Presentation Information
[s8HxLUFw-03]Agentic Earth System Digital Twins: From Prediction to Collaborative Scientific Discovery
*Jennifer Sleeman1 (1. APL, Johns Hopkins (USA))
The landscape of Earth system modeling has advanced significantly with the
advent of AI, giving rise to AI emulators, differentiable models, physics-informed
models, and foundation models. These approaches are changing not only how
we model and predict the Earth system, but also how AI can participate in
the scientific process. As these capabilities mature, an increasingly important
challenge is how to coordinate heterogeneous models, observations, experiments,
and scientific reasoning across domains and scales. Agentic AI offers a new
paradigm in which AI can progress from a scientific tool or assistant toward a
collaborator capable of designing experiments, coordinating analyses, evaluating
evidence, and helping scientists explore new hypotheses.
Within this emerging paradigm, we consider how Earth system digital twins
can evolve from predictive representations toward environments for collaborative
scientific discovery. We present an agentic digital twin framework that treats
scientific questions as executable workflows. Scientific questions are interpreted
to identify the observations, models, and computational experiments needed to
investigate them. Agents coordinate these workflows by acquiring data, executing
and evaluating models, comparing experiments, and synthesizing results.
When existing capabilities are insufficient, the framework can dynamically construct
new AI models, allowing the digital twin’s modeling capabilities to evolve
as new scientific questions emerge.
We demonstrate this approach through an Earth system tipping-point digital
twin that coordinates reduced-order models with AI-based methods for discovery
and forecasting to investigate abrupt transitions and interactions across
ocean, atmosphere, and land systems. Current applications include the Atlantic
Meridional Overturning Circulation, the Amazon, and the West African
Monsoon, providing environments for exploring hypothetical and counterfactual
scenarios and, ultimately, interactions and cascading behavior across Earth
subsystems. These capabilities point toward a broader role for digital twins in
AI-enabled science: not only predicting how the Earth system may evolve, but
enabling scientists and AI to collaboratively ask questions, conduct experiments,
evaluate evidence, and discover new understanding.
1
advent of AI, giving rise to AI emulators, differentiable models, physics-informed
models, and foundation models. These approaches are changing not only how
we model and predict the Earth system, but also how AI can participate in
the scientific process. As these capabilities mature, an increasingly important
challenge is how to coordinate heterogeneous models, observations, experiments,
and scientific reasoning across domains and scales. Agentic AI offers a new
paradigm in which AI can progress from a scientific tool or assistant toward a
collaborator capable of designing experiments, coordinating analyses, evaluating
evidence, and helping scientists explore new hypotheses.
Within this emerging paradigm, we consider how Earth system digital twins
can evolve from predictive representations toward environments for collaborative
scientific discovery. We present an agentic digital twin framework that treats
scientific questions as executable workflows. Scientific questions are interpreted
to identify the observations, models, and computational experiments needed to
investigate them. Agents coordinate these workflows by acquiring data, executing
and evaluating models, comparing experiments, and synthesizing results.
When existing capabilities are insufficient, the framework can dynamically construct
new AI models, allowing the digital twin’s modeling capabilities to evolve
as new scientific questions emerge.
We demonstrate this approach through an Earth system tipping-point digital
twin that coordinates reduced-order models with AI-based methods for discovery
and forecasting to investigate abrupt transitions and interactions across
ocean, atmosphere, and land systems. Current applications include the Atlantic
Meridional Overturning Circulation, the Amazon, and the West African
Monsoon, providing environments for exploring hypothetical and counterfactual
scenarios and, ultimately, interactions and cascading behavior across Earth
subsystems. These capabilities point toward a broader role for digital twins in
AI-enabled science: not only predicting how the Earth system may evolve, but
enabling scientists and AI to collaboratively ask questions, conduct experiments,
evaluate evidence, and discover new understanding.
1
Comment
To browse or post comments, you must log in.Log in