Nathaniel Chen
所有專案專案文章 · 英文

World modeling for fusion Fusion World Model

Learn how a plasma evolves. Help shape what happens next.

A world model for fusion, built from scratch to connect diagnostic measurements, learn plasma dynamics, and support better experimental decisions.

ContributionBuilt from scratch · multimodal modeling & scientific software

Project record2025–present · ongoing research program

ObserveSignals, images & inputs

Different instruments, one experiment.

RepresentA shared plasma state

Connect what the measurements tell us.

PredictWhat happens next?

Study evolution and possible interventions.

Fusion World Model · the research direction, from observations to dynamics and experimental decisions.

From measurements to an evolving plasma

A fusion experiment observes one plasma through many instruments. Each captures a different part of its behavior, at a different time scale. A useful model needs to connect those observations and describe how the system changes.

Fusion World Model is the broader project behind this work. The goal is to learn a shared account of plasma state and dynamics that can help researchers interpret measurements, anticipate changes, and investigate possible control decisions.

Built from scratch, across the research stack

I built Fusion World Model from scratch with Azarakhsh Jalalvand, Peter Steiner, and Kouroche Bouchiat. The project brings together scientific data organization, multimodal representation learning, and the software needed to study plasma behavior.

The early foundation-model presentation explored learning across diagnostic modalities through masked reconstruction. The data-organization paper examines how to bring heterogeneous measurements into that learning process, including the trade-off between temporal context and frequency resolution.

A research program that keeps growing

The papers and presentations below document different parts of the project. Each contributes to the larger aim of making a learned plasma model useful for scientific understanding and experimental decisions.

Ongoing work extends from representing the plasma toward understanding its evolution and response to interventions. Further papers and workshop contributions will document that progress as results become available.

Papers & presentations

A growing record of the work behind Fusion World Model.

Towards Large-Scale Heterogeneous Data Organization for Scientific Foundation Models: A Nuclear Fusion Case Study

2026 · Paper · DATA-FM workshop at ICLR 2026

Accepted · preprint available

Organizing heterogeneous diagnostic measurements for scientific foundation models.

Towards a Foundation Model for Fusion: Multimodal Representation Learning of Plasma State and Control

2025 · Poster · APS DPP 2025

Presented · poster available

An early study of learning a shared representation across plasma diagnostics.

Upcoming work

Planned publications & presentations

NeurIPS workshop contributions

Planned contributions from the project. Workshop names, titles, and links will be added as they are confirmed.

Further Fusion World Model papers

Additional papers will document the developing model, methods, and experimental findings.

Further reading

Essay: understanding the world through measurements

An introduction to the ideas behind the project · May 15, 2026