AMOC Sentinel

The ocean's conveyor belt is failing. We give scientists and policymakers a 1–5 year AMOC early warning system, built from 160 years of open data, explorable by anyone in minutes.

Comment

AMOC Sentinel

Try the live app

Inspiration

AMOC is the ocean current that keeps Europe warm, stabilises monsoons, and regulates sea levels on the US East Coast. It's slowing down, and nobody has a good early warning system for it.

Direct measurement only started in 2004. Multi-year forecasting still requires supercomputer runs most research groups can't afford. But 160 years of free sea surface temperature data from NOAA has been sitting there the whole time. We figured there had to be a signal in it.

What it does

You drag six sliders (AMO, NAO, ENSO, subpolar SST, subtropical SST, recent AMOC state) and get four forecasts: how strong or weak AMOC will be in 1, 2, 3, and 5 years. Each one comes with a risk tier and the main drivers behind it.

Three historical presets let you explore the full range: 1966 (documented weakening, cold subpolar gyre), 1997 (anomalously strong AMOC), and December 2024 (current conditions). The difference between them is immediate.

How we built it

We used the Caesar et al. (2018) SST fingerprint as our AMOC proxy, which pushes the observational record from 2004 back to 1854.

From there: 116 features per month (six SST regions, four climate indices, AMOC lags up to 48 months), four XGBoost models trained on 1857-2010, tested blind on 2018-2024. Each slider propagates to all its lag columns simultaneously, so the model sees a sustained shift rather than a one-month blip.

$$\hat{y}_{t+h} = f_h(\mathbf{x}_t), \quad h \in {1, 2, 3, 5} \text{ years}$$

All four models beat persistence by 47-55%.

Challenges we ran into

Model conservatism. XGBoost trained on 160 years of history mean-reverts hard. Pushing forecasts into extreme risk tiers requires years of accumulated forcing, not one bad slider reading. That's actually physically correct, but it makes for a less dramatic demo.

Accomplishments that we're proud of

We ran a SHAP attribution: Pearson correlation between monthly SHAP totals and SST anomalies at every Atlantic grid point. The pattern that came back was the Caesar fingerprint, warm subpolar, cool subtropical, the same one published climate scientists derived over decades of research.

The model learned it from data alone, without being told where to look. That felt like the real result.

What we learned

  • 160 years of free SST data has real multi-year AMOC signal. You don't need a supercomputer.
  • Skill vs persistence beats R² for low-variance test periods. Metric choice matters more than we expected.
  • SHAP on climate models is physical validation, not just explainability. It tells you if the model is right for the right reasons.
  • Propagating slider values across all lag columns is what makes a lagged model actually respond to user input.

What's next for AMOC Sentinel

  • Add EN4 ocean heat content and salinity (the strongest sub-surface AMOC predictors, currently missing)
  • Blend in the RAPID array directly rather than relying solely on the SST proxy
  • Automate monthly ERSSTv5 ingestion so the forecast always reflects the latest data
  • Add calibrated quantile forecasts so users see a range, not just a point estimate

Built With

Share this project:

Updates