Parameter estimation of high-altitude electromagnetic pulse based on physics and data co-supervised neural network.

Area of Science:

• Computational physics
• Artificial intelligence
• Electromagnetics

Background:

• High-altitude electromagnetic pulses (HEMP) require accurate parameter estimation for analysis.
• Conventional artificial neural networks (ANNs) face challenges in accurately modeling complex physical phenomena like HEMP.
• Existing methods struggle with uneven sample distribution in input spaces, impacting generalization.

Purpose of the Study:

• To propose a novel physics and data co-supervised neural network (PaDCoSNN) for estimating HEMP mathematical parameters from physical parameters.
• To improve the accuracy and generalization ability of neural networks in modeling HEMP characteristics.
• To develop a more representative training sample generation strategy.

Main Methods:

• Developed PaDCoSNN with input/output layers reflecting numerical characteristics.
• Embedded nonlinear physical equations into the loss function for supervised learning.
• Implemented a dual-scale probabilistic sampling strategy for improved training data representativeness.

Main Results:

• PaDCoSNN reduced relative errors by an order of magnitude compared to conventional ANNs.
• Achieved processing times under 1.5 seconds.
• Demonstrated strong generalization on out-of-scope samples with relative errors below 0.00001, compared to ANNs up to 1.7.

Conclusions:

• PaDCoSNN offers superior accuracy and generalization for HEMP parameter estimation.
• The proposed methods, including structural innovations and sampling strategies, enhance neural network performance.
• PaDCoSNN shows potential for parameter estimation without traditional sample-based training, opening new avenues in HEMP analysis.