Interpreting Data-Driven Weather Models
Analysis
This paper addresses the crucial issue of interpretability in complex, data-driven weather models like GraphCast. It moves beyond simply assessing accuracy and delves into understanding *how* these models achieve their results. By applying techniques from Large Language Model interpretability, the authors aim to uncover the physical features encoded within the model's internal representations. This is a significant step towards building trust in these models and leveraging them for scientific discovery, as it allows researchers to understand the model's reasoning and identify potential biases or limitations.
Key Takeaways
- •Applies interpretability techniques from LLMs to analyze data-driven weather models.
- •Identifies interpretable physical features within the model's internal representations.
- •Demonstrates the ability to probe and modify these features, leading to physically consistent changes in predictions.
- •Aims to increase trust and scientific value of data-driven physics models.
“We uncover distinct features on a wide range of length and time scales that correspond to tropical cyclones, atmospheric rivers, diurnal and seasonal behavior, large-scale precipitation patterns, specific geographical coding, and sea-ice extent, among others.”