Teaching AI how the world's forests grow.
The first global, ML-ready dataset of process-based forest carbon simulations, with benchmarks for training AI emulators of the land.
University of Maryland · University of Pittsburgh · Worcester Polytechnic Institute
01 One forest
1981, central Amazon: a young stand begins to grow. The Ecosystem Demography (ED) model, used in NASA's Carbon Monitoring System and the Global Carbon Budget, simulates its growth, mortality and reproduction from monthly climate, CO₂ and soil.
02 Fifteen starting ages
From 1-year saplings to 500-year-old stands, each grown in the same climate. Young forests race upward while old ones hold their carbon. An emulator has to learn both.
03 The planet
Repeated at every 0.5° land cell, across tropical, arid, temperate, cold and polar climates. Generating it took ED about 812 CPU-days on an HPC cluster.
04 The emulator
Every step is a training pair: 136 monthly drivers plus the current forest state in, next year's forest out. CarbonGlobe delivers it ML-ready, with tensors, splits, metrics and eight baselines.
05 What if?
Rolled out year after year, an emulator answers what-if questions: a different climate, a different starting forest. A global run that took ED 40 days of wall time takes about one GPU-hour.
06 CarbonGlobe in numbers
Why CarbonGlobe
AI weather models such as FourCastNet and GraphCast took off because open, ML-ready benchmarks existed. Forest carbon had nothing comparable. The best process-based model, ED, needs more than 100 inputs from many sensing products and about six weeks of cluster time for one global run.
CarbonGlobe closes that gap. We integrate the inputs, run and calibrate ED globally, and package the result for machine learning: inputs, targets, splits, scenarios and metrics, plus a suite of baseline emulators.
Forest ecosystems play a critical role in the Earth system as major carbon sinks that are essential for carbon neutralization and climate change mitigation. However, the Earth has undergone significant deforestation and forest degradation, and the remaining forested areas are also facing increasing pressures from socioeconomic factors and climate change, potentially pushing them towards tipping points. Responding to the grand challenge, a theory-based Ecosystem Demography (ED) model has been continuously developed over the past two decades and serves as a key component in major initiatives, including the Global Carbon Budget, NASA Carbon Monitoring System, and US Greenhouse Gas Center. Despite its growing importance in combating climate change and shaping carbon policies, ED's expensive computation significantly limits its ability to estimate carbon dynamics at the global scale with high spatial resolution. Recently, machine learning (ML) models have shown promising potential in approximating theory-based models with interesting success in various domains including weather forecasting, thanks to the open-source benchmark datasets made available. However, there are currently no publicly available ML-ready datasets for global carbon dynamics forecasting in forest ecosystems. The limited data availability hinders the development of corresponding ML emulators. Furthermore, the inputs needed for running ED are highly complex with over a hundred variables from various remote sensing products. To bridge the gap, we develop a new ML-ready benchmark dataset, CarbonGlobe, for carbon dynamics forecasting, featuring that: (1) the data has a global-scale coverage at 0.5° resolution; (2) the temporal range spans 40 years; (3) the inputs integrate extensive multi-source data from different sensing products, with calibrated outputs from ED; (4) the data is formatted in ML-ready forms and split into different evaluation scenarios based on climate conditions, etc.; (5) a set of problem-driven metrics is designed to develop benchmarks using various ML models to best align with the needs of downstream applications.
The dataset
Three stages, done jointly by carbon scientists and ML researchers, so you can start training on day one.
1Integrate 100+ inputs
Monthly meteorology (incl. 24-hour profiles), surface CO₂, soil properties and disturbance: 136 features per month at 0.5°.
2Simulate with ED
~812 CPU-days on AMD EPYC nodes, for 15 seed ages at every site.
3Package 7 targets
Height, AGB, soil C, LAI, GPP, NPP and Rh for every site, year and seed age.
X812,280 × 40 × 12 × 136sequences × years × months × driversv₀812,280 × 7initial forest stateY812,280 × 40 × 7annual targets to forecastTask: given X and v₀, forecast Y step by step for all 40 years.
Scenario 1
Train on 1/16 of sites (roughly a 4× coarser grid) and predict the rest, just as an emulator would fill in a finer grid.
Scenario 2
Scores split by Köppen–Geiger zone: tropical, arid, temperate, cold and polar. Where does an emulator break?
Scenario 3
Fifteen seed ages from 1 to 500 years test young, fast-growing stands and old, carbon-dense ones.
Problem-driven metrics
Averaging errors over 40 years hides what matters downstream. Next to RMSE and MAE, CarbonGlobe scores delta error, the error in year-over-year change, and cumulative error, how far a forecast has drifted by year 40.
Explore
Real CarbonGlobe targets on every 0.5° land cell. Pick a variable and a starting forest age, press play, and click anywhere on land to see that site's trajectory.
Drag to rotate · scroll to zoom · click land
Variable
Starting forest age
Year 2021
—
ED reference simulations at 0.5°. Stocks show the December state; LAI and fluxes are annual means. Values are quantized for the browser; download the dataset for full precision.
Benchmarks
The ED-specific DeepED leads on slowly accumulating stocks; attention models lead on seasonal fluxes. That gap is the open research problem.
Lower is better · click a column to compare models
Overall evaluation with noise-perturbation training, Table 1 of the paper. Colors rank the eight models within each variable.
Speed
A 40-year global forecast took ED 40.6 days of wall time on 20 CPU cores. Most emulators finish in about an hour on a single RTX A4500.
Impact
Weather and climate models already have fast AI counterparts. CarbonGlobe trains the land carbon component that can plug in next to them.
Couple
A fast, differentiable land carbon module that can sit inside AI or hybrid Earth-system models.
Scale
A 0.1° run would cost ED ~25× more. Emulators trained on coarse grids can fill in finer ones.
Explore
Sweep climate pathways, disturbance and reforestation plans fast enough for interactive what-if analysis.
Monitor
ED already informs the Global Carbon Budget, NASA CMS and Maryland's carbon inventory. Emulators make it continuous.
Get started
@inproceedings{NEURIPS2025_0b82662b,
author = {Wang, Zhihao and Ma, Lei and Hurtt, George and Jia, Xiaowei and Li, Yanhua and Li, Ruohan and Li, Zhili and Xu, Shuo and Xie, Yiqun},
booktitle = {Advances in Neural Information Processing Systems},
doi = {10.52202/085713-0261},
editor = {D. Belgrave and C. Zhang and H. Lin and R. Pascanu and P. Koniusz and M. Ghassemi and N. Chen},
publisher = {Curran Associates, Inc.},
title = {CarbonGlobe: A Global-Scale, Multi-Decade Dataset and Benchmark for Carbon Forecasting in Forest Ecosystems},
url = {https://proceedings.neurips.cc/paper_files/paper/2025/file/0b82662b6c32e887bb252a74d8cb2d5e-Paper-Datasets_and_Benchmarks_Track.pdf},
volume = {38, Main Conference},
year = {2025}
}