An approach of calculating winding pack smeared properties for TF magnets using artificial neural network

Abstract

Tokamak superconducting TF coils experience severe structural and thermal loads during operation. Hence optimizing the design of the conductor and the coil is crucial, making it inevitable to repeatedly try different conductor and WP configurations, and then perform structural analyses for the magnet. However, due to the fact that the magnet is huge in size, meanwhile the WP constituents are heterogeneous, complex and small in dimension, it is generally unrealistic to perform a global 3D structural analysis on a magnet model with detailed WPs. Smeared WPs with equivalent homogeneous properties are therefore commonly used to build the finite-element model. Usually, the smeared properties of WP are calculated by finite-element analyses, which are time costly and cumbersome. Here the ANN approach is proposed to calculate the WP smeared properties. First, the conductor and WP configurations are summarized, and a universal representative WP model is developed. Then, 4322 training cases are calculated by finite-element analyses with ANSYS. Next, the ANN model is built and the training cases are fed into it, after 298 iterations the ANN achieves convergence with excellent performance. Finally, the effectiveness of the trained ANN is demonstrated. Our results indicate that the ANN approach of calculating the WP smeared properties is reliable and can reduce the calculation time by 5 order of magnitude than the finite-element analysis.