Physical field-based device performance prediction method and device, server and medium

Abstract

This invention discloses a method, apparatus, server, and medium for predicting device performance based on physical fields, belonging to the field of device performance prediction technology. It includes: acquiring multiple sets of geometric parameters, real physical field distribution data, and real electromagnetic performance data of devices; training a sub-network using the real physical field distribution data as the first label to generate physical field prediction features; freezing the generated sub-network and training a prediction sub-network using the real electromagnetic performance data as the second label; in the prediction sub-network, the geometric main path generates coarse electromagnetic performance prediction features, and the image auxiliary path extracts low-dimensional features from the physical field prediction features, which are then fused through a gated weight mechanism to obtain a physical field correction amount, which is added to the coarse prediction to obtain a refined electromagnetic performance prediction result; using the trained model to predict the device to be predicted, and outputting the device performance prediction result. By introducing physical fields and constructing a decoupled collaborative architecture, high-precision device performance prediction and physical interpretability are achieved through gated weights.

Images