HELVETIC-AI

The problem

Formal climate detection and attribution compresses high-dimensional fields into scalar indices or a handful of leading patterns, trading spatial resolution for statistical power. That rigorously establishes that anthropogenic forcing is present, but it discards the geography through which the change actually manifests. HELVETIC-AI asks the question the other way round: treat each multi-variable climate state as a spatial image, train a classifier to tell the climate eras apart, then ask the network which regions and fields it used, so the signal comes back as a spatially explicit, physically interpretable fingerprint rather than a single number.

The substrate is what makes the question sharp. The states are drawn from a 5 km ICON-ESM-ER coupled run, the coupled ocean–atmosphere–land–cryosphere output of the EERIE project, currently among the highest-resolution coupled models available on decadal timescales. Three eras are labelled: a pre-industrial-style control, the historical transient (1950–2014), and a near-term future under SSP245 (2015–2050). If forcing has imprinted coherent spatial signatures on the coupled system, a network should be able to separate the eras from structure alone.

Approach

The classifier never sees the native fields directly. Each climate state is reduced to 8 coupled-system fields of anomalies from the control climatology, spanning the whole coupled system: sea-level pressure, near-surface temperature, mid-tropospheric geopotential height and total column water in the atmosphere; sea-surface temperature and sea-ice concentration in the ocean; snowpack and soil moisture on land. Those channels are stacked into an image and regridded onto a common European window, so the network reasons over a coherent multi-variable field rather than any one variable in isolation. An EfficientNet-B0 backbone does the classification. The point of the design is not the backbone but what comes after it: an explainability pass that turns a black-box label into an attribution map.

Results

The classifier reaches 94.7% test accuracy, but the per-class breakdown is the more honest read: control 100% · historical F1 92.2% · future F1 89.4%. The control is trivially separable and the network gets it exactly right; the real test is the historical-versus-future boundary, two adjacent forced states that a scalar index blurs together, and the per-class scores show it holds up there too.

The attribution is the deliverable, not the accuracy. Grad-CAM heat-maps over the classified images are spatially coherent rather than scattered: the network keys on physically meaningful regions, so the era label is grounded in real climate structure rather than a texture artefact. Read forward, that turns the classifier into a data-driven complement to formal attribution. Score an observed pattern under each era and take the likelihood ratio

$$\frac{P(\text{pattern} \mid \text{future})}{P(\text{pattern} \mid \text{historical})},$$

which asks, of a given state, how much more like the future climate it looks than the past.

> ENGINE_ROOM — The explainability pass, and one honest limitation

Grad-CAM is the working attribution method, targeted at the final EfficientNet feature block. It gives spatial saliency, where the network looks, cleanly, and for this question, where the forced signal lives geographically, spatial saliency is most of what is wanted.

Per-channel attribution, which of the eight fields drives the call, is the part that did not come for free. DeepSHAP, the method I first reached for, is incompatible with the EfficientNet backbone: an in-place-operation conflict in the framework breaks the backward pass it relies on. That is reported rather than worked around. The published fingerprints are Grad-CAM’s spatial maps; the per-channel decomposition is left as a named gap, reachable through an integrated-gradients pass or a SHAP-compatible backbone, not asserted as if it were already in hand.

What this demonstrates