Work

Physics-guided multi-task calculation of ECRH power deposition based on machine learning

Year: 2026
Type: article
Abstract: Accurate calculation of the electron cyclotron resonance heating (ECRH) power deposition is essential for localized heating and current-drive applications in tokamaks, such as the suppression of neoclassical tearing modes. Such calculations are conventionally carried out with ray-tracing codes, which are physically reliable but computationally expensive for the repeated evaluations required by real-time control. In this work, a physics-guided multi-task surrogate is proposed for fast ECRH power-deposition calculation. The model is trained on a reference database generated by the ray-tracing code TORAY under J-TEXT operating conditions, while physics-guided soft penalties are introduced to guide the reconstruction of physically meaningful deposition features. Using plasma and launch parameters as inputs, the model jointly predicts the radial power-deposition profile and the deposition location, represented by the normalized radial coordinate and the corresponding major radius and vertical coordinates. A two-stage training strategy is adopted: a stable fit is first established in the logarithmic profile domain, and then refined by physics-guided soft penalties on physically meaningful profile characteristics. The predicted radial deposition profile reaches a coefficient of determination of R2 = 0.976, while the deposition location is recovered with mean absolute errors (MAE) of 7.6×10−3 in normalized radial coordinate, 0.132 cm in major radius and 0.141 cm in vertical position. These results indicate that the proposed model provides a fast TORAY-referenced surrogate for ECRH power-deposition prediction within the scanned J-TEXT operating range, with the physics-guided penalties improving the preservation of physically meaningful profile characteristics.
Language: English
Open Access status: bronze
OpenAlex ID: W7235559163
FWCI:
Cites: 16
Cited by:
Related to: 10