplitude, the trained model is capable of retrieving with qualitative precision the flame describing function (FDF) of the studied flame. We show that the non-linear term [Formula: see text] is essential for capturing the gain saturation for high amplitudes of the input signal. All results are validated against CFD data.

No abstract