Uncertainty Quantification at the Molecular–Continuum Model Interface

Abstract

Non-equilibrium molecular dynamics simulations are widely employed to study transport fluid properties. Observables measured at the atomistic level can serve as inputs for continuum calculations, allowing for improved analysis of phenomena involving multiple scales. In hybrid modelling, uncertainties present in the information transferred across scales can have a significant impact on the final predictions. This work shows the influence of force-field variability on molecular measurements of the shear viscosity of water. In addition, the uncertainty propagation is demonstrated by quantifying the sensitivity of continuum velocity distribution to the particle-based calculations. The uncertainty is modelled with polynomial chaos expansion using a non-intrusive spectral projection strategy. The analysis confirms that low-order polynomial basis are sufficient to calculate the dispersion of observables.

Machine Learning Directed Optimization of Classical Molecular Modeling Force Fields

“…For example, surrogate-assisted optimization (also known as black-box or derivative-free optimization) uses computationally inexpensive surrogate model evaluations to emulate the outputs of a complex computer simulation, e.g., computational fluid dynamics, finite element analysis, or molecular simulations. Several different types of surrogate models have been successfully applied to molecular simulations for uncertainty quantification , and force field parametrization. ,− Linear regression response surface models were used to predict the optimal combination of scaling factors for the charge and Lennard-Jones (LJ) parameters of General AMBER force field (GAFF) to reproduce four properties of organic liquid electrolytes. While easy to implement and moderately successful at improving the force field’s accuracy for some of the properties, this method was limited by the choice of statistically significant parameters in the response surface .…”

Predictive scale-bridging simulations through active learning

“…While the effects of microscopic random fluctuations on coarse-grained models using hybrid algorithms have been studied for a few problems including linear diffusion 19 and the inviscid Burgers 20 , Navier–Stokes 21 , and Ginzburg–Landau 22 equations, direct atomistic-continuum scale-bridging models have been challenging to implement. This is primarily due to the disparate length and timescales resulting in inaccurate information transfer across scales and poor computational performance 23 . Although there are some studies on hybrid atomistic-continuum models (including coupling LBM and MD) for dense fluids, these focus on very simple systems, e.g., single-phase laminar flow past or through a carbon nanotube 24 .…”

Hybrid machine-learning-assisted quantification of the compound internal and external uncertainties of graphene: towards inclusive analysis and design

“…To capture the parametric uncertainties of the Lennard-Jones (LJ) inter-atomic potential, Angelikopoulos et al 27 implemented the Bayesian probabilistic framework in the parameters of the LJ potential and proposed an adaptive surrogate model to demonstrate less computationally expensive MD simulations of liquid and gaseous argon. Zimon ´et al 28 adopted a polynomial chaos expansion to depict the influence of variation in the LJ potential parameters on the molecular simulations of the shear viscosity of water. Similarly, Messerly et al 29 quantified and propagated the uncertainties associated with the LJ potential parameters for the prediction of critical constants of n-alkanes.…”