Deep Robust Adaptive Beamforming via Element-Wise Manifold Calibration and Regularized Response Projection

Limited snapshots and element-wise gain–phase mismatch jointly impair covariance estimation and array manifold accuracy in uniform planar arrays. This paper proposes a deep robust adaptive beamforming framework that combines statistical base-weight generation, element-wise array manifold calibration, and regularized response projection. The base-weight network extracts finite-snapshot covariance information, whereas the calibration network estimates a physically bounded element-wise complex-gain vector from covariance features and nominal direction context. Phase-aligned auxiliary supervision makes the calibration loss invariant to the unidentifiable common phase and is required only during training. The calibrated steering vectors define a closed-form minimum-distance projection that preserves the normalized base weight’s desired direction response while suppressing the calibrated interference responses. Across three training seeds, the method achieves 23.43 ± 0.06 dB output SINR and a −52.48 ± 0.09 dB average null level, improving the former by 5.64 dB and deepening the latter by 5.11 dB relative to the best-performing baseline under the main test distribution. Experiments on mismatch severity, input SNR, snapshot number, direction-of-arrival errors, controlled ablations, and computational cost characterize the performance and limitations of the method under the stated synthetic-array model.