Climate Surrogates for Scalable Multi-Agent Reinforcement Learning: A Case Study with CICERO-SCM

We achieve a >100x speedup in multi-agent climate policy reinforcement learning (RL) by replacing the CICERO-SCM climate simulator with a high-fidelity surrogate model. This surrogate captures multi-gas climate dynamics with near-simulator accuracy (global-mean temperature RMSE = 0.0004K) while running approximately 1000x faster per simulation step. Bypassing the core computational bottleneck, the surrogate enables regional agents to learn climate policies under multi-gas dynamics in scenarios where the original simulator is intractable. Our approach preserves policy fidelity (converging to the same solutions as the original simulator) and unlocks large-scale multi-agent experiments across alternative climate-policy regimes with high-fidelity climate response.