Rain is often treated as an outcome of climate: warm water and shifting winds help determine where it falls. In the North Atlantic, it can also be part of the machinery. Fresh water changes the ocean’s surface conditions, while atmospheric processes associated with precipitation redistribute heat. That makes a one-way story—ocean and atmosphere produce rain—potentially incomplete.[2]
A study by Andrew Nicoll and colleagues, published in August and highlighted by Oxford in October, used sparse equation discovery to connect Atlantic temperature variability, the North Atlantic Oscillation pressure pattern and regional precipitation. The chosen model has 26 terms across three coupled equations; rainfall interaction terms recur robustly across candidate models.[1][2]
What it means to discover an equation
Instead of asking a neural network for a forecast and then trying to explain the answer, equation discovery searches for a mathematical description of change. Candidate ingredients can include a variable itself, its square, or its product with another variable. A selection procedure looks for a compact combination that fits the measured or reconstructed dynamics. The resulting expression can be inspected, compared and challenged directly.[3]
A product term has a useful everyday interpretation. In a hypothetical equation, the effect of ocean warmth on an atmospheric variable might depend on whether the region is unusually wet or dry. The same temperature anomaly could then be associated with a different response under a different rainfall state. This is an illustration of an interaction term, not a new causal measurement from the study.
The attraction is not that three equations replace a global climate model. They make a focused relationship discussable. Researchers can ask which interactions remain when the mathematical model is simplified, which disappear, and whether the surviving signs and timescales make physical sense. A compact model earns its place by clarifying that problem, not by describing every process on Earth.[3]
The difficult part is deciding what the data know
Climate data used for fitting and earlier-period tests
| Role | Reanalysis and years |
|---|---|
| Training | ERA5: 1950–2022 |
| Validation | ERA20C: 1940–1949 |
| Held-out test | ERA20C: 1900–1939 |
That distinction is consequential here. The paper notes that ERA5 precipitation was not assimilated before 2010; comparison with GPCP monthly precipitation anomalies gave a correlation of 0.4 over 1982–2020. The rainfall signal is therefore a limitation of the evidence, not just an input label.[1]
The model also recovered a roughly 20-year timescale, interpreted as a damped ocean oscillation influenced by atmospheric forcing. A recurring timescale is not an appointment: it does not imply an Atlantic phase reversal exactly every two decades.[2]
Oxford’s account points to a constructive use for the result. Rainfall may help connect atmospheric heating and ocean freshwater effects within an interpretable description, rather than appearing only at the end of a forecasting chain. Those links supply hypotheses to examine with better observations and more detailed models.[2]
The scientific gain is a smaller object that can be argued with. Where an immense simulation can show a pattern, a short equation can expose the particular interaction proposed to explain it. Here, rain moves from the margin of the Atlantic story into the relationship that the next study has to test.
Sources & further reading
- New insights into decadal climate variability in the North Atlantic revealed by data-driven dynamical models Earth System Dynamics · 7 Aug 2026 · Accessed 17 Oct 2026
- Using AI-discovered equations to study climate variability University of Oxford, Department of Physics · 7 Oct 2026 · Accessed 17 Oct 2026
- Potential for equation discovery with AI in the climate sciences Earth System Dynamics · Accessed 17 Oct 2026