We present fast and accurate surrogate models that predict ion cyclotron resonance heating (ICRH) of a hydrogen minority in deuterium plasmas at ASDEX Upgrade (AUG) that can be used in fast transport modeling. Especially for medium-sized tokamaks with high power densities, such as AUG and SPARC, reliable ICRH modeling with full-wave codes like TORIC requires the coupling of a Fokker–Planck (FP) solver like SSFPQL, which is indispensable to account for the formation of a high-energy ion tail that can drastically alter the absorption behavior. For the first time, our surrogates reproduce the simulations provided by a coupled full-wave code TORIC, combined with a FP solver SSFPQL, demonstrating quantitative agreement not only on synthetic data but also on real AUG discharges that were entirely unseen during training. The models, implemented as feed-forward neural networks, capture collisional heating profiles for electrons and deuterium generated by the TORIC-SSFPQL codes, and reduce inference times from minutes to the order of s. To this end, we developed a complete machine learning pipeline that includes volumetric weighting of the radial heating profiles to emphasize physically relevant plasma regions during training. Finally, we integrate our Python-trained models into a C++ and Fortran-compatible Open Neural Network Exchange framework, demonstrating their suitability for cross-platform, real-time deployment in integrated modeling workflows.
Advances in numerical simulations of ion cyclotron heating of non-Maxwellian plasmas