Introduction to Meso-NH (CNRM/LAERO)

At Weatherwise, our operational infrastructure relies on a diverse suite of Numerical Weather Prediction (NWP) models to generate high-resolution environmental data. A critical component of this suite is Meso-NH, a non-hydrostatic research model developed jointly by the Centre National de Recherches Météorologiques (CNRM) and the Laboratoire d'Aérologie (LAERO)[cite: 61, 1296].

Meso-NH is distinct from standard operational models due to its ability to simulate atmospheric processes across a broad range of scales, from planetary waves down to turbulent eddies. This article outlines the technical architecture of Meso-NH, with a specific focus on its application in meter-scale forecasting and its recent adaptation for Graphics Processing Unit (GPU) infrastructure.

High-Resolution Simulation Capabilities

Meso-NH is a grid-point limited-area model based on a non-hydrostatic system of equations. This formulation allows the model to explicitly resolve vertical motions and atmospheric instability that hydrostatic models must approximate. To achieve high-resolution output essential for localized forecasting, the model employs two primary methodologies: grid nesting and Large-Eddy Simulation (LES).

Two-Way Grid Nesting

To capture the interaction between large-scale synoptic patterns and local weather events, Meso-NH utilizes a two-way interactive grid-nesting technique. This capability allows the simultaneous execution of multiple model domains with progressively finer horizontal resolutions. This is one of the techniques we use at Weatherwise to create high-resolution weather simulations.

• Downscaling: The coarse mesh fields (the "father" model) provide time-evolving boundary conditions to the fine mesh domain (the "son" model).
• Upscaling: Conversely, the fine mesh fields are averaged spatially and used to relax the coarse mesh fields in the overlapping area.

This bidirectional exchange ensures consistency between scales, enabling the simulation of complex interactions, such as the cascade from synoptic cyclones (scales >100 km) down to local wind gusts (scales <1 km).

Large-Eddy Simulation (LES)

Standard NWP models typically operate at resolutions where turbulence must be parameterized. However Meso-NH can function as a Cloud-Resolving Model (CRM) or Large-Eddy Simulation (LES). In LES mode, the model resolves the majority of the turbulent kinetic energy (up to 90%) explicitly, rather than relying on subgrid approximations.

• Resolution: This mode supports horizontal grid spacing on the order of 10 to 100 meters.
• Terrain Handling: The model utilizes a height-based terrain-following coordinate system, which is essential for accurately simulating flows over complex topography and urban canopies.

Physical Parameterizations

To handle subgrid processes, Meso-NH integrates a comprehensive set of physical parameterizations:

• Microphysics: The model employs bulk microphysical schemes. The one-moment ICE3 scheme predicts mass mixing ratios for five water species: cloud droplets, raindrops, pristine ice crystals, snow, and graupel. The advanced two-moment LIMA scheme predicts both mass mixing ratios and number concentrations, allowing for a prognostic representation of aerosol-cloud interactions.
• Surface Exchange: Surface fluxes for sensible heat, latent heat, and momentum are computed by the externalized SURFEX platform. SURFEX divides grid boxes into four tiles (land, town, sea, and inland water) to account for surface heterogeneity.
• Turbulence: The turbulence scheme utilizes a 1.5-order closure, solving a prognostic equation for subgrid turbulent kinetic energy (TKE).

Adaptation to GPU Infrastructure

As simulations move toward hectometric resolutions, computational costs increase exponentially. To maintain operational efficiency, Meso-NH has been ported to GPU architectures. The current GPU-enabled version, MESONH-v55-OpenACC, integrates OpenACC directives, optimized memory management, and revised communication protocols.

Performance and Efficiency

The porting of computationally intensive components—such as advection, turbulence, and microphysics—to GPUs has yielded significant performance gains.

• Speedup: Performance benchmarks on the Adastra supercomputer (AMD MI250X GPUs) demonstrate a 6.0x speedup compared to an AMD Genoa CPU configuration using 64 nodes. By utilizing single precision (32-bit floating point numbers), the speedup increases to approximately 19x.
• Advection: The advection routines specifically show a performance increase of approximately 23x on GPUs compared to CPUs.
• Energy Efficiency: The GPU implementation provides an energy efficiency gain of a factor of 2.3 compared to the CPU version on the same node configuration. This factor increases to 3.6 when using single precision.

Implementation Strategy

The GPU port utilizes a directive-based approach with OpenACC. This strategy allows a single source code to be maintained for both CPU and GPU architectures, preserving the readability of the original Fortran code.

• Memory Management: To mitigate the overhead associated with frequent memory allocation on GPUs, the standard use of automatic arrays was replaced with a pooled memory system. Large arrays are allocated once at initialization, and pointers are used to manage memory segments during execution.
• Custom Preprocessing: An in-house preprocessor was developed to manage architecture-specific loop optimizations, such as converting array syntax to nested loops or `do concurrent` constructs, without code duplication.

Reliability: Bit Reproducibility

For research and operational consistency, bit reproducibility is a strict requirement for Meso-NH. The model employs a library called `MPPDB_CHECK` to verify that results remain identical regardless of the parallel decomposition (number of processors).

This verification extends to the GPU implementation. Results produced by the GPU version are bit-for-bit identical to those produced by the CPU version. To achieve this, intrinsic mathematical functions (e.g., log, exp) were replaced with a custom bit-reproducible library to eliminate discrepancies arising from different compiler implementations.

Operational Applications

The combination of high-resolution physics and GPU acceleration allows Meso-NH to be applied to complex environmental challenges at the meter scale.

Extreme Weather Events

Meso-NH is capable of resolving the cascade of scales inherent in extreme weather. In "Giga-LES" simulations (utilizing over 2.1 billion grid points), the model has successfully represented scales ranging from synoptic cyclones (>100 km) down to local wind gust formation (<1 km). These simulations allow for the explicit representation of turbulent mechanisms leading to surface wind damage.

Coupled Earth System Modeling

Meso-NH supports online coupling with other Earth system models to handle complex feedback loops.

• Ocean and Waves: Through the OASIS3-MCT coupler, Meso-NH interacts with ocean models (e.g., NEMO, SYMPHONIE) and wave models (e.g., WAVEWATCH III). This coupling enables the study of air-sea fluxes during severe weather events without significant additional computational cost.
• Wildland Fire: The model couples with the ForeFire physical fire spread model. This two-way interaction accounts for fire-induced winds and the impact of atmospheric conditions on fire propagation at resolutions as fine as 50 meters.
• Atmospheric Electricity: The CELLS scheme (Cloud Electrification and Lightning Scheme) simulates the full life cycle of electric charges and lightning flashes within the model.

Conclusion

Meso-NH provides a robust framework for high-resolution atmospheric simulation. By leveraging recent developments in GPU computing—such as the geometric multigrid solver and OpenACC porting—and maintaining strict validation standards through bit reproducibility, it serves as a capable tool for resolving meter-scale phenomena within the Weatherwise modeling suite.