Decades of effort into climate modeling have resulted in an achievement of humanity, but also a prisoner of physics. Enormous supercomputers have calculated the laws of air sea ice and land, but at an enormous cost, and with the granularity of their grid constraining their models’ accuracy. But now AI is going to turbo-charge the numbers, and the impact will be far-reaching.
The Old Bargain and Its Limits
Conventional climate and weather models are first principles. They computationally fragment the atmosphere into three-dimensional cells, and then apply fluid-dynamics equations to each, progressing in the time direction. This method is precise but it is hamstrung by a stubborn compromise: the greater the resolution, the more computation is required. A global model that captures thunderstorms in their full complexity is far beyond the reach of most centres.
The implications are pragmatic. Clouds, convection, and turbulence take place at a scale that’s finer than the size of most grid cells, which is to say that modelers use “parameterizations” to build into the model educated guesses for how something should behave that they can’t model directly. The largest factors contributing to uncertainty in a climate prediction are the model’s use of those guesses.
Learning the Atmosphere From Its Own Memory
AI has a different approach. Instead of progressing through a series of physical equations, a machine-learning system simply learns the patterns of the atmosphere based on years’ worth of observed and reanalysis data. After it has been trained, it can produce a forecast in seconds on a single contemporary processor, something that previously took a supercomputer hours.
The results have been astonishing. Google DeepMind’s GraphCast and GenCast, Huawei’s Pangu-Weather, and the European Centre for Medium-Range Weather Forecasts’ (ECMWF) own AI forecast system have met or exceeded the performance of state-of-the-art physics-based models across many standard weather indicators, and used a tiny amount of energy to do it. Speed is not just useful. It’s a scientific tool. With a model that takes seconds, you can run thousands of times for huge ensembles that chart out the entire range of potential futures and quantify extreme events far more accurately than a handful of costly simulations ever could.
Hybrid Models: The Pragmatic Middle Path
I find it interesting that Google’s NeuralGCM, along with projects like Allen Institute’s ACE and the Climate Modeling Alliance are trying to reach this goal while staying physically consistent. Perhaps the most interesting work might not be pure AI at all. Hybrid models preserve a physics-based dynamical core but replace the weakest links cloud and convection schemes with neural networks that are trained on high-resolution simulations.
I think the main difference is that this is not weather forecasting, where Tuesday’s rain is a matter of short-term accuracy. Predicting the climate in 2100 demands that the model act in a way, for data that it has never seen before. That is where data-driven systems become most vulnerable.
The Problem of the Unprecedented
They can have difficulty when faced with new situations. Climate change causes the system to enter areas: higher temperatures, different ocean movement and less ice.
Models based on physics follow rules about energy, mass and movement because that is how they are built. A neural network follows what it has learned and it can end up in situations that do not make sense physically after long periods.
Interpretability and Trust
And then there’s opacity. When a conventional model produces a unexpected outcome, researchers can follow the mathematics and look at the parameters to see what led to the result. Deep neural networks are much more difficult to scrutinise. In a field where models are used to justify building seawalls, set insurance premiums and draft national adaptation plans, it’s not optional to understand why a model has said what it has said. Still, explainable-AI methods are improving and ultimately trust will be established more by open benchmarking, open data and independent validation than by results alone.
Democratizing the Forecast
Perhaps the least heralded yet is equity. State-of-the-art climate models have traditionally been limited to the world’s rich countries with their access to state-of-the-art computers. AI models, which can be run on cheaper hardware after being trained on more powerful equipment, could democratise high-quality regional prediction by empowering the m and the world ‘s most vulnerable regions to characterise the very typhoons, floods, and droughts they are most at risk from.
Downscaling a way of translating coarse projections for the whole globe into local detail at the street level is an area mainly ripe for growth. AI is capable of honing a hundred-kilometer-wide projection to a one-kilometer-level of detail for flood risk, heat stress, and crop stress and at a cost that was unaffordable before.
A Tool, Not an Oracle
This is not a time for machine vs physics framing. The most plausible view is, quite honestly, a hybrid: physics gives the mechanism and the ability to think about the new, with the speed, scale, and ability to model effectively what equations are inherently limited in us and in general. Neither will be enough on its own.
And a sobering caveat. Improved models won’t independently reduce carbon. Climate science has been solid for a while, and computer crunching power isn’t the limiting factor it’s the political and financial fortitude to do so that’s lacking. AI can illuminate what’s on the horizon and where, but it can’t inform us whether to prepare for it.
