GTL-NMDF: GAN-based transfer learning for non-hierarchical multi-fidelity data fusion

Multi-fidelity data fusion methods accurately simulate the responses of a physical system by fusing the data of different fidelity levels. However, many real-world engineering scenarios involve the non-hierarchical multi-fidelity data, consisting of a high-fidelity (HF) dataset and multiple low-fidelity (LF) datasets. In this paper, we propose a generative adversarial network (GAN)-based transfer learning model for non-hierarchical multi-fidelity data fusion (GTL-NMDF). The GTL-NMDF pre-trains its generator on each LF dataset and then fine-tunes it on the HF dataset by using the maximum mean discrepancy (MMD) distance between the HF samples and the non-hierarchical LF samples. In addition, three measurements based on the Minkowski difference are introduced to detect the GTL-NMDF’s training status, and serve as explainable loss functions for performance enhancement. Numerical experiments and engineering applications demonstrate the effectiveness and the superiority of the proposed GTL-NMDF method for handling non-hierarchical MDF tasks.