Abstract:We introduce DynAMO, a reinforcement learning paradigm for Dynamic Anticipatory Mesh Optimization. Adaptive mesh refinement is an effective tool for optimizing computational cost and solution accuracy in numerical methods for partial differential equations. However, traditional adaptive mesh refinement approaches for time-dependent problems typically rely only on instantaneous error indicators to guide adaptivity. As a result, standard strategies often require frequent remeshing to maintain accuracy. In the DynAMO approach, multi-agent reinforcement learning is used to discover new local refinement policies that can anticipate and respond to future solution states by producing meshes that deliver more accurate solutions for longer time intervals. By applying DynAMO to discontinuous Galerkin methods for the linear advection and compressible Euler equations in two dimensions, we demonstrate that this new mesh refinement paradigm can outperform conventional threshold-based strategies while also generalizing to different mesh sizes, remeshing and simulation times, and initial conditions.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to improve computational efficiency and the accuracy of solutions through Adaptive Mesh Refinement (AMR) techniques when numerically solving partial differential equations (PDEs). Traditional AMR methods usually rely solely on instantaneous error indicators to guide mesh refinement or coarsening. This results in the need to frequently readjust the mesh to maintain accuracy when dealing with time - dependent problems. Such frequent mesh adjustments not only increase computational costs but may also introduce additional errors. To this end, the paper proposes DynAMO (Dynamic Anticipatory Mesh Optimization), a new method based on multi - agent reinforcement learning. DynAMO aims to reduce the occurrence of future discretization errors by predicting future error propagation and adjusting the mesh in advance. This method can maintain a high solution accuracy over longer time intervals, reducing the frequency of mesh adjustments and thus significantly improving computational efficiency. Specifically, the main contributions and goals of DynAMO include: 1. **Predictive Mesh Optimization**: Unlike traditional AMR methods that are based solely on instantaneous error indicators, DynAMO can predict the propagation of future errors and adjust the mesh in advance, thereby optimizing the mesh before errors occur. 2. **Applicability to Complex Non - linear Systems**: DynAMO is particularly suitable for complex non - linear hyperbolic conservation law equations, for which there are usually no analytical solutions, so traditional AMR methods are difficult to apply effectively. 3. **Robustness and Generalization Ability**: DynAMO designs new observation and reward functions, enabling the policy to be robust and have generalization ability for different mesh sizes, remeshing times and initial conditions. 4. **User - Controllable Error and Cost Goals**: Users can control the relative refinement degree of the policy, that is, the error and cost goals, during evaluation, so as to better meet the needs of specific applications. Through these innovations, DynAMO can provide more efficient and accurate mesh optimization strategies in the numerical simulation of complex non - linear systems.

DynAMO: Multi-agent reinforcement learning for dynamic anticipatory mesh optimization with applications to hyperbolic conservation laws

DynAMO: Multi-agent reinforcement learning for dynamic anticipatory mesh optimization with applications to hyperbolic conservation laws

A Two-Dimensional Hydrodynamics Prediction Framework for Mantle-Undulated Propulsion Robot Using Multiple Proper Orthogonal Decomposition and Long Short Term Memory Neural Network.

Adaptive Swarm Mesh Refinement using Deep Reinforcement Learning with Local Rewards

Deep Reinforcement Learning for Adaptive Mesh Refinement

Swarm Reinforcement Learning For Adaptive Mesh Refinement

Better Neural PDE Solvers Through Data-Free Mesh Movers

Learning Controllable Adaptive Simulation for Multi-resolution Physics

Dynamic Mode Decomposition in Adaptive Mesh Refinement and Coarsening Simulations

A fast dynamic smooth adaptive meshing scheme with applications to compressible flow

An unstructured adaptive mesh refinement for steady flows based on physics-informed neural networks

A reinforcement learning approach for dynamic multi-objective optimization

Unstructured surface mesh smoothing method based on deep reinforcement learning

Asynchronous Distributed Smoothing and Mapping via On-Manifold Consensus ADMM

An offline multi‐scale unsaturated poromechanics model enabled by self‐designed/self‐improved neural networks

Mesh deep Q network: A deep reinforcement learning framework for improving meshes in computational fluid dynamics

Learning robust marking policies for adaptive mesh refinement

Mesh optimization using an improved self-organizing mechanism

GNNRL-Smoothing: A Prior-Free Reinforcement Learning Model for Mesh Smoothing

Reinforcement learning for anisotropic p-adaptation and error estimation in high-order solvers

Solving Expensive Optimization Problems in Dynamic Environments with Meta-learning