Goals of the project
The aim of this project is to construct a fast and accurate neural-network meta- model for reduced-complexity climate models. These models map greenhouse-gas emission trajectories into global mean temperature outcomes through simplified representations of gas cycles, radiative forcing, and energy balance dynamics. Reduced-complexity or simple climate models are widely used in integrated assessment, climate-risk analysis, and policy evaluation, where they are repeatedly embedded in economic models, optimization routines, and stochastic simulations.
Despite their relative simplicity compared to full Earth system models, the repeated numerical evaluation of such climate models can become a computational bottleneck in applications involving dynamic optimization under uncertainty, Monte Carlo simulations of climate risks, expectation formation and learning, or large-scale sensitivity and scenario analysis. The objective of the project is therefore to replace repeated numerical solves by a fast and differentiable surrogate, enabling analyses that would otherwise be computationally infeasible.
Project domain: Climate economics, Neural Networks
Sponsor: Stress testing chair (Ecole Polytechnique)