Geometry-Aware Conditional GAN for Sub-THz Channel Generation and Modeling in IIoT

Sub-terahertz (sub-THz) channel modeling for industrial Internet of Things (IIoT) is hindered by scarce measurements and strong scattering from metallic machinery. This letter proposes Geo-CGAN, a geometry-aware conditional GAN that synthesizes physically consistent power–delay–angular profiles (PDAPs) by injecting a geometric condition vector into both the generator and the discriminator. A two-stage transfer-learning strategy pretrains Geo-CGAN on abundant 28-GHz ray-tracing data and fine-tunes it on limited 220-GHz measurements with regularized adaptation. Measurements in a factory workshop show that Geo-CGAN achieves an average structural similarity index measure (SSIM) of 0.725. It also better matches the statistics of delay spread, angular spread, and path loss than TT-GAN and GAN-GRU, demonstrating proof-of-concept performance with limited measurements.