Three parallel algorithms for classical molecular dynamics are presented. The first assigns each processor a fixed subset of atoms; the second assigns each a fixed subset of inter-atomic forces to compute; the third assigns each a fixed spatial region. The algorithms are suitable for molecular dynamics models which can be difficult to parallelize efficiently—those with short-range forces where the neighbors of each atom change rapidly. They can be implemented on any distributed-memory parallel machine which allows for message-passing of data between independently executing processors. The algorithms are tested on a standard Lennard-Jones benchmark problem for system sizes ranging from 500 to 100,000,000 atoms on several parallel supercomputers--the nCUBE 2, Intel iPSC/860 and Paragon, and Cray T3D. Comparing the results to the fastest reported vectorized Cray Y-MP and C90 algorithm shows that the current generation of parallel machines is competitive with conventional vector supercomputers even for small problems. For large problems, the spatial algorithm achieves parallel efficiencies of 90% and a 1840-node Intel Paragon performs up to 165 faster than a single Cray C9O processor. Trade-offs between the three algorithms and guidelines for adapting them to more complex molecular dynamics simulations are also discussed.

Fast parallel algorithms for short-range molecular dynamics

Solving complex fluid-structure interaction (FSI) problems, which are described by nonlinear partial differential equations, is crucial in various scientific and engineering applications. Traditional computational fluid dynamics based solvers are inadequate to handle the increasing demand for large-scale and long-period simulations. The ever-increasing availability of data and rapid advancement in deep learning (DL) have opened new avenues to tackle these challenges through data-enabled modeling. The seamless integration of DL and classic numerical techniques through the differentiable programming framework can significantly improve data-driven modeling performance. In this study, we propose a differentiable hybrid neural modeling framework for efficient simulation of FSI problems, where the numerically discretized FSI physics based on the immersed boundary method is seamlessly integrated with sequential neural networks using differentiable programming. All modules are programmed in JAX, where automatic differentiation enables gradient back-propagation over the entire model rollout trajectory, allowing the hybrid neural FSI model to be trained as a whole in an end-to-end, sequence-to-sequence manner. Through several FSI benchmark cases, we demonstrate the merit and capability of the proposed method in modeling FSI dynamics for both rigid and flexible bodies. The proposed model has also demonstrated its superiority over baseline purely data-driven neural models, weakly-coupled hybrid neural models, and purely numerical FSI solvers in terms of accuracy, robustness, and generalizability.

Differentiable hybrid neural modeling for fluid-structure interaction

We consider non-spherical rigid body particles in an incompressible fluid in the regime where the particles are too large to assume that they are simply transported with the fluid without back-coupling and where the particles are also too small to make fully resolved direct numerical simulations feasible. Unfitted finite element methods with ghost-penalty stabilisation are well suited to fluid-structure-interaction problems as posed by this setting, due to the flexible and accurate geometry handling and for allowing topology changes in the geometry. In the computationally under resolved setting posed here, accurate computations of the forces by their boundary integral formulation are not viable. Furthermore, analytical laws are not available due to the shape of the particles. However, accurate values of the forces are essential for realistic motion of the particles. To obtain these forces accurately, we train an artificial deep neural network using data from prototypical resolved simulations. This network is then able to predict the force values based on information which can be obtained accurately in an under-resolved setting. As a result, we obtain forces on very coarse and under-resolved meshes which are on average an order of magnitude more accurate compared to the direct boundary-integral computation from the Navier-Stokes solution, leading to solid motion comparable to that obtained on highly resolved meshes that would substantially increase the simulation costs.

Using a deep neural network to predict the motion of under-resolved triangular rigid bodies in an incompressible flow

The goal of this work is to facilitate the design of gas turbine engine particle separators by reducing the computational expense necessary to accurately predict particle motion with machine learning (ML). It is well known that particle ingestion has many detrimental impacts on gas turbine engine operation. Particle ingestion is especially of concern in environments where dust, ash, or ice persist. Accordingly, devices called particle separators are used to remove the particles before they enter the engine. The goal of this work is to improve prediction of particle motion in separators. Ice particles are chosen for this study because of their relevance to civil aviation. The louver particle separator developed by [1] with computational studies is used as a baseline for this work. It has been shown that particle acceleration can be reasonably well-predicted in simple laminar flows using deep feed-forward neural networks [2]. However, there exists a gap for predicting particle acceleration in engineering flows such as the louver geometry of [1]. This is the topic of the current paper. The work builds upon work for laminar flows [2]. In the current work, the complex flow around the louver is decomposed into a laminar shear flow and laminar vortex, and results from [2] are used to analyze the decomposed flows. The accuracy of this decomposition for capturing particle trajectory in the louver is evaluated. 

Predicting Particle Acceleration Using Flow Feature Extraction in a Louver Particle Separator

An accurate prediction of the translational and rotational motion of particles suspended in a fluid is only possible if a complete set of correlations for the force coefficients of fluid-particle interaction is known. The present study is thus devoted to the derivation and validation of a new framework to determine the drag, lift, rotational and pitching torque coefficients for different non-spherical particles in a fluid flow. The motivation for the study arises from medical applications, where particles may have an arbitrary and complex shape. Here, it is usually not possible to derive accurate analytical models for predicting the different hydrodynamic forces. The presented model is designed to be applicable to a broad range of shapes. Another important feature of the suspensions occurring in medical and biological applications is the high number of particles. The modelling approach we propose can be efficiently used for simulations of solid-liquid suspensions with numerous particles. Based on resolved numerical simulations of prototypical particles we generate data to train a neural network which allows us to quickly estimate the hydrodynamic forces experienced by a specific particle immersed in a fluid.

/pdf/a-finite-element-neural-network-framework-for-modeling-55zw4tvl91.pdf

A Finite Element/Neural Network Framework for Modeling Suspensions of Non-spherical Particles

An accurate prediction of the translational and rotational motion of particles suspended in a fluid is only possible if a complete set of correlations for the force coefficients of fluid-particle interaction is known. The present study is thus devoted to the derivation and validation of a new framework to determine the drag, lift, rotational and pitching torque coefficients for different non-spherical particles in a fluid flow. The motivation for the study arises from medical applications, where particles may have an arbitrary and complex shape. Here, it is usually not possible to derive accurate analytical models for predicting the different hydrodynamic forces. However, considering for example the various components of blood, their shape takes an important role in controlling several body functions such as control of blood viscosity or coagulation. Therefore, the presented model is designed to be applicable to a broad range of shapes. Another important feature of the suspensions occurring in medical and biological applications is the high number of particles. The modelling approach we propose can be efficiently used for simulations of solid-liquid suspensions with numerous particles. Based on resolved numerical simulations of prototypical particles we generate data to train a neural network which allows us to quickly estimate the hydrodynamic forces experienced by a specific particle immersed in a fluid.

Martyna Minakowska

Papers

A Finite Element/Neural Network Framework for Modeling Suspensions of Non-spherical Particles

A finite element / neural network framework for modeling suspensions of non-spherical particles. Concepts and medical applications