Adaptive Multi-Fidelity neural Network: Benchmark tests and its application in Full-Field strain prediction for a wing box

Abstract

In engineering, high-precision numerical simulations and large-scale physical tests are often expensive and limited in number. Neural networks offer promising solutions, but traditional training methods rely heavily on abundant high-fidelity data, limiting their broader applicability. This paper proposes the Adaptive Multi-Fidelity Neural Network (Adp-MFNN), which adjusts the weights of the low- and high-fidelity networks adaptively for multi-fidelity prediction. Through extensive benchmarking on both purely numerical test cases and engineering problems, the potential of Adp-MFNN for multi-fidelity modeling has been preliminarily validated. Furthermore, a multi-fidelity framework employing two iterations of Adp-MFNN successfully extended sparse strain data from limited experimental load conditions to predict full-field strain under arbitrary conditions on a wing box. On one hand, the Adp-MFNN can effectively reduce the costs of large composite experiments. On the other hand, the Adp-MFNN method serves as a valuable complement to Digital Image Correlation (DIC) for large, complex, or internal structures, proving its effectiveness in engineering applications.