Validating physics-based (ASTRA/TRANSP), data-driven (D3D+AUG), and physics+data hybrid models for quantitatively accurate yet generalizable guidance for ITER operators

ITER operators will leverage predictive models to plan upcoming operational phases, rundays, and discharges. These models are developed based on prior experiments and will ideally be continuously updated as ITER data is collected. However, the reliability of even state-of-the-art simulations for predicting and extrapolating to a fusion-grade plasma is an open question [1]; and it is not obvious how to rigorously update models with data. Leveraging recent advances in AI and tokamak data+code availability and automation, information from experiments can be combined with human-made physical simulations to perform >10% better than either alone for predicting electron and ion temperature evolution in tokamak experiments. This “meta-learning” AI approach can be retrained in milliseconds and is interpretable, which could yield not only improvements in models used by ITER operators, but also a robust mechanism for continuously updating models as ITER data is collected [2]. We use data from DIII-D and AUG, and transport simulations by both ASTRA and TRANSP (using e.g. TGLF). We consider only predictions for energy and current-timescale 1D profiles in tokamaks (temperature, density, safety factor q, Zeff), but the methodology applies to any system for which both experimental data and simulations are available.

[1] Abbate J. et al. (2024) Phys. of Plasmas 31, 042506

[2] Abbate J. et al. (2025) Nucl. Fusion 65, 056014