Inverse Design of Bonding Wire Array Based on Multifidelity Data-Enabled Neural Networks

Bonding wire arrays (BWAs) play an important role in multichip systems. In particular, the inductance,
$Q$
value, and loss of BWA are closely related to the performance of the chip systems. However, in industrial applications, full-wave numerical design approaches bring data burden and high computational costs, particularly for large-scale BWA. To alleviate the problem, we propose the multifidelity data-enabled neural networks for inversely designing BWA, which consist of an inversion network and a residual network. Specifically, we first introduce an approximate analytical model of BWA and perform low-fidelity data collection with the analytical model to train the inversion network. Subsequently, a small set of high-fidelity data is used to train the residual network, which is targeted to learn the residual between the analytical solutions and high-fidelity results. During the testing process, we cascade the two trained networks. Intensive numerical results are used to test the proposed method, where the average testing accuracy reaches 94.30% with 5000 tests. To demonstrate the effectiveness of the proposed method, the inversely designed BWA is fabricated and measured experimentally, where the measured results satisfy the performance demands. Furthermore, a power amplifier (PA) using bonding wires as matching inductors is designed, fabricated, and measured by using the proposed method.