Provenance-preserved DL framework for intrinsically-isolated mm-wave CDRA MIMO

Abstract

Abstract This paper presents a provenance-preserving deep-learning framework for computational design screening and full-wave validation of high-isolation millimeter-wave cylindrical dielectric resonator antennas in a two-port MIMO configuration. The framework integrates Ansys HFSS, Python-based sampling, deep-learning prediction, and blockchain- and IPFS-based provenance within a reproducible computational workflow. An HFSS-Python pipeline generates 300 parametric CDRA geometries, each represented through binary image encoding and associated electromagnetic response data, including the reflection coefficient ( $$S_{11}$$ ) and the HFSS-exported total radiated electric-field response ( $$r_E$$ ). CNN, LSTM, CNN–LSTM, and multimodal Image–CNN–LSTM models are trained to predict radiation behaviour and rank candidate designs through pseudo-ECC-based screening. The selected CDRA is subsequently assessed through full-wave HFSS simulation in an orthogonal MIMO arrangement. At 27.96 GHz, the validated design achieves an $$S_{11}$$ of $$-20.79$$ dB, an $$S_{21}$$ of $$-50.13$$ dB, an ECC of $$7.78\times 10^{-8}$$ , diversity gain close to 10 dB, TARC of $$-21.09$$ dB, and CCL of 0.02185 bits s $$^{-1}$$ Hz $$^{-1}$$ . The orthogonal layout also improves isolation by about 20.31 dB compared with the parallel arrangement. The generated simulation and learning artefacts are further registered through a local Ethereum development network integrated with IPFS to support authenticity, traceability, and tamper-evident record keeping. The study demonstrates a practical computational route for automated and verifiable antenna-design screening, linking data-driven optimization with provenance-preserved management of electromagnetic design artefacts.