Abstract
Traditional electromagnetic (EM) scattering solvers typically require complete knowledge of an object's profile to predict resulting fields (forward scattering) or exhaustive measurement data to reconstruct an unknown object (inverse scattering). However, many practical scenarios involve "incomplete" profiles where only a portion of the geometry is known. This paper introduces a cascaded deep learning (DL) framework that bridges this gap. By utilizing two interconnected neural networks, the system first reconstructs the missing geometry and then calculates the full scattering characteristics. Results demonstrate that this approach significantly outperforms traditional inverse solvers in accuracy while providing a 60x speedup over conventional numerical methods.
Introduction
In computational electromagnetics, problems are generally categorized into forward and inverse scattering. Forward solvers predict fields from known geometries, while inverse solvers reconstruct geometries from measured fields. Real-world applications often encounter a hybrid scenario: a "known part" of the scatterer exists (e.g., a substrate or a partially scanned component), but the "unknown part" must be accounted for to determine the total scattering field.
Traditional forward solvers cannot operate with incomplete profiles, and traditional inverse solvers, such as the Subspace-based Optimization Method (SOM), often fail due to the ill-posed nature of the problem when measurement data is sparse. This research proposes a DL-based solution that uses limited "compensation" data—scattering from as few as one or two angles—to complete the object profile and predict the resulting fields.
Cascaded Deep Learning Architecture
The core of the proposed method is a cascaded architecture consisting of two sub-networks based on a 3-D U-Net design.
- Network-I (The Completer): This network functions as an inverse solver. It receives the known profile ($\chi^{p1}$), limited scattering data ($E^{sca}_0$), and the field generated by the known part ($E^{p1}$) to predict the unknown geometry ($\chi^{p2}$).
- Network-II (The Predictor): Acting as a forward solver, this network takes the combined profile (known + predicted) and calculates the change in the total electromagnetic field ($\Delta E^{tot}$).
A critical innovation in this scheme is the use of physics-augmented inputs. Instead of relying solely on the raw incident field, the authors derived formulations that include the field produced by the known part ($E^{p1}$). This provides the network with explicit information about the coupling between the known and unknown components, leading to more physically consistent predictions.
Experimental Validation
The model was trained on 15,000 samples generated via the Method of Moments (MoM). Datasets included complex 2-D and 3-D shapes derived from MNIST and EMNIST digits, as well as dielectric metasurfaces consisting of silicon nanoblocks on silica substrates.
Performance Metrics
The DL scheme demonstrated remarkable robustness and efficiency:
- Accuracy: In 2-D benchmarks, the system achieved a Mean Relative Error (MRE) of 9.32% for profile reconstruction and 2.68% for total field prediction.
- Computational Efficiency: The DL prediction for a 3-D problem took approximately 0.025 seconds, compared to 1.5 seconds for a traditional MoM solver—a 60-fold increase in speed.
- Generalization: The model successfully processed shapes and material contrasts not seen during training, indicating it learned underlying physical principles rather than simple pattern matching.
Discussion and Conclusion
The cascaded DL approach effectively regularizes the ill-posed inverse scattering problem by learning structural priors from training data. While traditional solvers like SOM exhibited errors as high as 68.27% under sparse data conditions, the DL framework maintained high precision.
Future work will focus on extending the model to handle multiple incident wave directions and determining the minimum observation data required to meet specific accuracy thresholds. This framework represents a significant step toward real-time EM analysis in complex, partially known environments.
Summary updated