An Improved Data Conditioning Method for Neural Network Modeling of Superconductor Critical Currents and n-Values

As high-temperature superconductors (HTS) are increasingly being considered for ac applications, it is essential to model the ac losses accurately. Conventional Kim-type critical current models often fail to capture the temperature dependence or high-field anisotropy of HTS materials. By contrast, artificial neural networks (ANNs) can capture parameter variations over large temperature and magnetic field ranges. However, training ANNs properly is challenging when the available experimental data are scarce or scattered. This article introduces an interactive method to transform raw experimental datasets into smooth and conditioned training sets. These extensive datasets are excellent for training ANN models of both critical current and power-law index parameters with standard machine learning tools. The method has been validated on three HTS products. All R-values exceed 0.96, and the models show no spurious behavior. Self-field fluctuations stay below $\mathbf {\pm }$0.71%, and there are no excessively steep partial derivatives. In comparison, ANN models trained directly on the raw data yield self-field fluctuations up to $\mathbf {\pm }$22.9%, and the partial derivatives reach unacceptable values for the n-value models. These findings indicate that ANN-based HTS models can improve modeling accuracy in ac applications by at least one order of magnitude.