The Water Cycle Was Never Four Problems
A single transformer that has never heard of Darcy's law now outperforms the physics models hydrology spent decades assembling.
For most of a century, climate science treated the water cycle as a set of separate problems, and for good reasons. Evapotranspiration, precipitation, soil moisture, and runoff each answer to different forces, get caught by different sensors, and get written down in different equations. So we built different models for each, tuned to its own physics, and stitched the pieces together. ERA5-Land, the reanalysis the field treats as its global benchmark, is exactly this: a multi-stage workflow of physical models, each an expert in its one variable. The implicit contract was that accuracy at global scale required domain knowledge built into the architecture itself.
That decomposition has a price, paid twice. Errors cascade through the hand-offs: a bias in precipitation seeps into soil moisture before it ever reaches runoff. And because these models have to run over the whole planet, they run coarse, smoothing the map until any drought or flood that plays out inside a single watershed becomes invisible. The coupling that matters most, how evapotranspiration draws down soil moisture, how saturated ground sheds runoff, is approximated at the seams rather than learned.
Now consider BERTH, which ignores all of it. It does not know Darcy's law. It has no separate encoder for evapotranspiration, no physics kernel for infiltration. It takes daily satellite radiance, six bands of surface reflectance, a handful of atmospheric variables, the shape of the terrain, and folds them into a single shared 256-dimensional space, then projects all four water-cycle components out at once, at 30-metre resolution. Pre-trained on satellite-derived data and fine-tuned against flux-tower and gauge measurements, it just learns the joint signal.
The results are the uncomfortable part. Against ERA5-Land, BERTH lifts the correlation with reality by about 14 percent and cuts normalized error by about 20 percent, holding across land-cover types and terrain. Against the best standalone remote-sensing products, the specialists, each devoted to a single variable, it is more accurate on evapotranspiration, precipitation, and soil moisture individually. Learning the four together did not blur them toward an average. It sharpened each one. The shared representation was capturing something the separate physical models had been leaving on the table.
One number should slow you down, though: runoff. BERTH has no independent ground truth for it. There is no global network of river gauges to train against, so runoff is pinned only by water-balance closure, the bookkeeping insistence that what comes in must go out. The estimate is physically coherent and completely untested. And that exposes the real question. Whether BERTH has learned the true joint physics of the water cycle, or merely a web of correlations that happen to hold in the satellite era, is something the paper cannot yet settle. The two possibilities look identical right up until the climate hands the model conditions its training never saw, which is precisely the moment, a flood forming in a basin no gauge ever watched, when you would most want to trust the answer.
The biggest unresolved issue is whether BERTH’s impressive skill at reproducing observed water‑cycle components stems from having captured the underlying physics of evapotranspiration, precipitation, soil moisture and runoff, or merely from exploiting statistical patterns that may break down when the model encounters climate conditions outside its training set.