Past Talks

Previous conference and workshop talks.

Accelerating Implicit Contact via Matrix-Free Operators on Non-Conforming Interfaces

In continuum solid mechanics, robust and accurate simulation of mesh-to-mesh contact remains a challenging problem to scale to modern GPU-based architectures. In this work, we present a novel matrix-free strategy for evaluating finite element integrals of nonlinear functions and their linearizations consistently across non-conforming and solution-dependent interfaces. To avoid assembling complex coupling matrices for segmented face intersections, we establish a single layout of dynamically determined quadrature points shared by both contacting meshes. By utilizing performance-portable libCEED operators to extract the necessary points for each surface cell, we can map quantities directly across the interface without copying the underlying vector of values. We demonstrate the efficiency of this zero-copy, matrix-free approach for mesh-to-mesh contact simulation in Ratel, our high-performance implicit solid mechanics library.

Grain-Resolving Simulation of Mock-High-Explosives with Ratel Implicit MPM

Ratel is a matrix-free, high-order, implicit finite element (FEM) and material point method (MPM) solid mechanics package. In this work, Ratel is used to simulate consolidation experiments for mock plastic-bonded explosives (mock-PBX) manufacturing. The consolidation experiments consider the confined compression of a 5 mm-diameter cylindrical die filled with prills—clumps of crystalline mock-high-explosive grains coated with a nitro-plasticized polymer binder. Due to the granular nature of the materials, the irregular distribution of each material within the sample, and the presence of large voids, the material point method is well-suited to this problem. The distribution of grain size is bimodal with peaks between 150-300 um and less than 45 um; thus, resolving the fine grains requires on the order of 1 billion material points. We discuss the methods for initializing material properties from voxelized computed tomography (CT) data and tracking deformation of the material points, challenges of scaling to hundreds of millions of points, and current simulation results.

Addressing Numerical Challenges in Frictional Contact Simulation for Finite-Deformation Solid Mechanics

The numerical simulation of contact phenomena in implicit, finite-deformation solid mechanics codes presents significant challenges due to the introduction of nonlinear and non-smooth operators. These complexities necessitate specialized linear and nonlinear solvers to be performant at scale. In this context, our finite element package, Ratel, distinguishes itself by employing high-order matrix-free methods, contrasting with the prevailing industry standard of low-order solvers applied to sparse assembled Jacobian matrices. Ratel capitalizes on the robust solver infrastructure provided by the Portable Extensible Toolkit for Scientific Computing (PETSc) and integrates the flexible matrix-free library libCEED, enabling highly scalable performance across both CPU and GPU architectures. Ratel supports level-set based frictional contact, which can be enforced by either Nitsche’s method or a penalty method. This talk will elucidate approaches to several numerical challenges associated with these contact formulations, particularly for high-order and matrix-free methods. These challenges include the computation of material stresses on contact surfaces, the solution of the asymmetric, indefinite, and/or poorly conditioned Jacobian matrices, and the numerical instability and slow nonlinear solver convergence behavior due to non-smooth friction models.

Using Decentralized Learning to Reduce Communication in Column-Partitioned, Multi-Agent Systems

Multi-agent systems introduce new challenges to distributed computing, such as unreliability and a need for data localization, which require robust decentralized learning methods capable of minimizing communication overhead. In multi-agent systems, each agent typically stores local, time-series data columns which must be communicated to other agents in order to apply traditional, row-partitioned distributed learning algorithms; such data-sharing is infeasible in unreliable or communication-delayed environments. State-of-the-art, column-partitioned decentralized learning methods avoid such communication bottlenecks through aggregation of approximate local optimization results between neighbors over a less connected network topology. In this talk, we will focus on the recent advances and outstanding challenges of decentralized learning for column-partitioned multi-agent systems.