How Do Climate Scientists Use Artificial Intelligence?
Last week's conversations at a National Academies workshop centered on a simple prompt with big implications: how do we use AI for climate action? Before we get to action, it helps to be precise about how and why AI already shows up in climate research.
Short answer: AI is one more tool in the kit. It complements, not replaces, physics-based climate models.
AI is another tool, not a replacement
Climate models are built on equations that represent the physics of the atmosphere, ocean, ice, land, and vegetation. AI, mostly machine learning, learns statistical patterns from data and makes predictions from those patterns.
That distinction matters. AI can learn that a warmer eastern Pacific usually coincides with a warmer Pacific Northwest during El Niño, but it isn't "following" the physics as a model does. Its predictions are pattern-based, which means they can be insightful and fast, but technically un-physical.
If you want a refresher on El Niño's global fingerprints, see NOAA's overview of ENSO teleconnections here.
Where AI fits in current research
- Climate emulators: AI systems that learn the input-output behavior of a climate model and reproduce its responses far faster. The AI2 Climate Emulator (ACE) can run about 100x faster and about 1000x more energy-efficient than a traditional model, while closely matching its outputs.
- Downscaling: Climate models operate on coarse grids that miss local detail. AI learns relationships between large-scale patterns and local observations to estimate city-scale changes in temperature or precipitation instead of broad regional averages.
- Hybrid modeling: Researchers combine AI components with physical models to correct biases, accelerate components, or infer sub-grid processes.
- Efficiency: The AI used in climate research is typically compact compared with large language models. Most projects don't require massive data centers.
Case study: filling data gaps for attribution
Attribution science needs long, reliable records to estimate how much climate change altered the odds or intensity of an extreme event. Many regions in the Global South lack complete historical data, which weakens studies where the need is greatest.
Machine learning can help close those gaps. By training on daily temperature from multiple high-resolution climate models to learn the structure of extreme heat, then predicting values at missing locations, researchers can reconstruct plausible records and compare them against independent observational datasets. Better records strengthen attribution, climate litigation, and Loss and Damage assessments.
From science to action: promising uses already in play
- Predict the final size of a wildfire from ignition and detect fires early from imagery and sensors.
- Advise sustainable farming practices and guide crop selection under shifting climate conditions.
- Recommend urban tree planting for heat mitigation and equity.
- Support water managers working to keep major aquifers within sustainable limits.
For context on cross-sector efforts, see the National Academies' Roundtable on AI and Climate Change overview.
What to watch out for
- Distribution shift: AI learns from past data. Future climate states may sit outside that experience. Purely data-driven projections can struggle if the system moves into new regimes.
- Black-box behavior: Many ML models don't explain themselves. Use explainable AI to probe what signals models rely on and whether those signals make scientific sense.
- Trust and rigor: Publish methods, make code and data accessible when possible, use fixed external benchmarks, and document uncertainty and training data quality. Reproducibility builds confidence.
Practical steps for research teams
- Start with problems dominated by pattern recognition: downscaling, bias correction, emulation, quality control, event detection.
- Pair AI with physics: add constraints, use hybrid loss functions, or post-filter with physical checks.
- Treat data like code: version datasets, track provenance, and record preprocessing choices.
- Interrogate models: run sensitivity tests, apply feature attribution, and stress-test on out-of-sample periods and regions.
- Quantify uncertainty: use ensembles, calibration, and coverage metrics that matter for decisions.
- Mind compute and energy: prefer compact architectures and mixed precision; benchmark runtime and energy usage.
- Share openly: pre-register evaluation plans, release model cards, and invite independent replication.
Bottom line
AI won't replace physics-based climate models, and it shouldn't. It adds speed, detail, and new ways to work with messy data. Used with care-clear evaluation, transparency, and physical reasoning-it can move climate science and climate action forward.
If you're strengthening AI skills for research workflows, here's a curated set of AI courses by skill level and focus areas: Complete AI Training: Courses by Skill.
Your membership also unlocks:
