Now showing 1 - 2 of 2
  • Publication
    Massively Parallel Molecular-Continuum Flow Simulation with Error Control and Dynamic Ensemble Handling
    (2022-01) ;
    Wittmer, Niklas
    In coupled molecular-continuum flow simulations, molecular dynamics (MD) simulations exhibit thermal fluctuations. Finding a way to minimize the impact of these fluctuations on the CFD solver, e.g. in terms of stability, and to control the corresponding statistical error plays a key role in order to obtain reliable results. In this paper, statistical error analysis is employed for MD simulations to determine the statistical error in flow velocities and the number of MD data samples to bound this error. The corresponding error estimator is augmented by a dynamic ensemble handling approach, which allows to couple a variable number of MD simulation instances to a single CFD solver. The ensemble members can be simulated independently from each other over separate coupling time intervals, enabling a high level of (MPI-based) parallelism. Adding or removing MD simulations to/from the ensemble allows to regulate the error and keep it under a prescribed threshold. All functionality is implemented in the massively parallel macro-micro-coupling tool (MaMiCo). We validate our approach by coupled molecular-continuum Couette flow simulation for liquid argon and provide scalability tests on up to 131.072 cores. The computational overhead for handling the dynamic MD ensemble is found to be rather negligible.
  • Publication
    Metadata only
    Open boundary modeling in molecular dynamics with machine learning
    (2020) ;
    Wittmer, Niklas
    Molecular-continuum flow simulations combine molecular dynamics (MD) and computational fluid dynamics for multiscale considerations. A specific challenge in these simulations arises due to the “open MD boundaries” at the molecular-continuum interface: particles close to these boundaries do not feel any forces from outside which results in unphysical behavior and incorrect thermodynamic pressures. In this contribution, we apply neural networks to generate approximate boundary forces that reduce these artefacts. We train our neural network with force-distance pair values from periodic MD simulations and use this network to later predict boundary force contributions in non-periodic MD systems. We study different training strategies in terms of MD sampling and training for various thermodynamic state points and report on accuracy of the arising MD system. We further discuss computational efficiency of our approach in comparison to existing boundary force models.