Physics-Informed AI Tackles Drought Forecasting in a Stressed Transboundary Basin

In one of the most water-stressed corners of Asia, where the Helmand River flows out of Afghanistan’s Hindu Kush highlands toward the vanished wetlands of the Sistan region on the Iranian border, a new study has tested whether the latest generation of artificial intelligence can see drought coming before it devastates farms and wetlands. The research, published in Earth Science Informatics, introduces a forecasting framework called PhysicsSolver, a Transformer-enhanced physics-informed neural network, and pits it against a well-established statistical machine learning approach known as Support Vector Regression combined with the Response Surface Method, or SVR-RSM. The target of both models is hydrological drought, the slow-motion crisis that unfolds not when rain fails but when rivers and reservoirs run low, and which is notoriously difficult to predict in basins where human decisions, upstream dams, and shifting climate patterns scramble the historical record.

The study’s setting could hardly be more consequential. The Zabol Basin sits at the downstream end of the transboundary Helmand River system, a landscape where decades of drought, upstream water diversion, and geopolitical tension have combined to drain the once-vast Hamun Lakes into salt flats. Communities in Iran’s Sistan and Baluchestan province depend on the timing and volume of Helmand flows for agriculture, drinking water, and protection from the region’s infamous dust storms. In such a basin, a reliable drought forecast is not an academic luxury; it is the difference between managed adaptation and humanitarian emergency. Yet forecasting here is uniquely hard because the river’s behavior is shaped by two countries, multiple dams, irrigation withdrawals, and a climate that is itself changing, all of which break the assumption that the past is a reliable guide to the future.

To quantify drought, the study relied on standardized indices computed from more than five decades of Helmand River streamflow data spanning 1961 to 2014. Three indices took center stage: the Standardized Runoff Index (SRI) and the Standardized Streamflow Index (SSI), both of which measure how far current water availability deviates from long-term norms, and a more ambitious third option, the Non-Stationary Standardized Streamflow Index (NSSI), which attempts to account for the fact that the statistical baseline itself shifts over time as human and climatic pressures reshape the river. The models were tasked with forecasting these indices at four time horizons: 1, 3, 6, and 12 months ahead. The inputs to the models were moving averages of streamflow and runoff, a technique that smooths out daily noise and lets the algorithms focus on the persistent signals that carry drought information across seasons.

The headline result is encouraging for seasonal water managers. For the stationary indices, SRI and SSI, both models performed remarkably well, with correlation coefficients between 0.95 and 1.00 and Nash-Sutcliffe Efficiency values, the standard hydrological yardstick that compares model predictions to a simple average-based baseline, ranging from 0.88 to 0.99. The sweet spot for both approaches was the 6- and 12-month scales, precisely the horizons at which seasonal drought monitoring is most useful for planning reservoir releases, crop choices, and emergency water allocations. In other words, when the underlying drought signal follows relatively stable statistical patterns, modern machine learning, whether built on support vector mathematics or on attention-based Transformer architectures, can capture it with near-perfect fidelity. PhysicsSolver edged out SVR-RSM in most comparisons, but the margins were modest rather than transformative.

The real story, and the scientifically provocative one, lies in what happened when the models confronted the non-stationary NSSI. Here the tidy agreement collapsed. At the 1-month horizon, PhysicsSolver demonstrated a clear advantage, achieving a Nash-Sutcliffe Efficiency of 0.98 compared with 0.88 for SVR-RSM, suggesting that the physics-informed Transformer’s ability to encode physical constraints and attend to long-range temporal dependencies gives it genuine power for very short-term prediction even when the data-generating process is shifting underfoot. But at 3- and 6-month horizons, both models failed outright, producing negative efficiency values, which in hydrological practice means the forecasts were worse than simply guessing the historical mean. The finding is a sobering reality check for a field that has grown accustomed to celebratory performance metrics.

Why does non-stationarity break medium-term forecasting so completely? The answer lies in what the NSSI is trying to represent. A stationary index assumes that the probability distribution of streamflow is fixed, so a drought is simply an unusually low draw from a known deck of cards. The non-stationary index acknowledges that the deck itself is being reshuffled by upstream dam operations, changing irrigation demand, land-use shifts, and evolving climate patterns. When a model trained on historical data tries to forecast several months ahead, it must implicitly extrapolate how those human and climatic drivers will evolve, and neither a support vector machine nor a physics-informed Transformer, however sophisticated, can conjure information about future dam releases or geopolitical water-sharing decisions that is not present in the training data. The physics constraints embedded in PhysicsSolver help it stay physically plausible, but they cannot substitute for knowledge of anthropogenic forcing.

The architecture behind PhysicsSolver deserves attention because it represents a broader movement in the geosciences. Physics-informed neural networks embed physical laws, such as mass conservation or flow equations, directly into the training objective, penalizing solutions that fit the data but violate known physics. The Transformer component, borrowed from the deep learning revolution in language modeling, uses attention mechanisms to weigh which parts of the historical record matter most for a given prediction, allowing the model to capture long-range temporal dependencies that older recurrent architectures struggle with. The concept was originally developed for solving and forecasting partial differential equations, and its adaptation to drought indices is part of a wave of hybrid approaches seeking to combine the flexibility of data-driven learning with the reliability of physical understanding, particularly valuable in data-scarce regions where pure machine learning risks learning spurious correlations.

For the Zabol Basin and the millions who depend on the Helmand, the practical implications are twofold. First, the study validates a workable toolkit for seasonal drought monitoring: agencies can use either model, at 6- to 12-month scales, to anticipate drought conditions with high confidence, providing lead time for water rationing, crop switching, and international coordination. Second, and more soberly, the study shows that medium-term forecasting of drought in human-dominated basins remains an open problem that no amount of architectural cleverness alone can solve. The authors’ conclusion is explicit: substantial methodological advances are needed before non-stationary drought can be forecast reliably at the 3- to 6-month horizons where early warning would matter most. That likely means incorporating covariates that explicitly represent anthropogenic pressures, such as upstream reservoir storage, irrigation withdrawals, and climate oscillation indices, rather than expecting streamflow history alone to carry the signal.

The broader lesson resonates far beyond the Iran-Afghanistan border. Transboundary basins cover nearly half of the world’s land surface and supply water to some two billion people, and many of them, like the Helmand, are experiencing exactly the kind of compound human-climate stress that renders historical statistics unreliable. As climate change accelerates and water infrastructure multiplies, the assumption of stationarity that underpins much of hydrology is eroding everywhere. Studies like this one perform a valuable service by mapping, with honest numbers, where the current generation of AI tools succeeds and where it hits a wall. PhysicsSolver’s near-perfect short-term performance on non-stationary indices hints that physics-informed architectures are the right direction of travel; its equally dramatic failure at medium horizons tells researchers precisely where the next breakthrough must come from. In the arid lands of Sistan, where the Hamun wetlands have already paid the price of unforecastable drought, that breakthrough cannot arrive soon enough.

Subject of Research: Physics-informed machine learning for hydrological drought forecasting in the transboundary Zabol Basin

Article Title: PhysicsSolver: A physics-informed transformer for hydrological drought forecasting in the transboundary Zabol Basin

Article References: Piri, J. (2026). PhysicsSolver: A physics-informed transformer for hydrological drought forecasting in the transboundary Zabol Basin. Earth Science Informatics, 19(10), Article 167. https://doi.org/10.1007/s12145-026-02214-7

Image Credits: AI Generated

DOI: 10.1007/s12145-026-02214-7

Keywords: hydrological drought, physics-informed neural network, Transformer, Zabol Basin, Helmand River, transboundary water, non-stationarity, drought indices, machine learning, streamflow forecasting, Sistan, water resource management