A Vision for Simulation-based, Multi-fidelity Digital Twins in Fusion Energy

Fusion energy is a unique field where the combination of rich multimodal diagnostics and multi-physics, multi-fidelity simulation are vital to progress. Here we present several efforts and techniques pursued towards building out faithful digital twins, based on a range of simulations covering differing physics and levels of fidelity. These digital twins can be leveraged for eventual use in downstream tasks such as design optimization, scenario planning, experiment interpretation, and control.

Work done through a multi-institution SciDAC project called StellFoundry is building up a framework for faithful digital models of stellarators, covering the self-consistent physics predictions from the fusion power core all the way to engineering calculations utilizing neutron and heat loads. Tools and schema for coupling these codes in HPC environments for multi-physics simulation capability are being developed, in addition to advanced optimization techniques and AI surrogates, to include higher-fidelity simulation in optimization loops.

Digital twins can be comprehensive or focused on targeted subsystems. In another work we create an AI model for accelerated predictions of the HEAT code, used for calculations of the divertor heat load on the upcoming SPARC tokamak. For a fixed divertor design, this model can predict CAD-level, detailed divertor heat loads, useful for engineering operation monitoring, and reduction into control algorithms.

Grounding digital models with data from the physical asset is a key requirement for digital twins, to account for gaps in physical realism of simulation models. Simulation-based inference based on AI models enables fast Bayesian inference of digital twin physics information from a combination of diagnostics, and opens a path for grounding of the models in addition to uncertainty quantification.