This paper tests whether choosing better training data can make machine‑learning climate emulators generalize to new situations. The authors show that, for a simple climate model, one carefully optimized emissions scenario can train an emulator that predicts temperature changes better than an emulator trained on six standard ScenarioMIP pathways. In short, the idea is to design the training scenarios themselves, not just the neural network, to improve out‑of‑sample performance.

The team frames training data generation as an optimal experimental design problem and uses a differentiable Simple Climate Model (SCM) based on the FaIR (Finite amplitude Impulse Response) approach. “Differentiable” means they can compute how small changes in the training emissions affect the emulator’s prediction error, using automatic differentiation. They run an iterative four‑step loop: train a base emulator on an initial emissions path, test it on target scenarios, compute the sensitivity of the test loss to the training data, and then update the training emissions by stochastic gradient descent. The emulator they use is a simple multilayer perceptron (a basic neural network) that maps emissions to global mean surface temperature.

The optimized training scenario produced higher predictive skill even though it was smaller than the standard ScenarioMIP training set. The optimized emulator also better reproduced temperature responses when forcings differed structurally from training data, and it could separate the temperature effects of different forcing agents (for example, greenhouse gases versus aerosols) without being trained on single‑forcing experiments. To test whether the idea transfers to more complex models, the authors ran the optimized scenarios through the MITEarthSystemModel (MESM), an Earth system model of intermediate complexity, and found that those outputs yielded a more skillful emulator than using the ScenarioMIP outputs.