MERL-GPR is a simulated ground penetrating radar dataset generated using the open-source finite difference time domain tool for electromagnetic simulation gprMax. The dataset consists of 400 two dimensional underground structures with a domain size of 0.5m x 0.5m. The structures are composed of three layers, where the top layer is air with depth of 0.15m, and the bottom two layers are ground material with a total depth of 0.35m. The source is located 0.1m above the ground and emits a standard Ricker wavelet source with center frequency of 1GHz. The depth of the second ground layer d2 is sampled from a uniform distribution U(0.1, 0.3) and the depth of the first ground layer is d1 = 0.35 – d2. The first ground layer has a permittivity sampled from U(3, 5) and the second ground layer has permittivity sampled from U(5,10). Two cylinders are embedded int eh second ground layer. One cylinder is composed of air (permittivity 1), whereas the permittivity of the second cylinder is sampled from a uniform distribution U(3,10). Both cylinders have radii sampled from U(0.03, 0.06). From the time domain data generated by gprMax, we apply the Fourier transform and extract the wavefields with frequencies within the [0.5GHz, 1.5GHz] band discretized over 50 frequencies. At a Glance The size of the unzipped dataset is ~1.72GB The data directory contains both the freespace response as well as the total wavefield measured at every pixel in the computational domain. The complex frequency coefficients of the source wavelet are also provided. Two trained models are provided in the forward_model sub-directory, one model corresponds to the vanilla FNO architecture with 10 layers and the other model corresponds to the proposed BornFNO architecture also with 10 layers. A pretrained autoencoder is also provided under the priors sub-directory. Other Resources Pytorch code for training the models and solving the inverse problem is available at https://github.com/merlresearch/DeepBornFNO. Citation If you use MERL-GPR in your research, please cite our paper: @InProceedings{ Zhao_2023ICASSP, author = {Qingqing Zhao and Yanting Ma and Petros Boufounos and Saleh Nabi and Hassan Mansour}, title = {Deep Born Operator Learning for Reflection Tomographic Imaging}, booktitle = {Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, year = 2023, month = June } Copyright and License The MERL-GPR dataset is released under CC-BY-SA-4.0 license. All data: Created by Mitsubishi Electric Research Laboratories (MERL), 2023 SPDX-License-Identifier: CC-BY-SA-4.0