The research, led by Fabien Courvoisier of ETH Zurich together with colleagues at ETH and the Swiss Federal Institute for Forest, Snow and Landscape Research WSL, puts a Long Short-Term Memory network, or LSTM, through one of the most demanding stress tests yet devised for machine learning in hydrology. LSTMs are recurrent neural networks equipped with trainable gates that control how information is stored and discarded across time steps, allowing them to capture long-range dependencies such as the delayed release of water from snowpack or soil moisture storage. Because they learn these relationships directly from data, they can be trained rapidly and applied across hundreds of catchments without laborious site-by-site calibration, which is precisely what makes them attractive, and potentially risky, for climate projections.

The danger is well documented. A purely data-driven model imposes no physical constraints, so when it is forced with inputs outside its training domain, it can behave in physically implausible ways. Previous studies have shown LSTMs occasionally projecting runoff increases under strong warming, in contradiction to evapotranspiration theory, or failing to reproduce extreme floods because the network’s internal gates saturate when rainfall exceeds anything in the training record. Recognizing these pitfalls, the team designed their evaluation to detect exactly such failures. They trained the model on daily runoff observations from 96 minimally impacted Swiss catchments drawn from the CAMELS-CH database, using gridded precipitation and temperature products from MeteoSwiss along with static catchment descriptors covering topography, land cover, soil properties, and basin area.

A crucial innovation was the treatment of glaciers. Rather than assuming a fixed glacier extent, the researchers fed the network a time-varying glacier fraction derived from established glacier evolution simulations based on the GloGEMflow model, applied consistently to both the machine learning model and the process-based benchmark. This matters because glacier retreat is dramatic under the high-emission RCP8.5 scenario used in the study: of the 76 study catchments with more than half a percent glacier cover in the historical baseline, over 80 percent had fallen below that threshold by the end of the century in the median projection. Training used an eight-fold spatial cross-validation scheme combined with a strict temporal split, holding out the final years of the record, so the model’s skill was tested on both unseen basins and unseen time periods.

For the projections, the trained LSTM was driven with the CH2018 climate scenario ensemble, Switzerland’s official set of downscaled and bias-corrected regional climate simulations, spanning 1981 to 2100. Fourteen model chains under RCP8.5 provided the forcing, the strongest signal available for probing robustness. The results were benchmarked against Hydro-CH2018, a national simulation dataset produced with the semi-distributed process-based model PREVAH, which ran on identical climate inputs and glacier projections. This side-by-side design ensures that any divergence between the two models stems from differences in model structure rather than from different assumptions about the climate or the cryosphere.

The headline finding is reassuring for the machine learning camp. The LSTM achieved a median Nash-Sutcliffe efficiency of 0.75 on the held-out test catchments, slightly better than PREVAH’s 0.72 on similar basins, and its projections remained remarkably stable across all fourteen climate chains. Annual correlations between the two models’ runoff simulations reached medians near 0.95 for the 307 projection catchments, with little degradation between the historical period and the end of the century. Both models reproduced the canonical alpine signals: wetter winters as snowfall shifts to rain, pronounced summer drying as meltwater reserves dwindle and evapotranspiration rises, and drying that intensifies with elevation as snow- and glacier-fed regimes lose their seasonal buffers.

The agreement extended to the catchment scale. In six representative basins ranging from the highly glaciated Rosegbach to the lowland rain-fed Venoge, the ensemble-median trajectories of runoff change tracked each other closely, with both models agreeing on the sign of change by 2100 in every regime. The largest discrepancies appeared exactly where hydrology is most complex: glacier-fed and high-alpine snow catchments, where PREVAH generally simulated stronger winter increases and the machine learning model sometimes projected stronger annual drying. The researchers attribute these differences to contrasting representations of snow-to-rain partitioning, melt timing, and storage release, rather than to any fundamental instability in the neural network’s response to out-of-distribution forcing.

Extremes tell a more cautionary story. Comparing how each model translates heavy three-day precipitation into peak flows revealed that PREVAH responds with much steeper runoff-precipitation slopes, while the LSTM systematically dampens both flood peaks and low flows, a regression-to-the-mean behaviour consistent with gate-activation saturation reported in earlier studies. The national maps confirm near-continuous underestimation of annual maxima by the network, a bias that is spatially consistent across the ensemble rather than random. Intriguingly, the LSTM projected less severe annual minima than PREVAH, and since the process-based model is known to sometimes over-dry under high evaporative demand, the machine learning estimate of future low flows may even be the more moderate of the two, though the authors urge caution in that interpretation.

The study does not claim that neural networks should replace physical models, and the authors are explicit that agreement between two structurally different approaches increases confidence in a signal without proving either model correct. What it does demonstrate is that, with sufficiently broad and diverse training data, a data-driven model can generalize learned hydroclimate relationships to a warmer future at national scale, at least for mean flows and seasonal regimes. Combined with the LSTM’s computational efficiency, which makes running large ensembles of climate scenarios almost trivial compared to process-based simulation, the results position these networks as a powerful complementary tool for water resource planning. The remaining frontier is clear: taming the tails of the distribution, where floods live and where current architectures still fall short. Hybrid approaches that embed physical constraints into the learning process, along with transfer learning and richer training datasets, are the most promising routes toward closing that gap.

Subject of Research: Stability of LSTM neural network runoff projections under climate change scenarios in Switzerland

Article Title: Assessing the stability of LSTM runoff projections in Switzerland under climate scenarios

Article References: Courvoisier, F., Kraft, B., Haddad, Y. Y., Zappa, M., & Gudmundsson, L. (2026). Assessing the stability of LSTM runoff projections in Switzerland under climate scenarios. Hydrology and Earth System Sciences, 30(18), 5873-5900. https://doi.org/10.5194/hess-30-5873-2026

Image Credits: AI Generated

DOI: 10.5194/hess-30-5873-2026

Keywords: LSTM, hydrology, runoff projections, climate change, Switzerland, machine learning, glacier retreat, CH2018, PREVAH, alpine catchments, RCP8.5, flood extremes