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

Recent research has emphasized the need to adapt life cycle inventories and life cycle impact assessments to account for changes in future scenarios. This is mostly achieved by using Integrated Assessment Models (IAMs), which combine environmental and economic data to develop prospective life cycle inventories (LCIs). However, running complex IAMs to simulate scenarios is computationally expensive, and the resulting data is not always aligned with the geographical scope of background inventories. This study explores the potential of machine learning (ML) to create prospective water-scarcity characterisation factors by forecasting the AWARE factor for ~ 9700 watersheds globally. Historical time series of water-scarcity characterisation factors are generated using the global freshwater model WaterGAP v2.2d and the AWARE method. Several ML models are trained on these historical datasets and benchmarked using symmetric mean absolute percentage error (sMAPE). The best-performing model, N-Beats, is then used to forecast AWARE values through 2032. The results demonstrate that ML can produce spatially and temporally explicit forecasts with reasonable accuracy (median sMAPE ~ 28%). However, the models primarily capture seasonal patterns rather than long-term structural trends and the results are sensitive to the quality and representativeness of the training data. This study highlights both the potential and the limitations of time series forecasting for developing prospective characterisation factors in life cycle assessment. Graphical abstract