Weather forecasting faces a notorious "blind spot" between two weeks and two months. In the short term (0–10 days), numerical weather prediction (NWP) models excel by solving fluid dynamics equations based on current atmospheric states. In the long term (months to years), climate models provide skill by tracking slow-moving boundary conditions like ocean temperatures.
The window in between—known as subseasonal forecasting—is often called the "predictability desert." In this regime, the chaotic nature of the atmosphere has caused initial conditions to fade, yet long-term climate drivers have not yet become dominant. Recent research is bridging this gap by using machine learning to extract subtle signals from this high-noise environment.
The Predictability Gap
As Jessica Hwang, Paulo Orenstein, and Lester Mackey describe in Improving Subseasonal Forecasting in the Western U.S. with Machine Learning, the subseasonal window (2–6 weeks) is a transitional zone. Traditional physics-based models struggle here because small errors in initial measurements grow exponentially, a phenomenon known as atmospheric chaos.
To overcome this, researchers are shifting from purely deterministic simulations to hybrid and data-driven approaches. The goal is not necessarily to replace physics, but to use machine learning to identify "sources of opportunity"—slow-moving physical processes that provide a memory effect in the climate system.
Learning from Slow Drivers and Teleconnections
The key to subseasonal skill lies in identifying variables that change more slowly than the daily weather. These include:
- Oceanic Indices: Large-scale patterns like the El Niño-Southern Oscillation (ENSO).
- Atmospheric Oscillations: The Madden-Julian Oscillation (MJO), a pulse of clouds and rain that moves around the equator.
- Land Surface States: Soil moisture and snowpack.
Yongkang Xue and colleagues have demonstrated in the LS4P-I project that land surface temperatures, particularly on the Tibetan Plateau, act as a "first-order" driver for global precipitation weeks later. Machine learning models are uniquely suited to capturing these "teleconnections"—long-distance links where a temperature anomaly in one part of the world triggers a weather event in another. For example, Ioannis Prapas developed TeleViT, a transformer-based architecture that uses "asymmetric tokenization" to weigh these global teleconnections against local environmental data to predict wildfires months in advance.
Adaptive Bias Correction
Rather than building models from scratch, a powerful emerging strategy is to "fix" existing physics models. Dynamical models often have systematic biases—they might consistently predict too much rain in a certain valley or fail to capture how a mountain range blocks wind.
Soukayna Mouatadid and Judah Cohen introduced Adaptive Bias Correction (ABC), a framework that uses machine learning to post-process dynamical model outputs. By learning the historical errors of these models, ABC can boost temperature forecasting skill by up to 90%. This hybrid approach combines the structural integrity of physical laws with the statistical precision of deep learning.
Overcoming the Spring Predictability Barrier
One of the most significant hurdles in subseasonal-to-seasonal (S2S) forecasting is the "Spring Predictability Barrier." During the Northern Hemisphere spring, the signals that drive major climate patterns like ENSO tend to weaken, causing forecast accuracy to plummet.
To break this barrier, researchers like Chuang Rui are using temporal embeddings and self-attention mechanisms. By explicitly "teaching" the model about the time of year through embeddings, these neural networks can learn to navigate the periods of instability where traditional models lose their way.
Generative Modeling for Local Precision
Even when global models get the large-scale patterns right, they are often too coarse to be useful for local decision-making. Maximilian Springenberg and Noelia Otero addressed this with DiffScale, a diffusion-based generative model.
Diffusion models, famous for generating realistic images, are used here to "downscale" coarse global forecasts into high-resolution local predictions. Because diffusion models learn the underlying distribution of weather patterns, they can generate physically plausible local details (like wind gusts) that are statistically consistent with the large-scale climate signal, effectively turning a blurry global forecast into a sharp local one.