Near-field E-MIMO beamforming now predicts only stream powers: 1.11 ms direct inference, with feasibility reaching 100% on the retained test set after verification and repair
Synopsis
Beamforming for near-field extreme massive MIMO integrated sensing and communication usually requires repeatedly solving nonconvex constrained optimization, and online latency grows with array size. PATBO fixes the communication and sensing beam directions geometrically and lets a Transformer predict only the normalized stream powers, trained with Cramér–Rao bound (CRB) sensitivity weighting. On a retained test set with 1024 antennas, 4 users and 2 targets, direct inference reaches 19.54% joint feasibility, physics-constrained power repair raises it to 87.79%, and optimizer fallback for the remainder brings it to 100%, at 1.11 ms per sample.
Interpretation
Beamforming shifts from predicting the full complex beamforming matrix to predicting only the power split across communication and sensing streams, with beam directions constructed directly from geometry. Earlier learning-based designs mostly output high-dimensional beamforming variables directly, or focus on communication and sensing performance without explicitly handling the localization sensitivity introduced by near-field CRB constraints. With 1024 antennas, 4 communication users and 2 sensing targets, directly predicting the full complex beamforming matrix requires real-valued outputs, whereas PATBO predicts only six normalized stream powers.
The Transformer keeps a separate token for each user and target and applies no global pooling before the output layer, so each power variable stays tied to a specific entity. Pooled Transformer variants compress entity tokens into a single representation before generating stream powers, removing that one-to-one correspondence. On the same dataset and test split, token-preserving sensitivity-aware PATBO reaches 19.54% direct joint feasibility; the specific figures for the pooled variants and the MLP are not given in the supplied material, so the comparison stays qualitative.
Physical feasibility is checked explicitly after inference; infeasible predictions go through low-dimensional power repair, and those still infeasible go to a constrained optimizer. Prediction accuracy and constraint satisfaction are not equivalent, and samples near the CRB feasibility boundary can cross it even with small power errors. Of 560 directly infeasible samples, 475 recover feasibility through repair, raising joint feasibility from 19.54% to 87.79%; after fallback for the remaining 85, the retained test set reaches 100% feasibility.
Perspective
The evaluation uses 8 GHz, a 1024-antenna uniform planar array, 4 communication users and 2 sensing targets, with users and targets in the 8–12 m radiative near field; 4632 of 5000 random scenarios are retained and split 3242/694/696 into training, validation and test. The authors propose future work on dynamic mobility, channel uncertainty and real-time implementation.
The final 100% feasibility is an empirical result on the retained test set, not a universal guarantee for unseen realizations; for the 85 samples still infeasible after repair, the offline evaluation uses the stored feasible reference solution, while an online implementation would invoke the constrained optimizer. The compared SQP results represent the evaluated local numerical solver, not a globally optimal or universally feasible optimizer.