Table of Contents

1. Introduction
2. FourCastNet3 Architecture
3. Scalability and Efficiency
4. Performance Comparison
5. Spectral Fidelity and Calibration
6. Inference using Earth2Studio
7. Recommendations
8. Availability of FCN3 Code

Introduction

FourCastNet 3 (FCN3) enables fast and accurate large ensemble weather forecasting, competing with leading ML models like GenCast and surpassing traditional systems like IFS-ENS. FCN3 provides medium-range to subseasonal forecasts with remarkable calibration and spectral fidelity up to 60 days ahead.

FourCastNet3 Architecture

FCN3 utilizes a fully convolutional, spherical neural operator structure based on spherical signal processing, with the addition of local spherical convolutions to spectral convolutions for anisotropic filters. This design ensures both accuracy and computational efficiency through NVIDIA CUDA implementation.

Scalability and Efficiency

Spatial operations like convolutions are distributed using NVIDIA Collective Communications Library (NCCL), allowing training on up to 1,024 GPUs with domain, batch, and ensemble parallelism. FCN3 on a single NVIDIA H100 outperforms GenCast and IFS-ENS, generating a 15-day forecast in a minute.

Performance Comparison

FCN3 matches GenCast in medium-range forecasting accuracy and surpasses IFS-ENS, excelling in speed and computational efficiency. The model maintains spectral properties at all scales, capturing wind intensity variations accurately.

Spectral Fidelity and Calibration

FCN3 preserves atmospheric spectral signatures, reproducing real-world weather patterns sharply even at extended lead times. The model accurately represents wind intensity magnitude across different scales, demonstrated by angular power spectral density analysis.

Inference using Earth2Studio

Running FCN3 inference is simplified with Earth2Studio. A code snippet for a 4-member ensemble inference is provided for easy execution, giving detailed results for analysis.

Recommendations

For optimal FCN3 performance, installing torch-harmonics with custom CUDA extensions and utilizing automatic mixed precision in bf16 format during inference is advised, with default settings in Earth2Studio.

Availability of FCN3 Code

The FCN3 checkpoint and distributed training/inference code are available on GitHub, alongside Earth2Studio for efficient inference processes.