Parameter identification for IGBT behavioural models using a hybrid method of graph neural networks and differential evolution

To address the long-standing challenges of low efficiency and reliance on manual expertise in parameter identification for Insulated-Gate Bipolar Transistor (IGBT) behavioural models, this paper introduces a paradigm-shifting hybrid framework that unifies Graph Neural Network (GNN) and the Differential Evolution (DE) algorithm for parameter extraction. Based on a classical IGBT behavioural model, a large-scale synthetic dataset of characteristic curves were generated by perturbing six core parameters ([a0,a1, ...,b2]), establishing a rigorous foundation for training. A key innovation of this work is the transformation of the two-dimensional output characteristic curves into graph-structured data, enabling the construction of a GNN to directly learn the mapping from waveform graphs to model parameters. Subsequently, the prediction from the GNN is utilized as a high-quality initial solution for the DE algorithm to perform fine-grained optimization, effectively overcoming the convergence challenges faced by traditional optimization methods. After extensive training and cross-validation, the framework was validated using real experimental data, demonstrating order-of-magnitude improvements in efficiency and accuracy compared with traditional iterative fitting approaches. This work introduces a generalizable data-driven tool for automated semiconductor device modelling, bridging the gap between empirical waveform data and physics-informed behavioural models.