DeepFusion Accelerator for Fusion Energy Sciences in Disruption Mitigations
Researcher
Sponsoring Agency:
U.S. Department of Energy
Summary
Artificial Intelligence (AI), specifically, Deep Learning (DL) promises a revolution in how physics and engineering can be seamlessly bridged for authentic predictive control and predictive design of engineering systems like fusion power reactors. This project aims to develop and apply fundamental DL/AI methodologies that would enable predictive design and predictive control in fusion energy systems. Although the immediate focus of the proposed research would be on the specific application of tokamak disruption mitigation, our proposed DL/AI methodologies and tools would be broadly applicable to fusion science and engineering. Our innovation centers around the construction of high-fidelity surrogate models that are of low computational cost at online deployment, at the expense of high cost in offline training using extreme scale computing. The key innovation is the explicit control on the fidelity of the surrogate model. Once such high-fidelity low-cost surrogate models are in place, they would allow us to develop and deploy a framework that solves inverse problems efficiently, taking advantage of the differentiability of deep neural network (DNN) surrogates. This framework would facilitate predictive control and predictive design, and will be demonstrated for the specific problem of runaway avoidance and minimization in disruption mitigation design. A key feature of the proposed DL-based framework for inverse problems is its ability to account for uncertainties in model parameters and measurement, which can be parameterized and learned by DNN surrogate as well.
Research Area
Term
September 1, 2023–August 31, 2025