[Paper] Scalable Mesh Coupling for Atmospheric Wave Simulation

Source: arXiv - 2603.02971v1

Overview

The paper presents a scalable mesh‑coupling algorithm that lets two independent atmospheric solvers exchange data across overlapping computational meshes in real time. By enabling a globally consistent solution for in‑situ coupled wave simulations, the technique opens the door to higher‑fidelity weather and climate models that can run efficiently on modern supercomputers.

Key Contributions

• General‑purpose overlapping‑mesh interpolation that works with any pair of solvers sharing a common physical domain.
• Scalable implementation built on the p4est library, achieving near‑linear strong‑scaling up to tens of thousands of MPI ranks.
• Robust handling of non‑conforming meshes, including adaptive refinement on one side while the other remains coarse.
• Performance evaluation on a realistic atmospheric wave case, showing < 5 % overhead compared to running the solvers independently.
• Open‑source reference implementation (MIT‑licensed) that can be plugged into existing CFD/atmospheric codes.

Methodology

1. Domain decomposition – Each solver owns its own forest of octrees (or quadtrees) that discretizes the atmosphere. The two forests overlap in a “coupling region.”
2. Ghost‑layer construction – A lightweight MPI communication step builds a shared list of interface cells, regardless of mesh resolution.
3. Interpolation kernel – For every cell in the overlap, the algorithm finds the donor cell(s) from the other mesh and performs a high‑order polynomial interpolation. The kernel is designed to be embarrassingly parallel: each MPI rank works on its local subset of overlap cells.
4. Load‑balanced communication – Using p4est’s space‑filling curve ordering, the authors redistribute overlap data to keep communication balanced even when one mesh is heavily refined.
5. In‑situ coupling loop – The two solvers advance a time step, invoke the interpolation routine, and then continue with the updated boundary values, all without writing intermediate files.

Results & Findings

• Strong scaling: On the NERSC Perlmutter system, the coupling routine maintained > 80 % parallel efficiency from 256 up to 32 k MPI ranks.
• Weak scaling: Adding more cells to the overlap region increased runtime linearly, confirming the algorithm’s O(N) complexity.
• Accuracy: Benchmarking against a reference solution showed that the interpolation error stayed below 1 % of the wave amplitude, even with a 4× resolution mismatch between meshes.
• Overhead: The coupling step added only 3–5 % to the total simulation time, far less than the I/O cost of traditional checkpoint‑based coupling.

Practical Implications

• Hybrid modeling pipelines – Weather agencies can now run a high‑resolution mesoscale model together with a coarse global model in a single job, preserving fine‑scale features without costly data staging.
• Adaptive refinement on‑the‑fly – Developers can trigger mesh refinement in regions of interest (e.g., thunderstorms) while the surrounding domain stays coarse, with the coupling algorithm handling the data exchange automatically.
• Multi‑physics extensions – The same framework can be reused for coupling atmospheric dynamics with ocean, land‑surface, or chemical transport models, accelerating the development of Earth system simulations.
• Reduced storage costs – Because the coupling is performed in‑situ, there is no need to write large intermediate fields to disk, cutting both storage requirements and post‑processing time.

Limitations & Future Work

• Assumes structured‑grid solvers – The current implementation relies on octree/quad‑tree based meshes; extending to completely unstructured finite‑element meshes will require additional mapping logic.
• Single‑direction interpolation – The paper focuses on one‑way data transfer; bidirectional coupling (e.g., feedback loops) is not yet benchmarked.
• Physics‑agnostic – While the algorithm is mathematically sound, coupling complex source terms (e.g., moist convection) may need customized interpolation stencils.
• Future directions include adding support for GPU‑accelerated solvers, exploring higher‑order conservative interpolation schemes, and integrating error‑controlled adaptive refinement driven by the coupling error itself.

Authors

• Hannes Brandt
• Tim Griesbach
• Matthew Zettergren
• Scott Aiton
• Jonathan Snively
• Donna Calhoun
• Carsten Burstedde

Paper Information

• arXiv ID: 2603.02971v1
• Categories: cs.DC, cs.CE
• Published: March 3, 2026
• PDF: Download PDF