Page 548 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

initial results demonstrate that privileged information enables the efficient training of ML agents
while maintaining the simplicity of the Critic network's architecture. The Actor learns the control
policy directly from sensor and actuator data, which can speed up the controllers by eliminating
the intermediate step of magnetic flux distribution reconstruction. These advancements have
laid the groundwork for developing essential components of the PCS. We believe that the
intersection of ML and fusion science holds the key to unlocking sustainable, clean energy for
our planet’s future.

Partners

• University of California, San Diego, USA
• General Atomics (DIII-D National Fusion Facility), USA

2�2 Benefits of the use case

1. In this use case, we use AI-based reinforcement learning and plasma state exploration
to optimize plasma control, reduce disruptions, and enhance reactor efficiency. It helps
increase the stability and energy output of fusion reactors, making fusion a more reliable
and cost-effective clean energy source.
2. Our use case leverages advanced RL control modules and digital twin technology to
modernize fusion reactor infrastructure. By enabling real-time plasma management and
predictive diagnostics, our solution accelerates fusion commercialization and supports
the development of next-generation energy.
3. We actively collaborate with leading research institutions, including the University of
California in San Diego, Columbia University, University of Seville, and the DIII-D National
Fusion Facility, as well as private fusion developers. By working with early adopters and
industry stakeholders, we accelerate the validation, deployment, and scalability of AI-
driven fusion control systems, contributing to the global clean energy transition.

2�3 Future Work

With additional resources, we plan to expand the list of controlled parameters and validate our
ML-based Plasma Control System across a broader range of fusion devices in collaboration
with early adopters. This would allow us to refine our models, enhance adaptability across
different reactor designs, and further demonstrate the system’s scalability and effectiveness.
To achieve this, we would require more experiments and extensive data analysis, alongside
the development of a simulation environment, which would enable us to train our ML models
for more diverse control tasks and shift the control paradigm to control high-level plasma
parameters, thereby defining overall plasma energy gain. This would improve the accuracy
and robustness of real-time plasma state predictions and control strategies.

3 Use Case Requirements

REQ-01: Access to Experimental Data

It is critical that there is availability of time-series data from operating fusion devices (e.g.,
magnetic sensors, actuator signals) for training and validating ML control models.

REQ-02: Tokamak Simulator

It is critical that there is a physics-based, high-fidelity simulator (such as our NSFsim) to model
plasma behaviour and evaluate control strategies in a tokamak environment.

512