A global-local attention network for uncertainty analysis of ground penetrating radar modeling

A global-local attention-based feature reconstruction (GLAFR) surrogate model is proposed for uncertainty analysis (UA) in ground penetrating radar (GPR) simulation. The uncertain inputs are converted to electric fields by the surrogate model instead of the full-wave simulation, and the uncertainty of output is quantified effectively. In the model, the global feature scaling (GFS) and the local feature reconstruction (LFR) are employed to obtain the long-term and short-term relationships of features. In addition, a new loss function is proposed to accelerate the convergence of the model for training data with a wider range of input disturbances. The validity of the surrogate model is verified by the UA result from the Monte Carlo method (MCM). Compared with existing deep learning methods, the proposed approach can efficiently get higher quality predictions. Meanwhile, the Sobol indices evaluated by GLAFR are in agreement with those of MCM which requires running the full-wave simulation one thousand times to converge.

Automated summary

A global-local attention-based feature reconstruction (GLAFR) surrogate model is proposed for uncertainty analysis (UA) in ground penetrating radar (GPR) simulation. The model effectively quantifies the uncertainty of the output by converting uncertain inputs to electric fields instead of using full-wave simulation. It employs global feature scaling (GFS) and local feature reconstruction (LFR) to capture long-term and short-term relationships of features, and a new loss function to accelerate convergence of the model. The validity of the surrogate model is verified by comparing with the UA result from the Monte Carlo method (MCM), and it outperforms existing deep learning methods in terms of prediction quality and Sobol indices evaluation.