A data-augmented multi-fidelity deep learning method and its application on AUV drag prediction

The application of deep learning for performance prediction in complex engineering systems is often hindered by the prohibitive computational cost of acquiring sufficient high-fidelity data. While multi-fidelity modeling can address this challenge, the effective fusion of disparate data source remains a key issue. To tackle the problem, this paper introduces a novel multi-fidelity surrogate model, MFS-DNNDA, which synergistically combines a deep neural network architecture with a data augmentation strategy. The framework utilizes a Kriging model to obtain the augmented dataset by generating high-quality samples in high-confidence regions. A specialized neural network fuses multi-fidelity data with the augmented dataset to deliver accurate predictions. Rigorous evaluations against five state-of-the-art multi-fidelity models are conducted using 12 mathematical benchmark problems and an autonomous underwater vehicle drag prediction case. This result validates its potential as an efficient tool for accelerating engineering design and optimization.