Scientific World Models: Toward Closed-Loop Scientific Intelligence
Abstract
Artificial intelligence is moving from task-specific scientific tools toward foundation models, generative design, research agents, and autonomous laboratories. Despite this progress, most systems remain organized around isolated predictions or local workflows. They rarely maintain an evolving account of scientific objects, predict the consequences of interventions, or learn systematically from computational and experimental evidence. This article introduces scientific world models as a core paradigm for closed-loop scientific intelligence. A scientific world model reconstructs the physical state of a scientific object and its corresponding epistemic state—what is known, uncertain, or contested—from partial, multimodal, and instrument-mediated observations. It models action-conditioned state transitions, explores plausible futures under scientific constraints, and supports the selection of measurements, simulations, experiments, and process interventions according to scientific value, information gain, cost, feasibility, and risk. The new evidence then updates the state, hypotheses, and subsequent actions. On this basis, we propose a three-layer architecture of scientific perception, cognition, and action. Perception turns the physical world into traceable scientific observations and state representations. Cognition combines data-driven models with scientific laws, mechanistic knowledge, and simulation to interpret current states and reason about possible futures. Action connects model reasoning to computation, instruments, experiments, and physical processes. We further distinguish the scientific world model core from foundation models, scientific agents, digital twins, and autonomous laboratories, and outline domain instantiations, maturity levels, evaluation criteria, and open technical challenges. This framework provides a common basis for developing scientific AI systems that can observe, reason, act, and improve through evidence.
Related
- LLM-AutoSciLab: Closed-Loop Scientific Discovery via Active Experimentation with LLMs
- Closed-loop formulation–process optimization under unknown feasibility for extended-release pellets
- Closed-loop Auto Research for Molecular Property Prediction: Discovering and Certifying Generalizable Improvements
- Closed-loop optimization in an industrial R&D lab
- Closed-loop photoreactor design enabled by machine learning and digital twins
BibTeX
@misc{wang2026scientific,
title = {{Scientific World Models: Toward Closed-Loop Scientific Intelligence}},
author = {Huan Wang and Zhenlin Zhao and Siming Dong},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009602/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009602/v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Scientific World Models: Toward Closed-Loop Scientific Intelligence
AU - Huan Wang
AU - Zhenlin Zhao
AU - Siming Dong
PY - 2026
DA - 2026/09/29
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009602/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009602/v1
AB - Artificial intelligence is moving from task-specific scientific tools toward foundation models, generative design, research agents, and autonomous laboratories. Despite this progress, most systems remain organized around isolated predictions or local workflows. They rarely maintain an evolving account of scientific objects, predict the consequences of interventions, or learn systematically from computational and experimental evidence. This article introduces scientific world models as a core paradigm for closed-loop scientific intelligence. A scientific world model reconstructs the physical state of a scientific object and its corresponding epistemic state—what is known, uncertain, or contested—from partial, multimodal, and instrument-mediated observations. It models action-conditioned state transitions, explores plausible futures under scientific constraints, and supports the selection of measurements, simulations, experiments, and process interventions according to scientific value, information gain, cost, feasibility, and risk. The new evidence then updates the state, hypotheses, and subsequent actions. On this basis, we propose a three-layer architecture of scientific perception, cognition, and action. Perception turns the physical world into traceable scientific observations and state representations. Cognition combines data-driven models with scientific laws, mechanistic knowledge, and simulation to interpret current states and reason about possible futures. Action connects model reasoning to computation, instruments, experiments, and physical processes. We further distinguish the scientific world model core from foundation models, scientific agents, digital twins, and autonomous laboratories, and outline domain instantiations, maturity levels, evaluation criteria, and open technical challenges. This framework provides a common basis for developing scientific AI systems that can observe, reason, act, and improve through evidence.
ER -