AI2's coupled climate emulator SamudrACE runs 1,100 years/day on one H100 vs E3SM's 28 on 105 nodes
allen_ai · x · 2026-09-05
AI2 and the E3SM team built SamudrACE-E3SMv3, a fully coupled atmosphere-ocean ML emulator that reproduces the pre-industrial control run of the E3SMv3 Earth system model it learns from.
Dramatic speedup
- E3SM simulates 28 climate years/day on 105 compute nodes;
- the emulator does 1,100 years/day on a single H100 GPU.
How it works
- The atmosphere (ACE) and ocean (Samudra) emulators pretrain separately on "perfect" E3SM inputs, then are joined and fine-tuned together;
- a key design choice is a stochastic atmosphere, whose randomness becomes the coupled system's natural variability;
- against 400 years of held-out E3SM data, temperature and precipitation biases are smaller than E3SM's own gap from real-world observations, and El Niño (ENSO) stays realistic across a 400-year run without drifting or collapsing.
Limitation: it captures daily rainfall up to the 99.99th percentile but underestimates the rarest tropical downpours — learning the shape of a distribution is not the same as reaching its far tail.
Related event: AI2's SamudrACE Simulates 1,100 Years of Climate on One H100(2 posts)→
More from Research
- 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
- Figure AI's humanoid video dataset INDEX growing by 2M clips per week — Distinct-Question-16 · 2026-09-05