Generative AI-Downscaling of Large Ensembles Project Unprecedented Future Droughts
Hamish Lewis, Neelesh Rampal, Peter B. Gibson, Luke J. Harrington, Chiara M. Holgate, Anna Ukkola, Nicola M. Maher · 2025
Read Generative AI-Downscaling of Large Ensembles Project Unprecedented Future Droughts by Hamish Lewis, Neelesh Rampal, Peter B. Gibson, Luke J. Harrington, Chiara M. Holgate, Anna Ukkola, Nicola M. Maher (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-21844.txt.
Introduction
Hamish Lewis and six co-authors project how drought in New Zealand may change as the climate warms. Trustworthy projections need many high-resolution climate runs, which are costly to produce. Their 2025 study uses a new data set: large groups of global climate models, sharpened to fine detail by generative AI. The AI was trained to mimic a physics-based regional model.
Across the model groups, rain becomes more variable, while average rain changes stay uncertain. Droughts grow more intense across most of the country. Changes in drought length and frequency stay hard to pin down in many regions because of natural variation and model disagreement. The big groups also reveal rare, extremely long droughts that small groups miss. Under a high-emissions path, some places could face droughts lasting around thirty months.
Catalog record
| CREATOR | Hamish Lewis, Neelesh Rampal, Peter B. Gibson, Luke J. Harrington, Chiara M. Holgate, Anna Ukkola, Nicola M. Maher |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 57 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-21844
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Generative AI-Downscaling of Large Ensembles Project Unprecedented Future Droughts Hamish Lewis, Neelesh Rampal, Peter B. Gibson, Luke J. Harrington, Chiara M. Holgate, Anna Ukkola, Nicola M. Maher Abstract Abstract Understanding how droughts may change in the future is essential for anticipating and mitigating their adverse impacts. However, robust climate projections require large amounts of high-resolution climate simulations, particularly for assessing extreme events. Here, we use a novel dataset, multiple large-ensembles of Global Climate Models (GCMs), downscaled to 12km using generative AI, to quantify the future risk of meteorological drought across New Zealand. The ensembles consists of 20 GCMs, including two single-model initial condition large ensembles. The AI is trained to emulate a physics-based regional climate model (RCM) used in dynamically downscaling, and adds a similar amount of value as the RCM across precipitation and drought metrics. Marked increases in precipitation variability are found across all ensembles, alongside highly uncertain changes in mean precipitation. Future projections show droughts will become more intense across the majority of the country, however, internal variability and model uncertainty obscure future changes in drought durations and frequency across large portions of the country. This uncertainty is understated using a smaller number of dynamically-downscaled simulations. We find evidence that extreme droughts up to twice as long as those found in smaller ensembles, could occur across the entirety of the country in the current climate, highlighting the value of long-duration downscaled simulations to sample rare events. These extremely long droughts increase in length in many locations under a high emissions SSP3-7.0 scenario giving rise to events around 30 months long in some locations. 1 Introduction Droughts are one of the most impactful natural climate hazards, causing significant excess mortality ( Douris and Kim 2021 ) , economic losses ( Naumann et al. 2021 ; Zaveri et al. 2023 ) , threatening water ( Sousa et al. 2018 ) and food security Vogel et al. 2019 ; Lesk et al. 2016 , and causing lasting environmental damage ( Vicente-Serrano et al. 2020 ) . Anthropogenic climate change is set to increase drought risk across many regions of the globe ( Singh et al. 2022 ; Falster et al. 2024 ; Samaniego et al. 2018 ; van der Wiel et al. 2023 ; Stevenson et al. 2022 ) . However, for some countries such as New Zealand, future drought risks still remain uncertain ( Lewis et al. 2025 ; Gibson et al. 2025 ) . Quantifying future drought risk in New Zealand is particularly difficult due to highly uncertain summertime precipitation projections, with both model uncertainty (uncertainty arising from the use of different climate models) and internal variability ( Deser et al. 2012 ; Deser et al. 2014 , natural fluctuations in the climate system,) obscuring the sign and magnitude of change across CMIP6 Global Climate Models (GCMs) ( Gibson et al. 2024a ) . New Zealand’s complex topography and land-sea contrasts means that GCMs often struggle to simulate both extremes, and finer scale climatic features, ( Rummukainen 2016 ; Gibson et al. 2024b ) . Thus, dynamically-downscaled climate projections are often required to produce actionable insights. Within he latest dynamically-downscaled CMIP6 projections for New Zealand ( Gibson et al. 2024b ) , only across 50-60% of the NZ land surface do five out of the six downscaled models agree on the sign of change of various meteorological drought metrics, while DJF precipitation sign agreement is approximately 30% ( Gibson et al. 2025 ) . Lewis et al. 2025 used a storylines approach to explore the two extremes of future precipitation projections on soil-moisture droughts, using the same downscaled projections as Gibson et al. 2025 . They found that increasing future precipitation alleviated the increased actual evaporative drying of a warmer world, leading to a small increase in overall drought severity. Decreasing future precipitation compounded increased actual evapotranspiration, leading to the average year in the future becoming comparable to the driest years of the current climate, with the worst future years exhibiting unprecedented severity. Hydrological projections of New Zealand river flows have also found similar uncertainty in the sign of change across downscaled models across CMIP5 models ( Mullan et al. 2018 ; Collins et al. 2018 ) . Studies which examine synoptic conditions responsible for meteorological droughts have found that these conditions are projected to become more prevalent under climate change ( Harrington et al. 2016 ; Gibson et al. 2016 ) . These studies highlight that there are competing factors involved in producing future drought outcomes. In the face of considerable uncertainty ( Gibson et al. 2024a ; Gibson et al. 2025 ; Collins et al. 2018 ; Lewis et al. 2025 ) and possibly unprecedented future droughts ( Lewis et al. 2025 ) , there is considerable interest in better understanding the possible future outcomes of drought across New Zealand — outcomes that cannot be fully quantified using a small ensemble of downscaled GCMs. Single Model Initial-Condition Large Ensembles (SMILEs) created by running a climate model multiple times with perturbed initial conditions ( Bengtsson and Hodges 2019 ; Maher et al. 2021 ) , are valuable tools used to quantify both uncertainty in climate projections arising from internal variability ( Deser et al. 2012 ; Deser et al. 2014 ) , and producing more rare events to robustly estimate their changes in frequency and intensity in a warming world ( Suarez-Gutierrez et al. 2018 ; Fischer et al. 2013 ; Haugen et al. 2018 ) . However, the coarse spatial resolution of GCMs means that they struggle to simulate key aspects of New Zealand’s climate ( Rummukainen 2016 ; Gibson et al. 2024b ) . This means that GCM SMILEs are not directly fit for purpose for actionable climate projections for New Zealand at the local scale. Regional Climate Model (RCM) SMILEs on the other hand would be more suitable for local scale projections, however the high computation cost leads to smaller ensemble sizes, with fewer models and scenarios selected ( Leduc et al. 2019 ; von Trentini et al. 2019 ; Aalbers et al. 2018 ) . This high computational cost further exacerbates challenges in producing robust climate projections faced by smaller countries with limited resources, such as New Zealand. Artificial Intelligence (AI) based RCM emulators (AI-RCME), which are are orders of magnitude faster than RCMs themselves, are an emerging solution to tackle these challenges ( Rampal et al. 2024a ; Doury et al. 2023 ; Chadwick et al. 2011 ) . AI-RCME are deep learning models trained to learn the relationship between large-scale GCM meteorological fields (winds, temperature, humidity) to already dynamically-downscaled surface variables ( Rampal et al. 2024a , precipitation, temperature, etc.) . Recent studies have demonstrated that AI-based emulators are capable of capturing observed, historical climate characteristics including means, variability, and extremes, as well as climate change signals for mean precipitation ( Doury et al. 2024 ; Rampal et al. 2025a ) and temperature ( Doury et al. 2023 ; Baño-Medina et al. 2024 ; Rampal et al. 2025b ) . The first use of SMILEs generated using AI-RCME to produce climate projections has recently been published, showing significant internal variability in projections of temperature and precipitation extremes at fine scales across New