Together, AI and human knowledge make better weather forecasts
Together, AI and human knowledge make better weather forecasts
Weather forecasts are complicated by the unpredictability and complexity of natural systems, and scientists' incomplete understanding of the dynamics at play in the Earth-ocean-atmosphere system. Recent CIRES-led work used a new machine learning technique — cross-attractor transforms (CATs) — to reduce forecast errors in Numerical Weather Prediction models.
The work suggests that "forecasts from this hybrid machine-learning approach are more accurate than purely data-driven methods when applied to an idealized system."
Read the paper in Geophysical Research Letters. And read the commentary about the work.
Schematic of a feed-forward neural net used to train pre- and post-processing transforms Trm and Tmr using a custom layer in the middle, representing imperfect model forecasts. The trainable nodes and weights are color-coded in red, whereas the non-trainable ones are in green. Hidden layers on each side of a fixed custom forecast layer are trained using a catalog of time-evolving reference system states indexed by + and -. The size of inputs, outputs, and the number of neurons and hidden layers are varied according to the dimensionality of the problem. Credit: Agarwal et al., 2025.