A Stacked Neural-Network-Based Multi-Fidelity Approach to Solve Inverse Scattering Problems
Conference
Sendrea, RE, Zekios, CL, Georgakopoulos, SV. (2026). A Stacked Neural-Network-Based Multi-Fidelity Approach to Solve Inverse Scattering Problems
. 2015 IEEE INTERNATIONAL SYMPOSIUM ON ANTENNAS AND PROPAGATION & USNC/URSI NATIONAL RADIO SCIENCE MEETING, 1201-1204. 10.1109/AP-S/USNC-URSI60190.2026.11675645
Sendrea, RE, Zekios, CL, Georgakopoulos, SV. (2026). A Stacked Neural-Network-Based Multi-Fidelity Approach to Solve Inverse Scattering Problems
. 2015 IEEE INTERNATIONAL SYMPOSIUM ON ANTENNAS AND PROPAGATION & USNC/URSI NATIONAL RADIO SCIENCE MEETING, 1201-1204. 10.1109/AP-S/USNC-URSI60190.2026.11675645
This work proposes a multi-fidelity (MF) learning framework for solving inverse scattering problems under datalimited conditions. Specifically, the proposed approach employs a stacked neural network architecture that combines physicsbased inverse scattering strategies with a data-driven correction model inspired by autoregressive techniques commonly used in forward modeling. By leveraging low-fidelity reconstructions and the available measured scattered field as structured inputs, the MF framework enables efficient and accurate reconstruction of hidden dielectric profiles in both soft- and hard-scattering scenarios. The proposed model is validated using synthetic examples, achieving relative reconstruction errors below 10% in out-of-distribution scenarios.
