Climate models have a blind spot: they're good at predicting what already happened, but shaky on what never did.
A new paper introduces ReRoute, a framework for getting machine learning models to answer counterfactual questions, like what would have happened under different CO2 levels, without training them on expensive controlled simulation data. The usual fix is to generate synthetic data where one variable is cleanly isolated from the rest, but that requires a simulator, computing power, and inherits whatever assumptions that simulator makes. ReRoute instead takes a model already trained on real observed data, locks the variable in question to a reference value, then reintroduces its effects through a known physical pathway while the model's own learned dynamics handle the rest. The researchers backed the method with a formal causal proof, checked with proof-verification software called Lean, and tested it first on a simplified fluid system where the correct answer was already known.
On a state-of-the-art climate emulator, ReRoute cut prediction error by 18.2 to 31.8 percent for severe CO2 scenarios outside the model's training range, without the cost of generating more simulated training data. It also held up on a model trained purely on historical weather records, where no ground-truth counterfactual exists to check against, better preserving the warming trend implied by real observed conditions. That matters because emulators are increasingly used to run fast what-if scenarios instead of full physics simulations, and a model that quietly falls apart outside its training data is worse than one that is simply slow.
The catch: in the one case that matters most, hypothetical climate futures, there is no exact answer to check it against, so for now this is a bet on the math holding, not a receipt from experiment.