AI climate emulator holds steady over 400 years, beating E3SM's own bias
allen_ai · x · 2026-09-05
Allen AI reports strong generalization results for its climate emulator: against 400 years of E3SM data it never trained on, the emulator's average climate holds up, with temperature and precipitation biases smaller than E3SM's own gap from real-world observations.
Key details:
- Training is staged: the atmosphere emulator (ACE) and ocean emulator (Samudra) pretrain separately on "perfect" inputs from E3SM data, then are joined and fine-tuned together
- A key design choice is a stochastic atmosphere — its randomness becomes the coupled system's natural variability
- El Niño (ENSO) stays realistic across the entire 400-year run without drifting or collapsing
More from Research
- Claude's Fermat Proof Passes Independent Rust Verifier: 1,052,234 Declarations, Zero Errors — imjustnewatai · 2026-09-05
- AI-enhanced adaptive virtual screening over billions of molecules lands in Nature Biotech — anshulkundaje · 2026-09-05
- F. Chollet: all AI will converge to symbolic learning as the optimally efficient form — burny_tech · 2026-09-05
- Anthropic model formalizes Fermat's Last Theorem in Lean, 13.4M lines closing 100-theorem benchmark — littmath · 2026-09-05
- GPT-6 Astra hits record 169 on Epoch AI's ECI, sweeping math and continual-learning benchmarks — rohanpaul_ai · 2026-09-05
- SemiAnalysis: Zhipu experiments with Loop Transformer as RL environment building becomes the new bottleneck — ricklamers · 2026-09-05