Brief

Researchers have introduced RealBench, a new benchmark designed to more accurately evaluate AI weather forecasting models under real-world operational conditions. Unlike previous benchmarks that relied on reanalysis data, RealBench uses low-latency operational analysis and in-situ observations, with a test set from 2025 to prevent data leakage. It also includes specific metrics for high-impact extreme events like heatwaves and tropical cyclones, revealing significant performance gaps compared to traditional benchmarks. AI

Researchers have developed a novel physics-informed generative solver designed to reconstruct complex physical fields from limited data. This method integrates data-driven learning with fundamental conservation laws, ensuring that generated states adhere to physical principles. The approach uses Martingale-Regularized Score Matching for stable prior learning and Physics-Informed Implicit Score Sampling to guide the generation process, demonstrating success in applications like acoustics and meteorological field reconstruction. AI