Climate science × symbolic AI

MOSAIC

Multi-scale Observation-driven Symbolic AI for Climate

MOSAIC is building an AI-assisted path towards better climate models. We bring observational data, symbolic equation discovery, differentiable simulation, and rigorous evaluation into a unified research loop. Our goal is to discover novel and interpretable climate parameterizations, test them at a global scale, and quantify the improvements.

Motivation

Many of the processes that shape climate projections (e.g. clouds, rain formation, aerosol activation, land-surface exchange, and boundary-layer mixing) occur at scales too fine for a global model to resolve directly. Scientists represent their aggregate effects with parameterizations: compact equations that traditionally take years of expert design and hand-tuning.

At the same time, satellites, field campaigns, reanalyses, and high-resolution simulations produce far more process-level information than researchers can manually distill. MOSAIC asks whether an AI system can help turn those observations into useful scientific discoveries. Can AI derive new, physically credible equations, implement them in a global atmospheric model, and calibrate them against observations to improve on hand-tuned baselines?

MOSAIC focuses especially on equation discovery over alternative approaches like neural network emulators. Symbolic equations provide unique benefits for interpretability, modifiability, and the ability to advance our understanding of earth system processes. We aim also to advance the science of physical constraint discovery and equation interpretability itself: how can we best guide equation discovery to satisfy physical constraints and identify genuinely interesting parameterizations?

Our research loop

The MOSAIC loop connects evidence to equation discovery, model integration, and global evaluation. Initial results inform the direction of subsequent experiments. In the longer term, model sensitivities also identify useful areas for new observations.

01

Curate the evidence

The agent works with matched satellite, reanalysis, field-campaign, and high-resolution simulation data. Initial sources include NASA MODIS, DOE Atmospheric Radiation Measurement data, ERA5, LASSO, and ClimSim.

02

Discover candidate equations

Physics-guided agents formulate hypotheses and orchestrate symbolic-regression and constraint-discovery tools. Candidates are screened for complexity, conservation, boundedness, stability, and interpretability.

03

Integrate and calibrate

Promising equations become modular, differentiable JAX components. They are inserted into JAX-GCM and tuned through model rollouts against observable quantities.

04

Evaluate and learn

Each candidate is compared with existing parameterizations on withheld observations, unfamiliar weather regimes, and longer global simulations. Diagnostics and failures become evidence for the next discovery cycle.

What makes MOSAIC possible

• A differentiable climate model. JAX-GCM (JCM) is an open, intermediate-complexity atmospheric model written in JAX. Automatic differentiation lets us tune a candidate parameterization through the model against observations and investigate how uncertain processes affect downstream climate metrics. Explore the code on GitHub or read the JCM v1.1 paper.
• Physics-guided AI agents. KeplerAgent provides a foundation for coordinating symbolic regression and symmetry discovery while reasoning about governing equations and physical constraints.
• Rigorous model evaluation. ClimateBench2 develops a common benchmark for climate models, combining physical-consistency checks, probabilistic scores against observations, and out-of-distribution tests. MOSAIC uses it to assess whether discovered equations improve global model skill and robustness.

Objectives

Multiple interpretable equationsDiscovered and integrated across targeted climate processes

Measurable improvementOver hand-tuned baselines on withheld observational data

Stable global behaviorUnder longer simulations and unfamiliar meteorological regimes

Reproducible open outputsIncluding code, equations, tests, benchmarks, and evaluation results

A broader path to observation-driven Earth-system modeling

MOSAIC’s initial program concentrates on demonstrating a rigorous end-to-end atmospheric workflow. From there, the same approach can extend into differentiable ocean, land and coupled Earth-system modeling. Because JAX-GCM exposes model gradients, MOSAIC can ask a second question: which additional observation would most reduce the uncertainty in a climate prediction? These sensitivity maps will guide future targeted observations, like turbulence-equipped Argo ocean floats.

Team

MOSAIC brings together climate modeling, Earth observations, scientific machine learning, symbolic discovery, land modeling, and model evaluation.

Duncan Watson-Parris UC San Diego · Project lead Climate model development, benchmarking, and satellite evaluation Rose Yu UC San Diego Physics-guided AI agents and symbolic discovery Pierre Gentine Columbia University Differentiable land modeling and scientific workflows Veronika Eyring University of Tübingen Earth-system evaluation and hybrid climate modeling Matthew Alford UC San Diego · Scripps Ocean turbulence observations and observing-system design Marshall Fisher UC San Diego Research prototype implementation

Build with us

Interested in interpretable AI for climate science?

We are growing an interdisciplinary team spanning climate science, machine learning, scientific software, and Earth observations. We welcome conversations with prospective researchers, engineers, collaborators, and research partners.

Get in touch

Selected references

1. Davenport, E. H., Madan, J. V., Gjini, R., et al. (2026). “JCM v1.1: A differentiable, intermediate-complexity atmospheric model.” Geoscientific Model Development, 19. doi:10.5194/gmd-19-6451-2026
2. Yang, J., Venkatachalam, O., Kianezhad, M., Vadgama, S., & Yu, R. (2026). “Think like a Scientist: Physics-guided LLM Agent for Equation Discovery.” arXiv:2602.12259. doi:10.48550/arXiv.2602.12259
3. Watson-Parris, D., Balaji, V., Bretherton, C. S., et al. (2025). “ClimateBench2.0: Probabilistic Climate Model Scoring.” AGU25 conference abstract. NASA record · Project repository