Key Takeaways
- • Professor Philip Stier of the University of Oxford, Director of the Intelligent Earth Centre, discussed AI-driven climate modelling with Man Group's Jason Mitchell on the podcast "A Sustainable Future" (recorded 5 October 2026).
- • Stier said climate models changed only gradually for decades, but high-resolution physical simulations and AI have now changed the field almost simultaneously, and the future shape of climate models in ten years is "totally up in the air."
- • Clouds cool the Earth by roughly 20 watts per square metre on net, about ten times the anthropogenic effect of CO2, so small changes in cloud behaviour could substantially offset or amplify warming.
- • Aerosols from air pollution seed cloud droplets; Stier said the resulting cooling from anthropogenic aerosols is about half a degree, and removing that pollution would push warming close to 2 degrees.
- • Mitchell framed the investment relevance around pricing extreme weather risk now, citing insurer premium inflation and coverage withdrawals, rather than only long-horizon climate scenarios.
Summary
In a podcast episode recorded on 5 October 2026, Jason Mitchell, CIO for Responsible Investment at Man Group, spoke with Professor Philip Stier of the University of Oxford about how artificial intelligence could change the understanding of climate risk. Stier is Professor of Atmospheric Physics and Director of the Intelligent Earth Centre at Oxford, and also heads the Climate Processes Group and serves on the steering group of the Oxford Climate Research Network.
Mitchell framed the financial relevance around timing. Traditional physics-based climate models are well suited to out-of-distribution extremes, events whose magnitude has not been observed, while AI is strong at in-distribution extremes grounded in historical data and reanalysis. Sustainable investors often model portfolio impacts 20 to 70 years out, a process that can become entangled in discount-rate debates. Mitchell argued that insurers already illustrate the near-term pricing problem, pointing to premium inflation and coverage withdrawals after extreme weather events.
Stier said climate models evolved steadily over the past 30 to 40 years, but two developments have now coincided: high-resolution simulations that explicitly resolve processes previously parameterised, and AI transforming weather modelling, Earth-observation analysis and the emergence of hybrid physics-AI climate models. He declined to predict what models will look like in a decade.
On the scientific stakes, Stier explained that clouds regulate Earth's radiation balance. They reflect sunlight and absorb outgoing infrared, with a net cooling effect of around 20 watts per square metre, roughly ten times the anthropogenic effect of CO2. Aerosols from air pollution act as cloud condensation nuclei, and more droplets make clouds more reflective. Stier estimated that anthropogenic aerosol cooling is about half a degree and that removing air pollution would take warming close to 2 degrees, beyond the 1.5-degree Paris threshold he said has already been reached.
Stier also described why small-scale cloud processes are hard to model: they span roughly 14 orders of magnitude, from nanometre-scale aerosol formation to global circulation. Because this cannot be fully resolved computationally, models rely on parameterisations with inherent uncertainty. Stier said AI may help improve these parameterisations and that constraining them with observations is an emerging frontier. In discussing the boundary between learning from data and enforcing physical laws, he noted that any climate model must obey physical laws, with the difference lying in whether this is enforced by construction through numerical solutions of physical equations or whether a model trained on sufficient observational data learns the physics itself.
Mitchell framed the financial relevance around timing. Traditional physics-based climate models are well suited to out-of-distribution extremes, events whose magnitude has not been observed, while AI is strong at in-distribution extremes grounded in historical data and reanalysis. Sustainable investors often model portfolio impacts 20 to 70 years out, a process that can become entangled in discount-rate debates. Mitchell argued that insurers already illustrate the near-term pricing problem, pointing to premium inflation and coverage withdrawals after extreme weather events.
Stier said climate models evolved steadily over the past 30 to 40 years, but two developments have now coincided: high-resolution simulations that explicitly resolve processes previously parameterised, and AI transforming weather modelling, Earth-observation analysis and the emergence of hybrid physics-AI climate models. He declined to predict what models will look like in a decade.
On the scientific stakes, Stier explained that clouds regulate Earth's radiation balance. They reflect sunlight and absorb outgoing infrared, with a net cooling effect of around 20 watts per square metre, roughly ten times the anthropogenic effect of CO2. Aerosols from air pollution act as cloud condensation nuclei, and more droplets make clouds more reflective. Stier estimated that anthropogenic aerosol cooling is about half a degree and that removing air pollution would take warming close to 2 degrees, beyond the 1.5-degree Paris threshold he said has already been reached.
Stier also described why small-scale cloud processes are hard to model: they span roughly 14 orders of magnitude, from nanometre-scale aerosol formation to global circulation. Because this cannot be fully resolved computationally, models rely on parameterisations with inherent uncertainty. Stier said AI may help improve these parameterisations and that constraining them with observations is an emerging frontier. In discussing the boundary between learning from data and enforcing physical laws, he noted that any climate model must obey physical laws, with the difference lying in whether this is enforced by construction through numerical solutions of physical equations or whether a model trained on sufficient observational data learns the physics itself.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.