ai extreme weather and climate by Zhi Li

In this episode of AI Extreme Weather and Climate, Allen and Sydney explore a major breakthrough in meteorological AI: predicting rare but high-impact events like typhoons. While foundation models like Pangu-Weather excel at everyday forecasts, they often stumble during extreme anomalies due to severe data imbalance. We dive into a newly proposed framework called Target Concept Tuning (TaCT) which acts like a "specialized meteorologist" inside the neural network. By using Sparse Autoencoders to untangle superposed features, TaCT automatically identifies the exact internal concepts that cause the model to fail during extreme weather. It then selectively fine-tunes only those specific concepts, dramatically improving typhoon forecasting accuracy without causing the model to "forget" how to predict normal weather patterns. Tune in to learn how making AI more interpretable is making our early warning systems safer and more reliable!

This paper introduces NeuralGCM, a hybrid atmospheric model that integrates machine learning with traditional differentiable physics to improve global precipitation simulations. Unlike older models that rely on high-resolution simulations for training, this framework is trained directly on satellite observations, specifically the IMERG dataset. By leveraging this observational data, the model effectively corrects common biases in extreme weather events and the diurnal cycle of rainfall. In comparative tests, the model outperformed the ECMWF ensemble in mid-range forecasting and showed superior accuracy over CMIP6 climate models. Additionally, the architecture is exceptionally efficient, running simulations at speeds orders of magnitude faster than conventional general circulation models. These findings suggest that hybrid neural models offer a more reliable and computationally accessible path for predicting future climate impacts.

The episode describes the Continuous Flow Operator (CFO), a novel neural framework for learning the continuous-time dynamics of Partial Differential Equations (PDEs), aimed at overcoming limitations found in conventional models like autoregressive schemes and Neural Ordinary Differential Equations (ODEs). CFO's key innovation is the use of a flow matching objective to directly learn the right-hand side of the PDE dynamics, utilizing the analytic velocity derived from spline-based interpolants fit to trajectory data. This approach uniquely allows for training on irregular and subsampled time grids while enabling arbitrary temporal resolution during inference through standard ODE integration. Across four benchmarks (Lorenz, 1D Burgers, 2D diffusion-reaction, and 2D shallow water equations), the quintic CFO variant demonstrates superior long-horizon stability and significant data efficiency, often outperforming autoregressive baselines trained on complete datasets even when trained on only 25% of irregularly sampled data.

This episode provides a comprehensive monograph on diffusion models, detailing their foundational principles through three unifying perspectives: the Variational View (related to VAEs and DDPMs), the Score-Based View (rooted in EBMs and Score SDEs), and the Flow-Based View (connecting to Normalizing Flows and Flow Matching). The core concept involves defining a continuous forward process that adds noise and then learning a corresponding reverse process—a Stochastic Differential Equation (SDE) or Probability Flow Ordinary Differential Equation (PF-ODE)—to transform noise back into data. Much of the discussion focuses on the mathematical equivalence of these different formulations, the tractable training objectives (like Denoising Score Matching), and advanced techniques for accelerating the slow sampling process, including sophisticated numerical ODE solvers (like DPM-Solver) and distillation methods (such as Consistency Models). Finally, the monograph explores the theoretical connection between diffusion models and Optimal Transport (OT), suggesting that diffusion is related to, but not generally equivalent to, solving the optimal transport problem.

This episode introduces RainSeer, a novel, structure-aware framework for reconstructing high-resolution rainfall fields by treating radar reflectivity as a physically grounded structural prior. The authors argue that existing interpolation methods fail to capture localized extremes and sharp transitions crucial for applications like flood forecasting. RainSeer addresses two main challenges: the spatial resolution mismatch between volumetric radar scans and sparse ground-level station measurements (AWS), and the semantic misalignment caused by microphysical processes like melting and evaporation between the radar's view aloft and the rain that reaches the ground. The framework employs a Structure-to-Point Mapper for spatial alignment and a Geo-Aware Rain Decoder with a Causal Spatiotemporal Attention mechanism to model the physical transformation of hydrometeors during descent, demonstrating significant performance improvements over state-of-the-art baselines on two public datasets.

This episode dives into FlowCast-ODE, a novel deep learning framework designed to achieve accurate and continuous hourly weather forecasting. The model tackles critical challenges in high-frequency prediction, such as the rapid accumulation of errors in autoregressive rollouts and temporal discontinuities inherent in the ERA5 dataset stemming from its 12-hour assimilation cycle. FlowCast-ODE models atmospheric state evolution as a continuous flow using a two-stage, coarse-to-fine strategy: it first learns dynamics on 6-hour intervals via dynamic flow matching and then refines hourly forecasts using an Ordinary Differential Equation (ODE) solver to maintain temporal coherence. Experiments demonstrate that FlowCast-ODE outperforms strong baselines, achieving lower root mean square error (RMSE) and reducing blurring to better preserve fine-scale spatial details. Furthermore, the model is highly efficient, reducing its size by about 15% using a lightweight low-rank modulation mechanism, and achieves the capability for hourly forecasting that previously required four separate models in approaches like Pangu-Weather.

Welcome to a new episode where we dive into AQUAH, the Automatic Quantification and Unified Agent in Hydrology! This groundbreaking system is the first end-to-end language-based agent specifically designed for hydrologic modeling.

cBottle, developed by NVIDIA, is a generative diffusion-based framework that acts as a generative foundation model for the global atmosphere. It directly tackles the challenge of petabyte-scale climate simulation data, which is currently almost impossible to access and interact with easily due to immense storage and data movement issues1....

This research evaluates the performance of the Aurora weather foundation model by using lightweight decoders to predict hydrological and energy variables not included in its original training. The study highlights that this decoder-based approach significantly reduces training time and memory requirements compared to fine-tuning the entire model, while still achieving strong accuracy. A key finding is that decoder accuracy is influenced by the physical correlation between the new variables and those initially used for pretraining, suggesting that Aurora's latent space effectively captures meaningful physical relationships. The authors argue that the ability to extend foundation models to new variables without full fine-tuning is an important quality metric for Earth sciences, promoting accessibility for communities with limited computational resources. They conclude that rich latent space representations allow for accurate predictions of new variables using lightweight extensions, advocating for future foundation models that encompass a broad range of physical processes.

When we talk about foundation models, what are we talking about? This is a reflection piece on foundation models by drawing an analogy from numerical solutions in fluid dynamics.

IBM recently released the first-of-its-kind geospatial intelligence any-to-any model TerraMind. In this podcast, we feature this new generative model and learn its capability of multi-modality. I believe there is a lot of potential with such a model.

Today, we are featuring a geospatial foundation model Prithvi, produced by NASA and IBM, one of the first foundation model in this space.

This research article introduces HydroGraphNet, a novel physics-informed graph neural network for improved flood forecasting. Traditional hydrodynamic models are computationally expensive, while machine learning alternatives often lack physical accuracy and interpretability. HydroGraphNet integrates the Kolmogorov–Arnold Network (KAN) to enhance model interpretability within an unstructured mesh framework. By embedding mass conservation laws into its training and using a specific architecture, the model achieves more physically consistent and accurate predictions. Validation on real-world flood data demonstrates significant reductions in prediction error and improvements in identifying major flood events compared to standard methods.

We are featuring three papers: