Multi-fidelity modeling method based on adaptive transfer learning

Abstract

A wide range of problems in engineering applications require high-fidelity data to reflect real physical phenomena. However, only a small amount of high-fidelity data can be maintained under data acquisition cost constraints, which makes it difficult to meet the high-fidelity data needs of engineering application scenarios. This paper presents a multi-fidelity transfer modeling method, namely the Adaptive Transfer Learning Net (AtNet), aiming to suppress the overfitting phenomenon caused by the scarcity of high-fidelity data samples in multi-source data fusion. This method constructs multiple intermediate domains similar to the characteristics of the high-fidelity domain based on the minimum transfer loss theory, divides the transfer task into multi-stage transfer processes, and introduces rich and diverse feature information for the model to suppress the overfitting phenomenon. Finally, AtNet is applied to the multi-fidelity modeling task of numerical examples. Compared with traditional multi-fidelity modeling methods, AtNet not only maintains a high prediction accuracy but also can still effectively reflect the laws of high-fidelity data after further reducing the high-fidelity training data. In particular, AtNet is applied to multifidelity modeling scenarios involving SG6043 airfoil aerodynamics data and ONERA M6 pressure distribution data, comparative experiments were conducted by reducing high-fidelity training data. The results demonstrate that AtNet can approximate the true physical laws of aerodynamics using minimal high-fidelity training data, thereby reducing the cost of high-precision aerodynamic parameter design.