Source-linked AI summary
Earth System World Model for What-If Simulations: A Case Study for Terrestrial Ecosystems
Zhihao Wang, Ruichen Wang, Ruohan Li, Lei Ma, George Hurtt, Xiaowei Jia, Gengchen Mai, Shaowen Wang, Yiqun Xie
TL;DR
Existing Earth-system emulators mainly forecast under prescribed forcings, leaving explicit intervention modeling limited for interactive what-if workflows. The paper introduces action-conditioned state-transition learning with transition-action pretraining and masked response learning, and experiments show preserved long-horizon emulation alongside controllable interventions and coherent coupled responses.
Problem
Most Earth-system emulators reproduce prescribed trajectories without an explicit mechanism for user-specified interventions.
Method
The framework derives label-free action supervision from natural state transitions and uses masked response learning to infer coupled variables under partial state edits.
Results
Experiments preserve strong long-horizon emulation performance while producing consistent user-specified action responses and reconstructing non-controlled variables under masking.
Takeaways & Limitations
The framework provides a first step toward interactive, intervention-aware Earth-system emulators for scientific what-if analysis.
Abstract
from arXiv · showhide
Machine learning emulators have become essential for accelerating expensive Earth-system simulations, but most existing approaches remain passive forecasters: they reproduce simulator trajectories under prescribed forcings without an explicit interaction mechanism for user-specified interventions. This limits their use in interactive scientific workflows and Earth-system digital twins, where users often need to explore how a system would respond if selected state components were changed. We propose an action-conditioned world-modeling framework for Earth-system emulation that reformulates simulator trajectories as supervision for controllable state-transition learning. The key idea is transition-action pretraining: naturally observed state changes are treated as label-free action supervision, allowing the model to learn both prescribed dynamics and action-conditioned responses without manually annotated interventions. We further introduce masked response learning to infer unobserved variables under partial state edits and learn coupled system dependencies. We test this framework on ecosystem dynamics across six global regions and multiple stand ages. Experiments show that the model preserves competitive long-horizon emulation accuracy while enabling controllable structural interventions and coherent responses in coupled ecosystem-cycle variables. These results suggest a practical route from passive Earth-system emulators toward interactive, intervention-aware scientific surrogates.
1 Introduction
The paper proposes extending passive Earth-system emulators with interactive, action-conditioned dynamics for user-specified state edits. Its framework derives action supervision from natural transitions and uses masked response learning to model partially observed coupled responses.
- Motivation: Most Earth-system emulators reproduce trajectories under prescribed forcings without explicitly supporting user-specified interventions.This limits their direct use for interactive what-if exploration.
- Motivation: World models add controllability by predicting how system states evolve under actions or interventions injected during rollout.This distinguishes action-conditioned dynamics from fixed-input forecasting.
- Approach: The framework reformulates passive input-output training as action-conditioned state-transition modeling for explicit user-edited ecosystem states.The study focuses on terrestrial ecosystems and rapid what-if exploration.
- Approach: Transition-action pretraining derives label-free action supervision from naturally observed state changes, avoiding manually annotated interventions.The approach treats observed transitions as supervision for learning controllable dynamics.
- Approach: Masked response learning improves stability with partially unobserved variables and supports partial state edits.Experiments cover ecosystem trajectories across heterogeneous regions from six continents.
2 Related Work
Prior work has developed fast Earth-system emulators and action-conditioned world models largely in separate research lines. This paper positions its framework as an early effort to bridge them for interactive, intervention-aware scientific emulation.
- Earth-system emulation: Deep learning emulators approximate expensive atmospheric, terrestrial, and hydrological models, including large-scale ecosystem simulations with substantial speed-ups.These systems primarily target efficient surrogate prediction of theory-based models.
- World models and controllable dynamics: World models learn compact state representations and action-conditioned dynamics for planning, control, and counterfactual interaction.They have been studied across reinforcement learning, robotics, video prediction, autonomous driving, and interactive generative environments.
- World models and controllable dynamics: Earth-system settings additionally involve physically coupled and partially observed system states.These properties motivate adapting interactive world-model ideas beyond conventional domains.
- Research gap: The paper is presented as among the first attempts to bridge Earth-system emulation and interactive world modeling.It does so through transition-action pretraining and masked response learning.
3 Method
The method reformulates passive ecosystem emulation as action-conditioned state-transition modeling, using observed state changes to supervise controllable interventions. A gated architecture and masked response learning support both prescribed rollouts and coupled responses to partial state edits.
- Problem formulation: The model represents environmental forcings and ecosystem states containing coupled structural and ecosystem variables such as canopy height, biomass, soil carbon, and carbon fluxes.
- Action-conditioned dynamics: User-specified actions edit selected state components, representing interventions such as logging, reforestation, disturbance, or land-management changes.
- Transition-action pretraining: Transition-derived actions convert simulator state changes into self-supervised action examples without manually annotated interventions, instantiated here as canopy-height changes.
- Transition-action pretraining: Training alternates between prescribed mode without actions and action-conditioned mode with transition-derived actions, allowing the same trajectories to supervise both behaviors.
- Action-conditioned dynamics: An action gate uses natural dynamics when no action is provided and the encoded action when one is specified, while decoding remaining variables from the resulting controlled state.
- Masked response learning: Masked response learning masks outdated non-controlled variables during partial interventions so the model can infer their coupled responses from the observed controlled state and action.
4 Experiment
The framework is evaluated on spatially separated ecosystem data across six regions and three forest ages, using long-horizon prescribed emulation, perturbed height actions, and masked response inference. Results indicate competitive baseline performance, controllable height responses, and coupled biomass behavior under partial observations.
- Experimental Setup: The evaluation uses 21,315 CarbonGlobe samples from six global regions, with young, intermediate, and mature forests and spatially separated training and testing locations.The dataset is based on 40 years of Ecosystem Demography model forecasting, with 85% training and 15% testing locations.
- Experimental Setup: The model combines five-year forcing histories with ecosystem states, recurrent temporal context, height-change prediction, and multi-variable response decoding.A GRU encodes temporal context; natural dynamics predict height changes before the decoder predicts remaining ecosystem responses.
- Prescribed Emulation Performance: The World Model achieves comparable or better long-horizon rRMSE than the dedicated baseline emulator and consistently outperforms persistence.This result holds across young, intermediate, and mature forest conditions and indicates that action-conditioned training does not compromise prescribed emulation.
- Controllability under Height Actions: Perturbed height actions produce ordered height responses, with stronger effects in young forests and weaker effects in mature forests.The age pattern is consistent with age-dependent growth dynamics and slower structural change near maturity.
- Masked Response Inference: Under height-only observation, aboveground biomass deviations remain close to zero after masking, matching the full-observation setting.Mature forests show especially weak non-height responses despite height perturbations, supporting inference of coupled responses from partial state information.
5 Conclusions
The study presents an action-conditioned world-modeling framework for Earth-system emulation that supports prescribed prediction and user-specified structural interventions. Its transition-based action supervision and masked response learning preserve long-horizon performance while producing consistent controlled and reconstructed ecosystem responses.
- Conclusions: The framework supports both prescribed prediction and user-specified structural interventions in learned Earth-system emulators.It derives action supervision from state transitions and uses masked response learning for coupled responses under partial state information.
- Conclusions: Transition-derived actions and masked response learning enable consistent responses to interventions and reconstruction of non-controlled variables under masking.Experiments on ecosystem trajectories show these capabilities alongside strong long-horizon emulation performance.