Princeton RL controller cuts DIII-D disruptions to two shots

An RL controller trained on a tearing-likelihood model kept DIII-D plasmas below the instability threshold; before RL control all but one shot disrupted, after it only two did (Nature, 21 Feb 2024).

2ⓘ

AI system

System

Deep reinforcement learning tearing-avoidance controller with a DNN tearability predictor (Keras model converted to real-time C)

Model

Not assessed

Release

Not assessed

Application developer

Princeton University / Princeton Plasma Physics Laboratory (Kolemen group), with Chung-Ang University

Role

Not assessed

Configuration

Not assessed

Inputs and outputs

Not assessed

Autonomy

Not assessed

Workflow

A database of past DIII-D discharges to train the tearability predictor. Real-time profile reconstruction and equilibrium fitting feeding the DNN. A plasma control system that accepts high-level beam-power and shape commands.

Human oversight

Not assessed

Relation to AI

The RL controller set the high-level actuator commands that the plasma control system executed. The disruption outcomes in the controlled shots are therefore attributable to it. The comparison is within one session, and the paper is authored by the system's developers. Independent evaluation is the journal's peer review and the OECD AIM registry entry.

AI contribution

The RL controller set the high-level actuator commands that the plasma control system executed. The disruption outcomes in the controlled shots are therefore attributable to it. The comparison is within one session, and the paper is authored by the system's developers. Independent evaluation is the journal's peer review and the OECD AIM registry entry.

Benefits

Outcome identifiers

Not assessed

Benefit

before RL control all but one shot disrupted, after it only two did

Category

Improved Task Accuracy

Status

Not assessed

Beneficiary groups

Organizations. Society.

Beneficiaries

Not assessed

Reach

Not assessed

Distribution and access

Not assessed

Duration

Not assessed

Expected?

Yes

Costs and harms

Not assessed

Measurement

Studies

Not assessed

Study design

Not assessed

Measurement

Not assessed

Setting and sample

Not assessed

Measures

before RL control all but one shot disrupted, after it only two did

Comparison

DIII-D discharges in the same experiment session without RL control

Results

Reference: all shots except one (193266: low-βN reference shown in Fig. 3b) were disrupted. Result: only two (193277 and 193282) were disrupted.

Additional value from AI

Not assessed

Limitations

Not assessed

Denominators

Not assessed

Materials for replication

Not assessed

Money

Kind of value

Not assessed

Amount

Not assessed

Currency and prices

Not assessed

Period and unit

Not assessed

Realization

Not assessed

Recipients and payers

Not assessed

Gross or net

Not assessed

Included costs

Not assessed

Comparison

Not assessed

Method

Not assessed

Return and payback

Not assessed

People and places

Participants

Developer: Princeton University / Princeton Plasma Physics Laboratory (Kolemen group), with Chung-Ang University. Operator: DIII-D National Fusion Facility, General Atomics.

Jurisdictions

Benefit: San Diego, United States. Developer: United States. Operator: United States.

Ownership and control

Not assessed

Dependencies

Not assessed

Cross-border connections

Not assessed

Geopolitical implications

Not assessed

Follow-up

Conditions for replication

A tearability model valid for the operating regime. Actuators (beam power, triangularity) able to move the plasma along the stable path. Session-level comparison shots to judge disruption rates.

Replication attempts

1 follow-up entries. Benefit recurred: 1.

Later findings

Later DIII-D papers from the same group describe successor tearing-mode controllers but no further result for this RL controller was found.

Availability

Not assessed

Institutional decision

Not assessed

Grounds

Not assessed

Action and result

Not assessed

Success criterion

Not assessed

Evidence

Source numbers

AIBD-0165-source-0001–0004

Citations

Numbered references below.

Language

Not assessed

Source relationships

Developer, Independent

Original evidence

Not assessed

Access and versions

Not assessed

Supported findings

Not assessed

Evidence strength

Not assessed

Missing evidence

Not assessed

Grades

Assessed benefits

Organizations. Society.

Depth

Organizations: 2. Society: Not assessed.

Reasons

Organizations: The DIII-D session obtained plasmas tracking a stable path with fewer disruptions and higher time-integrated beta. The paper calls it a path toward ITER-relevant scenarios. Society: No outcome beyond fusion research is documented.

Observed or intended

Not assessed

Rubric

Not assessed

Assessor and date

Not assessed

Alternative assessments

Not assessed

Fill

The makers' claim

Entry details

Kind

Not assessed

Field

Energy research (magnetic confinement fusion). Tokamak plasma control.

Purpose

The controller was built to avoid tearing instability while maintaining high-pressure plasma in DIII-D. The avoided disruptions were its intended use.

Dates

Date basis unassessed: 2024 (precision unassessed). Updated: 21 October 2026.

Stage

Not assessed

Version

SHA-256: 114ba554771a…

References

Other identifiers

Not assessed

Category mappings

Not assessed

Brief

Not assessed

Revisions

5 revisions. Latest: 21 October 2026.

Sources

  1. Peer-reviewed study Developer

    Seo, J., Kim, S., Jalalvand, A. et al., 'Avoiding fusion plasma tearing instability with deep reinforcement learning', Nature (2024), doi:10.1038/s41586-024-07024-9

  2. Press release Developer

    Princeton Engineering news, 'Engineers use AI to wrangle fusion power for the grid', 21 February 2024

  3. Registry label Independent author

    OECD AI Incidents Monitor, record 2024-02-24-7d77