Earth System World Model

01The model, live

Do something to this forest.

Pick an action and a year. The model reruns forty years of this site's real climate, and the forest you chose grows beside the one left alone.

Site

Forest

2021

—

in 2001

The response

Your scenario No action ED simulation

What is actually being computed?

The paper's model, all 519,047 parameters, runs in this tab on that site's real climate record, seeded by five years of the Ecosystem Demography simulation. An action is a change in canopy height; the model infers everything else. Verify it against the original PyTorch.

The forest is drawn from the predicted canopy height: trees still standing match the plot left alone, a loss is trees removed, and any recovery the model predicts grows back in the gaps. Which trees go, and the props, are illustration. Actions pass through a learned decoder, so a 2 m cut lands near 2.05 m, and values are December states.

02Digital twin

Now, a whole landscape.

256 forest cells of the western Amazon, where Brazil, Colombia and Peru meet, each one the paper's model running live. Log a stretch of forest, set a fire, replant, then press play.

Western Amazon2.5°N–5.5°S · 64.5–72.5°W

2021

Carbon above ground, 2021

…

Starts when you scroll here

in 2001

Each 0.5° cell (about 55 km) runs on its own, as in the paper: nothing spreads between cells. Imagery: NASA Blue Marble. Rivers and towns: Natural Earth.

03How it works

Emulators let you watch.
A world model lets you interact.

Video went from generators you watch to world models you can act in. Most Earth emulators, for weather and for ecosystems, are still the first kind: they replay a simulator under the climate they are given, with no input for a decision. This paper builds the second kind.

Earth emulators: the planet's weather, and DeepED's forest replayed from 1981 to 2021

This paper's Earth system world model: the forest you can act on, above

Ask both the same question. What if, in 2001, we

Passive emulator

ŷ_t+1 = F(y_1:t, x_1:t)

Replays the simulator under the forcing it is given. There is nowhere to put a decision.

World model this paper

ŷ_t+1 = F(y_1:t, x_1:t, a_t)

The same forecast, plus an input for actions.

Two ideas, tested on six continents.

The paper turns emulator training from passive input–output supervision into action-conditioned state transitions, without a single labelled intervention. Its three contributions:

Contribution 1

Transition-action pretraining

Simulator runs record no interventions, so every observed change becomes one: if canopy height rose 0.4 m in a year, that is a +0.4 m action. The same runs teach both modes, natural growth with no action and change under one.
Contribution 2

Masked response learning

An edit changes canopy height and leaves biomass, soil carbon and the fluxes out of date. Training masks them at the edited step, so the model learns to infer their coupled response from the edit alone.
Contribution 3

Six regions, three forest ages

Trajectories from six regions on six continents, young to mature stands: 21,315 samples, with test locations held apart from training. Accuracy holds, and responses follow the action.

The action-conditioned world model: forcing, state and an optional user action enter an encoder; an action gate selects either the encoded action or a natural-dynamics head; non-height variables are masked; the decoder produces a prescribed rollout or an action-conditioned response. — The framework (paper Fig. 1). An action gate gives one model two modes: with no action, a natural-dynamics head predicts growth; with one, the action sets the height and the rest is inferred.

04Results

Controllable, and no less accurate.

Forty-year rollout error at held-out cells, over seven variables. The world model beats the dedicated emulator at two of three stand ages and matches it at the third.

0.162

rRMSE, young forest vs 0.180 for the baseline

0.189

rRMSE, intermediate vs 0.211 for the baseline

7

coupled variables from a single edit

519k

parameters run in a browser tab

Table 1. Rollout mean rRMSE across 3 forest conditions. Lower is better.
Model	All time steps			Final time step
	Young	Inter.	Mature	Young	Inter.	Mature
Persistence	0.710	0.295	0.302	1.124	0.376	0.347
Baseline emulator	0.180	0.211	0.216	0.207	0.217	0.242
World Model	0.162	0.189	0.221	0.202	0.205	0.246

Height and aboveground biomass under 0, -1, -2 and -5 percent height actions for young, intermediate and mature stands. — Responses are ordered by action size, and strongest in young forests.

05Paper

Earth System World Model for What-If Simulations

A Case Study for Terrestrial Ecosystems. ACM SIGSPATIAL '26, Riverside, CA.

Zhihao Wang, Ruichen Wang, Ruohan Li, Lei Ma, George Hurtt, Xiaowei Jia, Gengchen Mai, Shaowen Wang, Yiqun Xie

University of Maryland · Rutgers University · UT Austin · UIUC

arXiv:2609.08855 DOI CarbonGlobe data

BibTeX

@inproceedings{wang2026earth,
  title     = {Earth System World Model for What-If Simulations:
               A Case Study for Terrestrial Ecosystems},
  author    = {Wang, Zhihao and Wang, Ruichen and Li, Ruohan and
               Ma, Lei and Hurtt, George and Jia, Xiaowei and
               Mai, Gengchen and Wang, Shaowen and Xie, Yiqun},
  booktitle = {Proceedings of the 34th ACM International Conference
               on Advances in Geographic Information Systems
               (SIGSPATIAL '26)},
  year      = {2026},
  address   = {Riverside, CA, USA},
  publisher = {ACM},
  doi       = {10.1145/3841645.3843371}
}

From passive emulators to interactive world models.

Two ideas, tested on six continents.

Transition-action pretraining

Masked response learning

Six regions, three forest ages