Report on Current Developments in Weather and Climate Forecasting

General Direction of the Field

The field of weather and climate forecasting is witnessing a significant shift towards the integration of advanced artificial intelligence (AI) techniques, particularly large-scale foundation models and transformer architectures. This shift is driven by the realization that AI emulators can rival, and in some cases outperform, traditional numerical weather prediction (NWP) models that rely on high-performance computing systems. The focus is now on developing versatile, multi-use case models that can handle a variety of forecasting tasks, from short-term nowcasting to long-term climate projections, with high accuracy and efficiency.

One of the key innovations is the development of foundation models specifically tailored for weather and climate applications. These models are designed to capture both regional and global dependencies in the input data, allowing for fine-resolution modeling of weather phenomena across different topologies. The use of encoder-decoder architectures, inspired by recent advancements in transformer models, is becoming prevalent, enabling these models to process large volumes of data and accommodate high token counts.

Another notable trend is the exploration of zero-shot super-resolution techniques in weather downscaling. Neural operators, which learn solution operators for partial differential equations (PDEs), are being critically evaluated for their ability to produce high-resolution outputs with higher upsampling factors than seen during training. This approach offers a more efficient alternative to traditional physics-based simulations, particularly in scenarios where high-resolution data is scarce or computationally expensive to generate.

The integration of space-time factorized transformer blocks and position-aware adaptive neural operators is also emerging as a promising direction. These innovations aim to reduce computational overhead while enhancing the model's ability to capture complex spatio-temporal dynamics. Data augmentation strategies are being employed to further boost performance and reduce training consumption, making these models more accessible and scalable.

Noteworthy Papers

• Prithvi WxC: Introduces a 2.3 billion parameter foundation model for weather and climate, demonstrating superior performance across multiple downstream tasks.
• WeatherFormer: Proposes a transformer-based NWP framework that significantly reduces computational overhead while approaching the performance of advanced physical models.
• Large Language Model Predicts Above Normal All India Summer Monsoon Rainfall in 2024: Fine-tunes a large language model for highly accurate monsoon rainfall prediction, achieving an RMSE percentage of 0.07% and a Spearman correlation of 0.976.