Abstract: To address the prohibitive per-iteration evaluation cost of evolutionary algorithms in Jiles-Atherton(J-A)hysteresis model parameter identification, a feature-directed surrogate-accelerated parameter identification method was proposed. A six-dimensional physical feature vector replaced hundreds of B-H loop sampling points as the identification target, eliminating inter-sample redundancy. A radial basis function(RBF)surrogate model with dimensionless-ratio input engineering was introduced to compress single-evaluation time to microsecond-scale inference. A feature-directed differential evolution(FD-DE)algorithm was proposed to exploit the RBF analytical gradient via weighted multi-feature direction with reverse annealing, that was, exploration dominated early search while gradient guidance was progressively introduced. Experiments on a J-A simulated dataset constructed from typical grain-oriented electrical steel parameter ranges showed R²>0.99 for all six features of RBF surrogate model, input engineering improved R² from 0.983 0 to 0.997 3, FD-DE accelerated convergence by 42% over standard differential evolution(DE), with all feature-level errors meeting hysteresis loop reconstruction requirements. The proposed method accommodated incomplete data and the surrogate model served multiple identification tasks after one training.

Key words: Jiles-Atherton model, parameter identification, surrogate model, differential evolution, directed mutation