I recently had the opportunity to attend the 2026 CMIP Community Workshop, an event focused on the future of global climate modeling. While a number of interesting and useful topics were covered throughout the week, I want to highlight the Advanced Diagnostics, Emulators, and Machine Learning for CMIP7 session. Speakers presented computational advancements across many areas, ranging from high-performance computing (HPC) systems to marine heatwaves, and showcased a variety of novel emulation and machine learning techniques. This side session demonstrated a clear need for computational efficiency, improvements to model evaluation, and rapid, impact-driven climate emulation as we approach CMIP7. Integrating these tools and machine learning frameworks will be crucial for delivering accurate, timely, and policy-relevant climate projections.
A huge thanks to the organizers for facilitating such a forward-thinking opportunity, and to the CMIP-IPO for their support that enabled me to attend the workshop!
Ocean modeling & diagnostics
Lessons from CMIP6 are highly relevant as we prepare for CMIP7. For example, simulations of sea surface temperatures in CMIP6 models are often too smooth. Applying continuous wavelet bias correction to these models corrects this bias, yielding simulated marine heatwaves that actually match the scale and duration of observational data. Similarly, the Ocean Model Intercomparison Project (OMIP) is leveraging CMIP6 experience to build a streamlined diagnostic framework for CMIP7 to support rapid evaluation needs, emulator development, and regional/polar process studies.
Emulating the global climate system
Climate emulators that target globally-averaged statistics, such as the carbon cycle and global mean surface temperature, are critical for assessing large-scale outcomes associated with potential mitigation pathways. The Minimal CMIP Emulator (MCE) provides an efficient tool for probabilistic climate and carbon-cycle projections, though modeling the high variability of land-use CO2 emissions remains a key challenge. Earth System Model (ESM) responses from the CMIP archive are actively used to calibrate the responses of climate emulators like the Finite-amplitude Impulse Response (FaIR) model. This process was central to the IPCC AR6 and will continue into AR7.
Local climate emulation
Modern statistical and machine learning methods have also enabled rapid emulation of changes in local climate fields based on potential policy decisions. Instead of solely focusing on improving predictive skill, new research emphasizes the importance of the training datasets themselves in ensuring that emulators capture a wide range of climate dynamics. This suggests that modeling centers could dedicate resources to running scenarios tailored for emulator training.
New tools like PRIME (land impacts) and ProFSea (sea level rise) are providing a first look at CMIP7 scenarios, helping researchers generate probabilistic ensembles to assess tail risks ahead of the full CMIP7 ESM releases. Additionally, the spatial emulator MESMER has been consolidated into an accessible Python package, enabling researchers to quickly assess natural variability, extreme precipitation, and climate overshoot scenarios. Machine learning also continues to show promise in this domain; for example, a new UNet emulator developed for Antarctica can accurately predict a century of daily surface mass balances after being trained on just twenty years of data.
Model evaluation
Despite the numerous benefits of the CMIP effort, our community still needs to be cognizant of the energy usage it demands. To address this, a new task team is identifying power consumption and HPC bottlenecks for CMIP7, creating a community database to benchmark resolution, core-hours, energy use, and estimated carbon footprints across platforms. Complementary efforts like ClimateBench2.0 will aid in model evaluation more generally. This benchmark applies across model types, from physical to hybrid to ML-based, and introduces a new scoring system that evaluates models not just on statistical fit, but on their physical consistency. Finally, the newly launched AI Model Intercomparison Project (AIMIP) is proving that AI-driven models can generate state-of-the-art forecasts at a fraction of traditional computing costs. The long-term goal of AIMIP is to support AI-augmented climate models as a part of traditional CMIP assessments.
Looking ahead to CMIP7
As we continue to improve our models and diagnostics for CMIP7, the insights from this session illustrate several promising, high-impact advancements. The rapidly expanding role of machine learning and emulation enables us to generate large, impact-relevant ensembles that were previously too computationally expensive. However, to fully leverage these techniques, we must proactively design and dedicate resources to new scenarios explicitly tailored to train these machine-learned models. As our modeling landscape diversifies to include AI-driven and hybrid systems, new, robust benchmarks will be essential. Initiatives like ClimateBench2.0 and AIMIP are critical to ensuring that both Earth System Models and emulators are held to high standards of physical consistency and historical accuracy. Embracing these complementary strategies will ensure that CMIP7 delivers the most reliable, actionable climate projections yet.
