Research > Metrics > AIMIP

AIMIP (PROTOTYPE)

The AI Model Intercomparison Project (AIMIP) is an open scientific initiative, primarily led by the Allen Institute for AI (Ai2), that benchmarks AI-based weather and climate models developed by groups such as NVIDIA, Google Research, and academic institutions. This page presents interactive plots based on the analyses described in the article “A PMP-Inspired Evaluation Framework for Assessing Deep-Learning Earth System Models (DL-ESMs)” referred below.

Understanding AIMIP: Traditional climate models use supercomputers to solve fluid-dynamics and thermodynamic equations over extended periods. AI models can emulate aspects of Earth's climate much more rapidly, but their outputs require standardized evaluation to determine whether they faithfully reproduce the historical climate and remain physically realistic. AIMIP Phase 1 provides a common experimental framework that includes:

• Common experimental setup: AI models have been trained with ERA5 historical observations from 1979-2014 and are used to generate global atmospheric simulations over the historical period using prescribed sea-surface temperatures and initial conditions, following the AMIP procedure.
• Standardized outputs: Models predict key atmospheric variables at multiple levels, including temperature, humidity, wind vectors, and precipitation.
• For more information, see https://allenai.org/blog/aimip

Interactive Portrait Plots

• AIMIP RMSE, Normalized by Median


• CMIP and AIMIP RMSE Comparison, Normalized by Median

References

Henn, Brian, et al. "AIMIP Phase 1: systematic evaluations of AI weather and climate models." arXiv preprint arXiv:2605.06944 (2026).

Pallotta, G., Duan, S., Bonfils, C., Lee, J., Goodnight, S., & Ullrich, P. (2026). A PMP-inspired Evaluation Framework for Assessing Deep-Learning Earth System Models. arXiv preprint arXiv:2604.06567.