The study, led by Ayse Ozer together with Afshin Shahbazi and Halit Apaydin, integrates CMIP6-based climate projections under the RCP 4.5 and RCP 8.5 emission scenarios with spatial interpolation techniques and a rich mix of meteorological and land use datasets. The goal was to project streamflow dynamics all the way to 2099, a horizon that stretches far beyond the capabilities of most operational forecasting systems. What sets the work apart is not merely the ambition of the timeline but the sheer scale of the model optimization effort. The team tested approximately 40,000 alternative configurations, spanning hyperparameter adjustments, architectural modifications and training strategies, in a systematic search for the most reliable predictors of future water availability.

The two modeling approaches at the heart of the study represent distinct strands of deep learning. LSTM networks, first introduced by Hochreiter and Schmidhuber in 1997, are recurrent architectures designed to capture long-term temporal dependencies through gated memory cells that regulate what information is retained, updated or discarded at each time step. This makes them naturally suited to hydrological time series, where the flow in a river today reflects rainfall events, snowmelt and soil moisture conditions accumulated over weeks or months. CNNs, by contrast, rose to fame in image recognition, most famously in the ImageNet classification breakthrough by Krizhevsky and colleagues. In hydrological applications, convolutional layers act as feature extractors, sliding filters across input sequences to detect local patterns and multi-scale structure in meteorological variables, effectively treating the input windows as one-dimensional signals rather than raw numbers.

Both architectures were trained to perform multi-step-ahead forecasts, meaning the models do not simply predict tomorrow’s flow from today’s conditions. Instead, they project streamflow across extended future horizons, a task that compounds error accumulation and demands that the networks learn genuine process dynamics rather than short-term autocorrelation. To give the models the richest possible picture of each sub-basin, the researchers enriched their inputs with spatially interpolated climate data and land cover variability, using techniques refined for sparse station networks to translate point measurements into spatially coherent fields. This spatial enrichment proved to be a decisive ingredient: deep learning models fed with interpolated inputs and land-use information significantly outperformed simpler configurations, indicating that hydrological intelligence depends as much on the quality and diversity of the inputs as on the sophistication of the network itself.

The performance results are striking. CNN models generally outperformed their LSTM counterparts in both sub-basins, achieving Kling-Gupta Efficiency (KGE) values between 0.80 and 0.85 during the testing phase. The KGE metric, widely regarded in hydrology as a more informative benchmark than the coefficient of determination, decomposes model skill into correlation, bias and variability components; values above 0.8 are typically interpreted as strong predictive capability. That a convolutional architecture, not classically designed for sequential data, edged out the temporally specialized LSTM underscores a recurring lesson in applied machine learning: with enough properly engineered inputs and careful tuning, feature extraction can matter as much as sequence modeling. The exhaustive configuration search, touching tens of thousands of design permutations, likely played a substantial role in pushing both architectures toward their performance ceilings.

But the study’s most consequential findings concern the future rather than the past. Once the optimized models were coupled with CMIP6-based climate projections under the two RCP pathways, the picture they painted of the Sakarya Basin toward 2100 was grim. In the E12A033 sub-basin, streamflow is projected to decline by approximately 37 percent under the moderate RCP 4.5 scenario and 42 percent under the high-emission RCP 8.5 scenario, compared to the historical baseline. In the E12A053 sub-basin, the decline is far more severe: reductions of 59 percent under RCP 4.5 and 62 percent under RCP 8.5. The divergence between the two sub-basins is itself instructive, showing that climate impacts on water resources will be spatially uneven even within a single river system, and that basin-by-basin assessment, rather than broad regional averages, is essential for planning.

The implications ripple outward from the Sakarya Basin. Türkiye sits in the Mediterranean climate zone, which the Intergovernmental Panel on Climate Change has repeatedly identified as a hotspot where warming and drying will outpace global averages. Declining streamflow of the magnitude projected here would strain drinking water supplies, hydropower generation and the irrigation that underpins agriculture in the region. The research suggests that when water availability in a sub-basin could be nearly halved, or worse, adaptive and data-driven management strategies are not optional refinements but necessities. Accurate long-term forecasts of the kind this framework produces give water managers something they rarely possess: a quantitative, locally calibrated expectation of what is coming, with enough lead time to restructure allocations, invest in efficiency and plan reservoir operations.

Methodologically, the study contributes to a rapidly growing literature pitting data-driven models against process-based hydrological simulators. Traditional physically based models require extensive calibration data and detailed knowledge of catchment characteristics, whereas deep learning approaches learn the rainfall-runoff relationship directly from observations. Recent comparisons in catchments across the United States and China have shown that machine learning techniques can match or exceed process-driven models in simulating streamflow, particularly when trained on rich, well-curated datasets. The Ankara team’s work extends this comparison into the climate-projection domain, demonstrating that deep learning can serve not just as a nowcasting tool but as a vehicle for translating global climate model output into actionable local hydrology through the end of the century.

The researchers are candid about the practical dimensions of their work. The R and Python code underlying the study and the datasets are available on request, and the framework’s reliance on publicly available CMIP6 archives, reanalysis-grade meteorological data and satellite-derived land use products means the approach is portable to other data-scarce regions facing similar climate stress. The combination of spatial interpolation to compensate for sparse station coverage with deep architectures tolerant of heterogeneous inputs offers a template for hydrological forecasting in the many parts of the world where dense monitoring networks remain a luxury.

As global emissions trajectories continue to determine which corner of the uncertainty envelope humanity inhabits, studies of this kind convert abstract climate scenarios into concrete, river-by-river consequences. The message from the Sakarya Basin is unambiguous: even under a moderate emissions pathway, the water flowing through these Turkish rivers by century’s end could be a fraction of what it is today. Whether that projection becomes a planning document or a post-mortem will depend on how quickly water authorities, in Türkiye and across the climate-sensitive Mediterranean, act on the warnings now flowing out of the neural networks.

The innovative framework employs a dynamic gating mechanism that intelligently combines multiple expert models, each specialized in different hydrological patterns, to produce a more precise simulation. Unlike conventional static ensemble models, the dynamic gating system adapts to changing environmental conditions by assigning varying weights to each expert’s output depending on the input data context. This flexibility allows the model to better mimic the complex, non-linear dynamics of natural streamflow.

One of the key technical breakthroughs lies in the framework’s interpretability. In many machine learning applications, predicting streamflow with high accuracy comes at the cost of understandability. The new MoE model incorporates an interpretable gating network that provides clear insights into which expert models dominate under specific hydrological scenarios. This interpretability fosters trust and provides actionable knowledge for hydrologists and policymakers.

In extensive testing across diverse geographic regions and climatic conditions, the dynamic-gated MoE framework consistently outperformed existing state-of-the-art models. It exhibited higher predictive accuracy for daily streamflow fluctuations, including extreme events such as floods and droughts. Such precision is vital for early warning systems and for optimizing reservoir operations in the face of climate variability.

The model integrates seamlessly with existing hydrological data sources and computational infrastructures, making it scalable and practical for real-world deployment. Its ability to process large datasets efficiently stems from its modular design, in which individual expert models handle subsets of data features or temporal patterns.

The research team emphasized the potential of this framework beyond streamflow simulation. Given its modular and adaptive architecture, the dynamic-gated MoE could be extended to other environmental forecasting problems, such as rainfall prediction, sediment transport, or water quality monitoring. Its success demonstrates the transformative power of combining machine learning adaptability with domain-specific hydrological expertise.

This development arrives at a critical time when accurate environmental modeling is necessary for effective climate change mitigation and adaptation strategies. As extreme weather events become more frequent, reliable predictions of water system behaviors are essential for safeguarding communities and ecosystems.

The researchers who led this study expect that their dynamic-gated Mixture-of-Experts framework will inspire further innovations, blending artificial intelligence with earth science disciplines to tackle complex environmental challenges. The future of hydroinformatics looks promising with such tools enhancing our understanding and management of freshwater resources worldwide.

Subject of Research: Improvements in daily streamflow simulation using machine learning frameworks

Article Title: A dynamic-gated Mixture-of-Experts framework improves and interprets daily streamflow simulation

Article References:

Yuan, W., HU, S., Zhan, C. et al. A dynamic-gated Mixture-of-Experts framework improves and interprets daily streamflow simulation.
Commun Earth Environ (2026). https://doi.org/10.1038/s43247-026-03799-z

Image Credits: AI Generated