AI will not replace climate science — but it can make it fit for an age of extremes

As we approach COP30, in Brazil, ‘climate AI’ will be a buzzword in discussions on how to slow the advance of global warming. In this race against time, it is tempting to think of climate AI as a competition: can machine learning outperform traditional climate models?

This is the wrong contest: physics-based models remain the only way to explore futures we have never observed and guarantee that basic laws—mass, momentum, energy—are respected. AI, in turn, excels at finding patterns and links between different variables, accelerating computation, and turning torrents of observations into usable signals. The prize is combining the two so society gets understandable, reliable answers to critical questions.

Many popular machine-learning approaches assume tomorrow’s relationships will resemble those of yesterday. In a warming climate, that assumption fails, especially at the margins where we find extreme events. Flood-producing storms, multi-day heatwaves, and compound events are rare by definition; data are sparse; and rules of the game are shifting as oceans warm and circulation patterns adjust. Training our tools on yesterday’s averages will leave them wrong about tomorrow’s extremes. We must embed climate context into how AI learns and let physics set the guardrails.

What a good climate–AI partnership looks like

First, use AI to focus on extremes. Rather than running a model for decades and hoping rare events show up, guide simulations toward ‘interesting bits’ while keeping statistics honest. Target significant but rare events atmospheric-river landfalls, blocking highs, rain-on-snow patterns, then reweight results so return levels and exceedance probabilities remain unbiased. The goal is credible data tails tempered with uncertainty, rather than a pretty animation.

Second, learn patterns with the right context - for example present climate vs. a +1.5 °C world, El Niño vs. La Niña, blocked vs. zonal flow. When training and evaluating AI, check calibration with physics constraints within each context – a simple discipline turning a fragile model into one that travels better as the climate changes.

Third, turn attribution into a test: If an algorithm claims greenhouse gases or a particular sea-surface pattern increased risk by 30%, run a controlled ‘what-if’ world in a digitally twinned physics model, dialling the driver up or down, and see if the change matches.

Fourth, design for explanation from the start. Black-box shortcuts are not enough for public-interest decisions. If we change a driver, the explanation should predict the sign and size of the change. We should publish simple scorecards that models must pass before they are trusted.

Digital twins as the meeting place

'Digital twin’ is both easy to hype and misunderstand. At heart, a climate twin is a continuously updated ‘virtual Earth’ fusing observations, physics, and AI to run scientifically consistent ‘what-if’ worlds - and trace the consequences from atmosphere to rivers to people.

Two principles make twins especially valuable for extremes:

• State, not scenario. Instead of arguing about which emissions pathway gets us to +2 °C in which year, use warming-level states (present; +1.5/+2/+3 °C). Within each, sampling realistic variations of extreme events, and capture risk with uncertainty. Policymakers can act on this today.
• From events to impacts. Twins should not stop at rainfall - the question is not only ‘How much rain?’, but ‘How many people and which assets are at risk, with what inequities, and which interventions change that?’.

Trust in climate analytics will not be earned with glossy dashboards but with auditability – deploying robust tools that validate and record data, context, and metrics.

Why this matters now

Cities are facing bigger floods, hotter summers, and compounding stresses. Decision makers do not need another benchmark or leader board, but clear, comparable outputs. Done well, AI + physics can deliver these on timelines that match budgets and election cycles. They need a practical checklist for agencies and partners:

• Do the methods provide credible estimates of rare but damaging events?
• Are results calibrated within warming levels and teleconnection phases?
• Are attribution claims validated against controlled ‘what-if, worlds in a physics model?
• Do surrogates behave as expected and do explanations predict outcomes?
• Can you audit what was run, with what data, and what passed/failed?

Collaboration not competition

At the University of Birmingham, these ideas are already shaping research. Our scientists use AI to analyse water samples from over 50 UK lakes, uncovering how pollutants and climate change together drive biodiversity loss. We use AI to improve weather forecasts over India, and to build emulators for regional climate models over the Himalayas and Antarctica. Another project combines AI with hydrological models to improve predictions of how rivers respond to heavy rainfall and drought. Policy experts are exploring how AI and conflict data can be combined to model climate security risks - enabling decision-makers to plan and respond more effectively in times of crisis. These studies show how AI can uncover hidden connections and turn huge datasets into meaningful insights. Our teams are developing AI tools that respect physical laws, testing them against established models, and training new researchers who can bridge climate science and computer science. The result is smarter, faster, and more reliable ways to understand and prepare for a changing world.

AI will not replace climate science, and we should not want it to. Combined carefully with digital twins for counterfactuals, AI for data tail coverage and pattern learning, physics for constraints and credibility, and transparent audits for trust, it can make climate information fit for decisions in an age defined by extremes. That is a design choice that is socially responsible and politically actionable.

Professor Russell Beale, Dr Ruth Geen, Professor Gregor Leckebusch, Dr Martin Waehlisch, and Dr Martin Widmann.

Supporting contributors from wider CLIM(AI)TE group: Dr Justyna Bandola-Gill, Professor Slava Jankin, Dr Jianbo Jiao, Professor Ales Leonardis, Dr Leandro Minku, Dr Paolo Missier, and Dr Xiaocheng Shang.

Discover more about the work of Birmingham Institute of Sustainability and Climate Action (BISCA) and Institute for Data and AI (IDAI).