High resolution monthly precipitation isotope estimates across Australia from machine learning

Abstract. The stable isotopic composition of precipitation (δ²H_P, δ¹⁸O_P; 'water isotopes') is a powerful tool for tracking water through the atmosphere, as well as fingerprinting land-surface water masses and identifying water cycle biases in isotope-enabled climate models. Water isotopes also underpin our understanding of multi-decadal to multi-centennial water cycle variability via their retrieval from palaeoclimate archives. Water isotopes thereby increase our understanding of past and present – and hence future – water cycle variability. Understanding the drivers of spatial and temporal water isotope variability is a critical first step in applying these tracers for a better understanding of the water cycle. However, water isotope observations are sparse in both space and time. Here we develop and apply a machine learning (random forest) approach to predict spatially continuous monthly δ²H_P and δ¹⁸O_P across the Australian continent at 0.25° resolution from 1962–2023. We train the random forest models on monthly δ²H_P (n = 5199) and δ¹⁸O_P (n = 5217) observations from 60 sites across Australia. We also predict the deuterium excess of precipitation (dxs_P, defined as δ²H_P − 8*δ¹⁸O_P). Out-of-sample δ²H_P and δ¹⁸O_P prediction skill is high both geographically and temporally. Skill is slightly lower for the secondary parameter dxs_P, likely reflecting the larger reliance of spatio-temporal dxs_P variability on moisture source conditions. The random forest models accurately capture both the seasonal cycle of precipitation isotopic variability and long-term annual-mean precipitation isotopic variability across the continent, and outperform estimates from an isotope-enabled atmosphere general circulation model over an equivalent time period. We show that spatio-temporal variability in precipitation amount, precipitation intensity, and surface temperature are particularly important for monthly δ²H_P and δ¹⁸O_P variations across the continent, with local surface pressure also important for dxs_P. Drivers of site-level δ²H_P, δ¹⁸O_P, and dxs_P are more varied. Overall, the new random forest modelled dataset reveals clear spatial and temporal variability in δ²H_P, δ¹⁸O_P, and dxs_P across the Australian continent over the past decades – providing a robust foundation for hydrology, ecology, and palaeoclimate research, as well as an accessible framework for predicting water isotope values in other locations.