AI Predicts Water Scarcity for Nearly 10,000 Watersheds Worldwide

Every product we make, from a cotton T-shirt to a computer chip, carries a hidden water bill. Life cycle assessment, the accounting framework that environmental scientists use to tally these hidden costs, relies on characterisation factors: numerical multipliers that translate a cubic meter of water consumed in one place into a measure of genuine scarcity. The problem is that these factors are snapshots of the present, while the technologies we assess today will operate in the water-stressed world of 2030 or 2050. A new study published in the Journal of Industrial Ecology by Niklas Engberg of Delft University of Technology and colleagues takes a strikingly direct approach to this problem, training a deep learning model to forecast water scarcity factors for nearly ten thousand river basins around the globe, and in doing so exposing both the promise and the hard limits of letting algorithms peer into the environmental future.

The conventional route to forward-looking characterisation factors runs through Integrated Assessment Models, or IAMs. These are vast coupled simulations of the economy, energy system, land use and climate that researchers use to sketch coherent scenarios of how the world might evolve. IAMs have powered most prospective life cycle assessments to date, including a well-known 2022 study that generated future water scarcity factors using output from the IMAGE model. But IAMs come with baggage. They are computationally expensive, they bundle similar technologies and regions into coarse aggregate categories, and their deterministic pathways can clash with the fine-grained spatial resolution that life cycle assessment demands. The characterisation factors they produce often do not line up neatly with the geographic units used in background inventory databases such as ecoinvent. Engberg and his collaborators hypothesised that a data-driven alternative, one that learns directly from historical records rather than from chains of socio-economic assumptions, could sidestep some of these frictions.

Their target variable was the AWARE factor, the consensus method adopted by the water-focused life cycle initiative WULCA, which expresses the relative availability of water in a region as the ratio between a consumption-weighted world average of available water and the water available in that specific basin. A high AWARE value means water is scarce; the factor can be read as the surface-time equivalent needed to generate one cubic meter of unused water in that place. The official AWARE dataset covers only a single year, 2010, which is far too thin to train a machine learning model. So the team rebuilt the factor from scratch, month by month, from 1960 to 2016, using the global freshwater model WaterGAP v2.2d. For each of roughly 9,700 watersheds they computed an availability-minus-demand quantity that subtracts human water consumption and environmental flow requirements from natural runoff, then normalised it against the world average. Environmental water requirements were weighted according to the scheme of Pastor and colleagues, protecting low-flow months more stringently than high-flow ones. The same boundary conditions as the original method, capping factors at 100 and flooring them at 0.01, were applied throughout.

Before any forecasting began, the team checked whether their reconstructed historical factors were trustworthy. Compared against the published 2010 AWARE values, their WaterGAP-based calculations showed a correlation of 0.81, high but not perfect, largely because the original method relied on an older, inaccessible version of the hydrological model. More telling was the comparison with the IMAGE-based factors of the earlier IAM study, which correlated at only 0.53 with the original values. Many basins that were genuinely water-rich appeared substantially scarcer in the IAM-derived dataset. This discrepancy matters: it suggests that the choice of underlying model and scenario machinery can shift characterisation factors dramatically, an uncomfortable truth for a field that treats these numbers as authoritative multipliers.

With the historical time series in hand, the researchers turned to model selection. They benchmarked several algorithms, including the gradient boosting methods XGBoost and LightGBM, alongside deep learning architectures designed for time series. The winner was N-Beats, a neural network architecture built from stacks of fully connected blocks that decompose a signal into interpretable trend and seasonality components. N-Beats is what forecasters call a global model: rather than fitting one model per basin, as classical approaches like ARIMA would require across thousands of series, it learns shared temporal patterns across all basins simultaneously. The deployed network used 24 stacks of two blocks each, with 136 units per layer, a learning rate of 0.001 and a batch size of 1024. Training was disciplined with early stopping, which halts the process when validation loss stalls for five consecutive epochs, and a learning rate scheduler that halves the rate whenever performance plateaus. The input window spanned 72 months, six full annual cycles, long enough to capture hydrological seasonality without dragging in outdated patterns that invite overfitting.

Performance was measured with the symmetric mean absolute percentage error, or sMAPE, a metric that penalises over- and under-forecasting equally and ranges from 0 to 200 percent. On the 1,463 largest basins, those spanning at least three half-degree grid cells, the N-Beats model achieved a median sMAPE of about 25 percent, and more than 60 percent of basins came in below 30 percent. Some basins were forecast with remarkable precision, errors near 3 percent, while the worst performers exceeded 50 percent and occasionally approached 78 percent. XGBoost and LightGBM fared worse, with maximum errors above 100 percent and a larger share of poorly predicted basins. On raw accuracy, the deep learning approach clearly outclassed both its statistical ancestors and its tree-based rivals.

But accuracy on monthly values tells only half the story, and it is the half that matters less. When the team smoothed the forecasts into 30-month moving averages to reveal long-term trajectories, cracks appeared. The models tracked seasonal oscillations faithfully, yet their multi-year trends often diverged substantially from the validation data. In the Volga basin, the gap between forecast and reality at the end of the validation period reached roughly 20 cubic meters of world-equivalent water per cubic meter consumed. Even some basins with seemingly excellent sMAPE scores showed poor alignment in their long-term behaviour, a reminder that a model can nail the seasonal rhythm while missing the underlying melody. Attempts to forecast annual values directly produced unsatisfactory results altogether.

The final deployment, trained on the full 1960 to 2016 record, projected AWARE factors out to 2032. According to the forecasts, more than 70 percent of basins will show lower water scarcity in 2030 than in 2010, while roughly 400 basins that already had factors above 60 were predicted to climb toward nearly 100, the ceiling of the scale. Comparing these projections with the IMAGE-based factors for 2030 revealed substantial disagreement in many regions, particularly near the equator, in eastern North America and in western Australia, while South America, Europe and Central Africa showed closer agreement between the two methods. The authors are careful to note that the comparison is imperfect, since the two approaches rest on different hydrological models and the IAM version embeds explicit socio-economic pathways, in this case a middle-of-the-road scenario, whereas the machine learning forecasts simply extrapolate past dynamics with no policy or demographic assumptions baked in.

That absence of assumptions is both the method’s defining feature and its Achilles heel. Time series forecasting, by construction, cannot anticipate structural breaks: a new water management regulation, a leap in irrigation efficiency, a wave of desalination plants, or an unprecedented drought all lie invisible to a model trained only on historical scarcity values. The authors are refreshingly blunt about the implications. Machine learning forecasts, they conclude, are not a standalone alternative to IAMs; they excel at capturing seasonal patterns but cannot reliably predict the long-term structural trends that matter most for prospective life cycle assessment. They also flag a subtle but significant finding for the field at large: the world-average availability term that anchors the AWARE method shifted by about 30 percent between 2010 and 2011, which alone moves every basin’s factor by the same magnitude. Even without any machine learning, that observation argues for treating single-year characterisation factors with caution and for embedding historical trends into impact assessment.

Still, the study opens doors that IAMs cannot reach. The approach could be extended to other characterisation factors where local historical time series exist, such as land use change, and the authors point to covariates from climate model ensembles like ISIMIP and CMIP, variables such as surface temperature, groundwater recharge and leaf area index, as a route to forecasts that respond to climate signals rather than merely repeating them. Transfer learning offers another avenue: a pretrained network could be adapted by practitioners with local data to generate factors for regions the global model never saw. The recently updated AWARE 2.0 method, with improved consumption and environmental flow calculations, may provide cleaner training data for follow-up work. For now, the message is one of calibrated optimism. Deep learning can render the seasonal pulse of the planet’s watersheds with impressive fidelity, and it does so at a spatial resolution that scenario models struggle to match. What it cannot yet do is tell us where the current of change is heading when the change itself has never happened before. Bridging that gap, between pattern recognition and genuine foresight, is the next challenge for anyone hoping to forecast the future state of the environment.

Subject of Research: Machine learning forecasting of water scarcity characterisation factors for prospective life cycle assessment

Article Title: Forecasting future states of the environment with machine learning: a case study on water scarcity

Article References: Engberg, N., Blanco, C. F., Barbarossa, V., Bakker, C., Balkenende, R., Lian, J. Z., & Sprecher, B. (2026). Forecasting future states of the environment with machine learning: a case study on water scarcity. Journal of Industrial Ecology, 30(4), 1285-1298. https://doi.org/10.1007/s44498-026-00010-6

Image Credits: AI Generated

DOI: 10.1007/s44498-026-00010-6

Keywords: machine learning, water scarcity, life cycle assessment, AWARE, time series forecasting, N-Beats, characterisation factors, integrated assessment models, WaterGAP, hydrology, prospective LCA, deep learning