WorksEngineering
Physics-Aware Transformer-Based Beamforming Optimization for Near-Field E-MIMO ISAC Systems
Abstract
This paper proposes a physics-aware Transformer-based beamforming optimization (PATBO) framework for near-field extreme massive multiple-input multiple-output integrated sensing and communication (E-MIMO-ISAC) systems. PATBO combines regularized zero-forcing (RZF) communication directions and spherical-wave near-field sensing directions with learned stream-power allocation. A token-preserving Transformer maintains the correspondence between individual users/targets and their allocated powers, while a Cramér–Rao bound (CRB)-sensitivity-aware objective emphasizes power-allocation errors with greater localization impact. Physical feasibility is explicitly verified after inference, with infeasible predictions handled through physics-constrained power repair and optimizer fallback. For a E-MIMO array with antennas, the proposed structured representation reduces the learned output from real-valued coefficients of an unstructured complex beamformer to six stream-power variables. Direct PATBO achieves reference-rate retention and joint feasibility, which increases to after repair. The complete hybrid framework achieves reference-rate retention and constraint satisfaction on the retained test set, while requiring optimizer fallback for only of the scenarios. Direct PATBO inference requires ~ms per sample, corresponding to approximately and lower runtime than the evaluated cold-start and PATBO-warm-start sequential quadratic programming (SQP) implementations, respectively.
Introduction
Integrated sensing and communication (ISAC) is a key enabling technology for future sixth-generation (6G) wireless networks, allowing communication and environmental sensing to share spectrum, hardware, and waveform resources [1], [2]. In parallel, extreme massive multiple-input multiple-output (E-MIMO) provides large beamforming gains, high spatial resolution, and enhanced spatial multiplexing capability [3], [4]. Its enlarged physical aperture, however, significantly extends the radiative near-field region, where spherical-wave propagation introduces range-dependent spatial characteristics and enables joint angular–range beam focusing [5], [6]. These properties make near-field E-MIMO particularly attractive for simultaneous communication, sensing, and localization.
Recent studies have investigated near-field ISAC beamforming, localization, Cramér–Rao bound (CRB) analysis, wideband operation, and complexity-aware designs [7], [8], [9], [10], [11], [12], [13]. Related developments include modular extremely large-scale multiple-input multiple-output (XL-MIMO) hybrid beamforming, movable-antenna-assisted near-field ISAC, and practical ultra-wideband architectures [14], [15], [16]. Near-field beam acquisition, channel estimation, beam training, and localization have also received considerable attention [17], [18], [19], [20], [21]. These works demonstrate the benefits of near-field propagation, while also highlighting the increased complexity of beam management in large-aperture systems.
Near-field E-MIMO-ISAC requires the joint coordination of communication and sensing resources. Communication users require sufficient signal-to-interference-plus-noise ratio (SINR) and achievable rate, whereas sensing targets require adequate sensing SINR and localization accuracy. Existing ISAC beamforming methods therefore commonly employ iterative numerical optimization [22], [23], [24], [25], [26], [27], [28]. Although such methods provide effective benchmarks, repeatedly solving constrained nonconvex problems for changing user and target geometries can incur substantial online latency, particularly for E-MIMO systems with large-dimensional beamforming matrices.
Learning-based beamforming offers an alternative by mapping channel or geometry information directly to beamforming variables [29], [30]. Structured learning methods have further addressed constrained beamforming using penalty-dual, Transformer, graph-based, and related architectures [31], [32]. Learning-assisted ISAC beamforming has also been investigated through predictive, deep-unfolding, Transformer, and graph-learning approaches [33], [34], [35], [36], [37], [38]. These studies demonstrate the potential of learning to reduce the repeated computational burden of iterative beamforming optimization. However, most existing learning-based designs either predict high-dimensional beamforming variables directly or focus primarily on communication and sensing performance without explicitly addressing the strong localization sensitivity introduced by near-field CRB constraints. The joint consideration of near-field E-MIMO geometry, low-dimensional physically structured communication–sensing beamforming, CRB-sensitive learning, and explicit post-inference feasibility recovery therefore remains insufficiently addressed.
A further challenge is that prediction accuracy and physical feasibility are not equivalent in constrained ISAC beamforming. Accurately approximating a reference optimization solution does not necessarily guarantee satisfaction of coupled nonconvex constraints [31], [32]. This issue becomes particularly important in near-field localization, where the CRB depends on the transmit covariance, geometry-dependent steering-vector derivatives, and the resulting Fisher information matrix. Consequently, relatively small stream-power prediction errors can produce noticeable CRB degradation when the operating point lies close to the localization-feasibility boundary. A learning framework should therefore account not only for prediction accuracy but also for the physical consequences of prediction errors.
Direct prediction of the complete complex-valued beamforming matrix also becomes increasingly inefficient as the E-MIMO array size grows. For an -element array serving communication users and sensing targets, direct beamformer prediction requires complex, or equivalently real-valued, output variables. In contrast, exploiting physically meaningful communication and sensing directions allows the learning task to be restricted to the stream-power variables. Such a structured representation reduces the learning dimension while retaining the near-field spatial characteristics through physics-based beam construction.
To address these challenges, this paper proposes a physics-aware Transformer-based beamforming optimization (PATBO) framework for near-field E-MIMO-ISAC systems. PATBO constructs regularized zero-forcing (RZF) communication directions and near-field sensing directions from the system geometry, while the Transformer predicts only the normalized stream-power fractions. Geometry-aware entity tokens are processed using a token-preserving Transformer architecture that models inter-entity interactions without discarding the association between each user or target and its corresponding power variable. Safe-margin reference generation is employed to construct robust training targets, while CRB-sensitivity weighting emphasizes power-allocation errors having a stronger impact on localization performance. After inference, physical feasibility is explicitly verified; infeasible predictions are passed to a low-dimensional power-repair stage, and optimizer fallback is invoked when repair is unsuccessful.
The main contributions of this work are summarized as follows:
We develop a structured near-field E-MIMO-ISAC beamforming framework that combines RZF communication directions and spherical-wave near-field sensing directions with learned stream-power allocation. For , , and , this reduces the learned output representation from 12,288 real-valued beamforming coefficients to only six stream-power variables.
We propose a token-preserving Transformer with CRB-sensitivity-aware training. The architecture preserves the association between individual users/targets and their allocated powers, while emphasizing prediction errors having greater impact on near-field localization performance.
We develop a hierarchical verify–repair–fallback strategy for physical constraint satisfaction. Physics-constrained power repair increases joint feasibility from 19.54% to 87.79%, while the complete hybrid framework achieves 100% feasibility on the retained test set with optimizer fallback required for only 12.21% of the samples.
The remainder of this paper is organized as follows. Section II presents the near-field E-MIMO-ISAC system model and joint beamforming problem. Section III introduces the proposed PATBO framework. Section IV presents the numerical results, and Section V concludes the paper.
Notation: Boldface lowercase and uppercase letters denote vectors and matrices, respectively. The superscripts and denote the transpose and Hermitian transpose, respectively. The operators and denote the Euclidean and Frobenius norms, while denotes the absolute value of a scalar. denotes the identity matrix. The sets of real and complex matrices of size are denoted by and , respectively. The operators and denote the matrix trace and diagonal matrix construction, respectively, while denotes the real part of a complex-valued quantity.
System Model and Problem Formulation
Near-Field E-MIMO-ISAC System
We consider a downlink E-MIMO-ISAC system in which a base station (BS) equipped with an -element antenna array simultaneously serves single-antenna communication users and senses targets. The communication and sensing functionalities share the same transmit array, bandwidth, and transmit-power budget.
Let
denote the joint transmit beamforming matrix, where is the beamforming vector associated with communication user , and denotes the sensing beam associated with target .
The transmitted signal is
where and denote the normalized communication and sensing symbols, respectively. Assuming mutually uncorrelated unit-power streams, the total transmit power is . Accordingly, the transmit-power constraint is
where denotes the maximum available transmit power.
Near-Field Propagation Model
Due to the large physical aperture of the E-MIMO array, users and sensing targets may lie inside the radiative near-field region. Therefore, spherical-wave propagation is considered instead of the conventional plane-wave approximation [39]. Let denote the Cartesian position of antenna element , and let denote the position of a user or sensing target. The propagation distance between antenna and position is . A spherical-wave near-field response can be expressed as
where denotes the carrier wavelength. The amplitude term accounts for distance-dependent propagation attenuation, whereas the spherical-wave phase preserves the element-dependent propagation distance required for near-field focusing. Consequently, both range and angular information are retained in the array response [40].
For communication user , let denote the corresponding channel vector. The overall communication channel matrix is
Communication Performance
The received signal of communication user can be written as
where denotes additive noise with variance . Accordingly, the communication SINR is
The achievable rate of user is
and the system sum rate is
Each communication user is required to satisfy , .
Near-Field Sensing and Localization Model
In addition to serving the communication users, the E-MIMO transmitter simultaneously illuminates sensing targets. Let , , denote the three-dimensional Cartesian position of sensing target . Owing to the electrically large aperture of the E-MIMO array, the sensing targets may lie in the radiative near-field region. Therefore, the sensing model is constructed using position-dependent near-field steering vectors rather than conventional far-field angle-only steering vectors [7, 41].
For target , let denote its near-field steering vector. The corresponding normalized near-field focusing direction is constructed as
Unlike far-field beam steering, which primarily controls the angular direction of the transmitted beam, (10) exploits the spherical-wave propagation structure to focus the transmitted energy toward a specific three-dimensional position [6, 40].
Sensing SINR
The sensing quality of target is characterized by its sensing SINR. The dedicated sensing beam constitutes the desired sensing signal, whereas the sensing beams intended for the other targets are treated as inter-target sensing interference. Accordingly, the sensing SINR of target is given by
where denotes the sensing-noise power. In the considered sensing model, interference among the dedicated sensing streams is explicitly accounted for in (11).
To guarantee a minimum sensing quality for every target, the target-wise sensing-SINR constraint is
where denotes the minimum required sensing-SINR threshold. For performance evaluation, the worst-case sensing SINR over all targets is defined as
This metric determines whether all sensing targets simultaneously satisfy the required sensing-SINR condition.
Near-Field Localization CRB
In addition to the sensing-SINR requirement, localization accuracy is quantified through the CRB. Near-field sensing with extremely large arrays provides range-dependent geometric information that can be exploited for three-dimensional target localization, whose theoretical estimation accuracy can be characterized through the Fisher information matrix (FIM) and CRB [41].
The target location is specified by three Cartesian coordinates. Define the Jacobian of the near-field steering vector with respect to the target position as
In the numerical implementation, the spatial derivatives in (14) are evaluated using the central finite-difference approximation
where denotes the th Cartesian basis vector and is the perturbation step used in the numerical evaluation.
For a given joint beamforming matrix , define the transmit covariance matrix as . The FIM associated with the Cartesian position of target is then evaluated as
where denotes the effective target-reflection coefficient. The FIM is made numerically symmetric before evaluating the localization bound.
The Cartesian localization CRB for target is defined as
Hence, a smaller value of corresponds to greater position information and, consequently, a lower theoretical localization-error bound.
For robust numerical evaluation of (17), let , , denote the eigenvalues of the symmetrized FIM, with . A relative eigenvalue floor is introduced as , where in the numerical implementation. The effective eigenvalues are then defined as
Accordingly, the numerically robust CRB used in the simulations is evaluated as
For the sensing targets, the aggregate localization metric is defined as
The aggregate localization accuracy is constrained according to
where denotes the maximum allowable aggregate CRB. Consequently, the sensing component of the considered E-MIMO-ISAC system is governed by two complementary requirements: (12) guarantees sufficient target-wise sensing SINR, whereas (21) controls the three-dimensional localization accuracy. These two sensing requirements are incorporated jointly with the communication and transmit-power constraints in the beamforming optimization problem formulated next.
Joint Beamforming Problem
The general beamforming design objective is to maximize the communication sum rate while satisfying the communication QoS, sensing SINR, localization CRB, and total transmit-power requirements, consistent with the constrained beamforming formulations commonly considered for near-field ISAC systems [8], [13]. Before imposing the structured beam-direction parameterization adopted in this work, this design objective can be expressed in terms of the joint beamforming matrix as
Directly optimizing (22) over an unrestricted complex beamforming matrix would involve real-valued coefficients. In the proposed framework, however, both the numerical reference and PATBO employ the structured beam directions introduced in Section III. Consequently, the implemented optimization problem is reduced to the stream-power variables associated with these directions. PATBO further replaces repeated numerical optimization of these powers with learned prediction, followed by explicit physical verification and selective feasibility recovery.
Proposed Physics-Aware Transformer-Based Beamforming Optimization
Overview
The proposed PATBO framework separates the beamforming problem into two components: physically constructed beam directions and learned stream-power allocation. Rather than directly predicting the complete complex-valued beamforming matrix, the Transformer estimates only the normalized powers assigned to the communication and sensing streams.
For each scenario, PATBO performs four main operations:
geometry-aware token construction and Transformer-based power prediction;
physical communication and sensing beam-direction generation;
beamformer reconstruction and physical feasibility verification; and
adaptive physics-based repair and optimizer fallback when required.
The overall architecture of the proposed PATBO framework is illustrated in Figure 1. The near-field user–target geometry is first represented through geometry-aware tokens, which are processed by the PATBO Transformer to predict the normalized power allocation for the communication and sensing streams. The predicted powers are combined with physically constructed RZF communication directions and near-field sensing directions to form the structured beamformer. Its physical feasibility is then verified with respect to the communication QoS, sensing SINR, aggregate CRB, and transmit-power constraints. Infeasible predictions are passed to the physics-constrained power-repair stage, whereas feasible beamformers are directly accepted.

Figure 1. Proposed PATBO framework for near-field E-MIMO-ISAC combining token-preserving stream-power prediction, physics-based beamformer reconstruction, feasibility verification, and repair–fallback recovery.
Geometry-Aware Token Representation
Each communication user and sensing target is represented by an individual Transformer token. The communication-user token is defined as
where denotes the Cartesian position, denotes the distance from the BS reference point, and and denote the corresponding angular coordinates.
For sensing target , the token is
where , , and denote the standard deviations of the assumed target-position uncertainty along the Cartesian axes. The last token entry serves as a type indicator, where zero represents a communication user and one represents a sensing target. The complete Transformer input is
The range is , and the angles are and . For each feature dimension , the normalization parameters and are calculated exclusively from the training samples. The normalized feature is
The same training-derived statistics are applied to the validation and test sets.
Token-Preserving Transformer-Based Power Allocation
To capture the coupled interactions among communication users and sensing targets, PATBO employs a Transformer encoder based on multi-head self-attention. Transformer architectures are well suited to modeling dependencies among sequence elements and have recently been investigated for learning constrained wireless resource-allocation and beamforming mappings [42].
Each 10-dimensional entity token is first mapped into a dimensional latent representation. Let
denote the embedded token sequence.
The latent sequence is processed by two Transformer encoder stages, each employing four attention heads and a feed-forward dimension of 128. A dropout probability of 0.1 is applied within the network. After the final encoder stage, the resulting token representations are denoted as
Through self-attention, each token incorporates information from all communication users and sensing targets. Hence, the latent representation of a particular stream captures not only its own geometric characteristics, but also the coupled multi-user and multi-target interactions induced by the shared communication, sensing, localization, and transmit-power requirements.
A key feature of the proposed architecture is that the individual token representations are preserved after the Transformer encoders. In particular, no global pooling operation is applied across the token dimension. Instead, each final token representation is processed using a shared token-wise readout:
where consists of a 64-unit fully connected layer, a Gaussian error linear unit (GELU) activation, dropout, and a scalar output projection.
The resulting entity-specific logits are jointly normalized through a softmax operation:
Therefore, the predicted normalized stream-power vector is
which satisfies
The corresponding physical stream-power vector is
The token-preserving readout is particularly important in the considered problem because each output power corresponds to a specific physical entity. Global aggregation before the output layer can suppress this entity identity, whereas the proposed readout maintains the one-to-one association between the th communication-user or sensing-target token and its corresponding stream-power variable. At the same time, the preceding self-attention layers continue to model interactions among all entities.
Physics-Based Beam-Direction Construction
The communication beam directions are generated using RZF, which provides a practical tradeoff between multi-user interference suppression and noise enhancement [43]. The unnormalized RZF direction matrix is
where the regularization parameter is selected as
where denotes the nominal communication-power level used for constructing the RZF directions in the reference beamformer. Each communication direction is normalized as
For sensing target , the near-field sensing direction is given by the normalized focusing vector in (10). Thus, the complete direction matrix is
Beamformer Reconstruction
The Transformer-predicted beamformer is reconstructed as
Since every physical beam direction satisfies and the Transformer output satisfies (32), it follows that
Hence, under the unit-norm beam-direction construction, the total transmit-power requirement is structurally enforced by the simplex-constrained output parameterization.
This representation also significantly reduces the learned output dimension. Direct prediction of the complete complex beamforming matrix requires
real-valued outputs, whereas PATBO predicts only real-valued normalized stream powers. Therefore, the learned output-dimensionality reduction factor relative to an unstructured full-beamformer predictor is
For the considered , , and configuration, direct prediction of the complex beamforming matrix would require real-valued outputs, whereas PATBO predicts only six normalized stream powers. This corresponds to a reduction in learned output dimensionality.
This comparison concerns the neural-network output representation. The numerical reference optimizer employed for label generation also operates over the same structured stream-power variables while keeping the physical communication and sensing beam directions fixed.
Safe-Margin Reference Generation
A neural predictor trained using reference solutions located close to the CRB feasibility boundary may violate the original CRB requirement even when the corresponding power-allocation prediction error is small. To provide additional robustness, PATBO generates reference training labels using an adaptively tightened CRB requirement.
Let denote the minimum aggregate CRB obtained under the physically constructed beam directions while satisfying the communication-QoS, sensing-SINR, and transmit-power requirements. Let denote the aggregate CRB of the baseline structured beamformer. An adaptive operational CRB threshold is first constructed as
where the values 0.50, 0.70, 0.85, 0.95, and 1.00 are tested in order for until a feasible reference solution is obtained.
To provide a safety margin for training while preventing the tightened threshold from falling below the achievable CRB region, the safe threshold is defined as
Algorithm 1 Safe-Margin Reference Generation
Require: , target positions, , , , and
Ensure: ,
Construct the communication and sensing beam directions according to (34)–(37).
Minimize the CRB over the stream powers subject to communication QoS, sensing SINR, and transmit-power constraints to obtain .
Evaluate the aggregate CRB of the baseline structured beamformer.
Construct according to (42).
Construct according to (43).
Solve the constrained reference power-allocation problem under .
Normalize the optimized powers as .
return , .
where is the CRB safety factor and provides a small numerical buffer above the minimum achievable aggregate CRB . The inner maximum prevents the tightened threshold from falling below the feasible CRB region, while the outer minimum ensures that the safe threshold does not exceed the original operational threshold. The complete safe-margin reference-generation procedure is summarized in Algorithm 1.
CRB Sensitivity Estimation
A small power-allocation error can have substantially different physical consequences depending on the corresponding stream. PATBO therefore estimates the local sensitivity of the aggregate CRB to each normalized stream-power variable.
Let denote the aggregate CRB obtained using normalized power allocation . For stream , a local simplex-constrained directional sensitivity is approximated using a central finite difference:
where .
Following each positive or negative perturbation, the power vector is projected back onto the unit simplex to preserve and . Consequently, should be interpreted as a coupled local directional sensitivity under the fixed-total-power simplex constraint, rather than as an unconstrained partial derivative with respect to a single independent power variable. For numerical stability, the CRB gradient is normalized by the corresponding safe CRB threshold:
CRB-Sensitivity-Aware Training
Let denote the normalized reference stream power for entity in sample , and let denote the corresponding PATBO prediction. The stream-wise prediction error is
Although a conventional mean-squared-error objective treats all power deviations equally, power-allocation errors can have substantially different physical consequences in near-field localization. In particular, a relatively small perturbation of a CRB-sensitive stream may cause a larger localization penalty than a comparable perturbation of a less-sensitive stream.
To reflect this behavior during training, the CRB-sensitivity weight is defined as
where is the normalized simplex-constrained CRB sensitivity defined in (45), and controls the contribution of the physics-based weighting.
For a mini-batch containing samples, the proposed PATBO training objective is
Hence, the proposed objective assigns larger penalties to prediction errors associated with streams having stronger local impact on the aggregate localization CRB. The loss remains a weighted regression objective over the normalized stream powers, while the physically constructed communication and sensing beam directions are preserved outside the neural predictor.
The final PATBO formulation deliberately employs only the CRB-sensitivity weighting in the neural loss. The safe CRB threshold is retained during reference generation to construct robust training targets, but the safe-margin value is not used as an additional sample weight in the final training objective. The complete CRB-sensitivity estimation and PATBO training procedure is summarized in Algorithm 2.
Physical Feasibility Verification
After reconstructing the Transformer-predicted beamformer according to (38), its feasibility is evaluated using the original operational constraints rather than the tightened CRB threshold used for reference generation. For a candidate beamformer , define the feasibility indicator
where indicates that satisfies all original physical constraints, and otherwise. If , the Transformer output is directly accepted; otherwise, physics-constrained repair is invoked.
Physics-Constrained Power Repair
If the Transformer prediction violates one or more original constraints, PATBO invokes a low-dimensional physics-constrained power-repair stage. The physical beam directions remain fixed, while only the real-valued stream powers are optimized.
Algorithm 2 CRB-Sensitivity-Aware PATBO Training
Require: Training tokens , reference powers , safe CRB thresholds , perturbation , and sensitivity factor
Ensure: Trained PATBO parameters
for each training sample do
for do
Perturb by .
Project the positive and negative perturbed power vectors onto the unit simplex.
Evaluate the aggregate CRB for both perturbed allocations.
Compute the finite-difference sensitivity using (44).
Normalize the sensitivity using (45).
Compute using (47).
end for
end for
repeat
Predict normalized stream-power allocations for a mini-batch using the token-preserving Transformer.
Compute using (48).
Update the network parameters.
Evaluate the validation loss.
until the validation stopping criterion is satisfied
return Parameters corresponding to the best validation loss
Let denote the Transformer-predicted physical power vector. The repair problem is formulated as
Here, .
The repair problem therefore operates over only real-valued stream-power variables, while the physically constructed beam directions remain fixed. Consequently, the dimensionality of the repair optimization is independent of the number of antenna elements .
The first repair attempt is initialized as . If this attempt fails to recover feasibility, a second sensing-oriented initialization is employed. Let denote the communication-stream power. A fraction of is removed proportionally from the communication streams and distributed equally among the sensing streams. The repair problem is then solved again from this modified initialization.
In the implementation, the constrained power-repair problem is solved using sequential quadratic programming (SQP), a standard numerical approach for constrained nonlinear optimization [44]. Since repair is invoked only for infeasible Transformer predictions, its computational cost is avoided whenever direct neural inference already produces a feasible physical solution.
Hybrid PATBO Inference
Although physics-constrained repair can recover a substantial portion of infeasible Transformer predictions, a repair solution is not assumed to exist for every realization. Therefore, the final PATBO framework includes a constrained optimizer fallback for cases that remain infeasible after physics-based repair.
Let denote the repaired solution and denote a feasible solution returned by the constrained optimizer fallback. The final hybrid beamformer is selected according to
In the offline retained-test-set evaluation, the stored feasible reference solution is used as the fallback realization. Therefore, the reported 100% final test-set feasibility should be interpreted as an empirical result on the retained evaluation set rather than as a universal guarantee that every unseen realization can be recovered by the proposed direct or repair stages.
Hence, PATBO follows a hierarchical computational strategy. Transformer inference provides the lowest-latency solution. Physics-constrained power repair is activated only when necessary, while the reference optimizer is reserved for scenarios that cannot be recovered through low-dimensional repair. The resulting architecture therefore provides multiple operating modes with different latency–reliability tradeoffs. Direct Transformer inference offers the lowest online latency, whereas physics-constrained repair and reference optimization can be invoked when stronger feasibility guarantees are required.
An additional advantage of the structured formulation is its favorable scaling with the array size. Although increasing enlarges the physical communication and sensing beam directions, the Transformer continues to predict only stream-power variables. Hence, for fixed numbers of users and targets, the learned output dimension is independent of . This property is particularly relevant to E-MIMO systems, where direct prediction of the complete beamforming matrix becomes increasingly challenging as the array size grows. The proposed structure therefore retains the array-dependent near-field propagation characteristics in the physical beam construction while confining learning to the lower-dimensional resource-allocation problem.
Simulation Setup and Numerical Results
Simulation Setup
The proposed PATBO framework is evaluated using a UPA with antennas, communication users, and sensing targets at GHz. The antenna spacing is , and the users and targets are randomly located in the radiative near-field region at ranges of 8–12 m. The target-position uncertainty along each Cartesian dimension is independently generated between 0.01 and 0.10 m. The communication and sensing beam directions are constructed using RZF and spherical-wave near-field focusing, respectively. The remaining system parameters are summarized in Table I.
| Parameter | Value |
| Number of BS antennas, | 1024 |
| UPA configuration | |
| Number of communication users, | 4 |
| Number of sensing targets, | 2 |
| Carrier frequency, | 8 GHz |
| Antenna spacing, | |
| Total normalized transmit power, | 100 |
| Noise power, | 1 |
| Minimum user rate, | 1 bit/s/Hz |
| Minimum sensing SINR | 5 dB |
| User/target range | 8–12 m |
| Target-position uncertainty | 0.01–0.10 m |
| CRB safety factor, | 0.95 |
| Generated scenarios | 5000 |
| Retained valid scenarios | 4632 |
| Training/validation/test split | 3242/694/696 |
Table I. Simulation parameters
A total of 5000 random scenarios are generated. A realization is retained only when the structured reference optimization produces a valid solution satisfying the imposed physical constraints. This results in 4632 valid realizations (92.64%), which are divided into 3242 training, 694 validation, and 696 test samples. The safe-margin procedure of Section III-F uses ; over the retained dataset, the mean reference, safe-threshold, and original-threshold CRBs are 185.3078, 186.4777, and 191.8442, respectively.
The geometry-aware tokens are constructed according to (23)–(25), with feature standardization based exclusively on the training set. The final token-preserving PATBO configuration and training parameters are summarized in Table II.
| Parameter | Value |
| Input features per token, | 10 |
| Number of tokens, | 6 |
| Model dimension, | 64 |
| Attention heads | 4 |
| Feed-forward dimension | 128 |
| Transformer encoder layers | 2 |
| Token-wise hidden dimension | 64 |
| Dropout probability | 0.1 |
| Mini-batch size | 32 |
| Maximum epochs | 100 |
| Initial learning rate | |
| Early-stopping patience | 15 |
| CRB sensitivity factor, | 0.5 |
| Finite-difference perturbation, |
Table II. PATBO network and training parameters
Training and Power-Prediction Performance
The final token-preserving PATBO model is trained using the CRB-sensitivity-aware objective in (48). Early stopping
terminates training after 51 epochs, with the best validation loss of obtained at epoch 36.
On the independent 696-sample test set, the predicted normalized stream-power allocations achieve a mean-squared error of . Using the common sample-wise normalized-error metric adopted for the learning-baseline comparison, the corresponding power-allocation NMSE is dB. The softmax output satisfies the unit-simplex constraint numerically, with a maximum power-sum error of .
Figure 2 compares the average normalized stream-power allocation of PATBO with the reference solution. PATBO closely reproduces the overall communication–sensing power distribution, with most of the available power assigned to the communication streams and a smaller fraction allocated to the sensing.
Figure 2. Average normalized stream-power allocation of the reference solution and PATBO over the 696 retained test samples.
Nevertheless, agreement in average power allocation or conventional NMSE does not by itself imply physical feasibility. The localization CRB is highly sensitive to certain coupled perturbations of the stream-power vector, and therefore relatively small prediction errors can move a realization outside the feasible CRB region. This motivates the explicit constraint evaluation presented next.
Communication and Sensing Performance
Table III summarizes the performance of direct token-preserving PATBO before feasibility recovery. The reference and PATBO achieve average communication sum rates of 17.9204 and 18.0966 bit/s/Hz, respectively, corresponding to 100.98% rate retention. The slightly higher learned rate should not be interpreted as outperforming the constrained reference, since some PATBO predictions violate the localization constraint.
| Metric | Reference | PATBO |
| Sum rate (bit/s/Hz) | 17.9204 | 18.0966 |
| Rate retention (%) | 100.00 | 100.98 |
| Aggregate CRB | 181.9202 | 199.9954 |
| CRB change (%) | – | +9.94 |
| Communication QoS (%) | 100.00 | 100.00 |
| Sensing SINR (%) | 100.00 | 99.86 |
| Original CRB feasibility (%) | 100.00 | 19.54 |
| Safe CRB feasibility (%) | 100.00 | 13.79 |
| Power feasibility (%) | 100.00 | 100.00 |
| Joint feasibility (%) | 100.00 | 19.54 |
Table III. Performance and feasibility of direct PATBO
The reference and PATBO aggregate CRBs are 181.9202 and 199.9954, respectively, corresponding to a 9.94% increase. Communication QoS and transmit-power feasibility are 100%, and sensing-SINR feasibility is 99.86%. In contrast, only 19.54% of the direct predictions satisfy the original CRB constraint, which also limits the overall joint feasibility to 19.54%. The tighter safe CRB threshold is satisfied by 13.79% of the predictions.
These results identify localization CRB as the dominant feasibility bottleneck. Despite an NMSE of −11.1670 dB and near-reference communication performance, small errors in the allocated powers can significantly affect the FIM and CRB, both of which depend on geometry. Explicit post-inference feasibility verification is therefore required.
Constraint Feasibility and Physics-Constrained Repair
Physics-constrained repair is applied to the 560 direct PATBO predictions that violate at least one original constraint. The repair stage preserves the RZF communication and near-field sensing directions and optimizes only the stream-power variables according to (50).
Of the 560 repair attempts, 475 successfully recover feasibility, corresponding to a conditional success rate of 84.82%. Together with the 136 directly feasible PATBO predictions, this increases joint feasibility from 19.54% to 87.79%, as summarized in Table IV. The average relative power adjustment among successfully repaired samples is 17.403%. Thus, repair recovers most direct failures without altering the physics-based beam directions.
| Constraint | Direct PATBO | After Repair |
| Communication QoS (%) | 100.00 | 100.00 |
| Sensing SINR (%) | 99.86 | 100.00 |
| CRB (%) | 19.54 | 87.79 |
| Power (%) | 100.00 | 100.00 |
| Joint feasibility (%) | 19.54 | 87.79 |
Table IV. Constraint feasibility before and after repair.
Repair also improves the localization performance. The average aggregate CRB decreases from 199.9954 for direct PATBO to 190.5222 after direct acceptance and repair, reducing the CRB degradation relative to the reference from 9.94% to 4.73%. The corresponding average sum rate is 17.9573 bit/s/Hz (100.21% reference-rate retention), while the power-allocation NMSE improves from −11.1670 to −13.5352 dB.
Nevertheless, 85 of the 560 repair attempts remain infeasible. These realizations, corresponding to 12.21% of the complete test set, are subsequently handled by the optimizer-fallback stage.
Hybrid PATBO Performance
The complete hybrid PATBO framework combines direct inference, physics-constrained repair, and optimizer fallback. Among the 696 retained test realizations, 136 are directly feasible, 475 additional realizations are recovered through repair, and the remaining 85 require fallback. Consequently, joint feasibility progresses from 19.54% for direct PATBO to 87.79% after repair and finally to 100% after fallback, as illustrated in Figure 3.
Figure 3. Joint-feasibility progression across direct PATBO, physics-constrained repair, and optimizer fallback.
The final hybrid beamformer achieves an average sum rate of 17.9004 bit/s/Hz, corresponding to 99.89% reference-rate retention. Its average aggregate CRB is 184.4167, only 1.37% above the reference value of 181.9202. Thus, the CRB degradation decreases from 9.94% for direct PATBO to 4.73% after repair and to 1.37% for the complete hybrid output.
The corresponding operating-mode distribution is shown in Figure 4. Direct PATBO provides the final solution for 19.54% of the samples, successful repair handles an additional 68.25%, and only 12.21% require optimizer fallback. Hence, 87.79% of the retained test realizations are handled without invoking the fallback optimizer.
Figure 4. Distribution of PATBO inference operating modes over the 696-sample test set.
The reported 100% final feasibility is an empirical result on the retained test set rather than a universal guarantee. For the 85 realizations that remain infeasible after repair, the offline evaluation uses the stored feasible reference solution; an online implementation would instead invoke the constrained optimizer.
Computational Complexity and Runtime Analysis
PATBO replaces repeated numerical optimization with direct prediction of the stream-power variables while retaining the physically constructed beam directions. As established in Section III, this learned output dimension is independent of the antenna-domain beamformer dimension for fixed and .
Runtime is evaluated using the actual PATBO inference, SQP, and repair routines. Single-sample PATBO latency is measured after warm-up and averaged over repeated inference trials, while batched throughput is measured over the complete test set. Cold-start and PATBO-warm-start SQP are evaluated on a deterministic 100-sample subset using equal-power and PATBO-predicted initializations, respectively. The measured results for are summarized in Table V and Figure 5.
| Method | Mean Runtime | Outcome |
| PATBO, single sample | 1.1085 ms | Direct prediction |
| PATBO, batched | 0.0367 ms/sample | Throughput |
| Cold-start SQP | 0.9512 s | 39% feasible |
| PATBO-warm SQP | 0.8641 s | 50% feasible |
| Physics repair | 2.4011 s | 84% success |
Table V. Measured runtime of PATBO and numerical optimization
Figure 5. Mean per-sample runtime of PATBO inference, SQP, and physics-constrained repair.
Direct PATBO requires an average single-sample inference latency of 1.1085 ms, compared with 0.9512 s for cold-start SQP and 0.8641 s for PATBO-warm-start SQP. Thus, direct PATBO is approximately 858× and 779× faster than cold-start and PATBO-warm-start SQP, respectively. Batched PATBO further reduces the average processing time to 0.0367 ms/sample, although this value represents batch throughput rather than single-sample online latency. Physics-constrained repair requires 2.4011 s on average and is therefore employed selectively for feasibility recovery.
The outcome column in Table V complements the runtime comparison. Cold-start and PATBO-warm-start SQP achieve 39% and 50% feasibility, respectively, indicating that PATBO predictions can provide a useful initialization for numerical optimization. Physics-constrained repair successfully recovers 84% of the directly infeasible samples in the runtime benchmark. These SQP results represent the evaluated local numerical solver rather than a globally optimal or universally feasible optimizer. Overall, the results support the proposed hierarchy: low-latency direct PATBO inference is attempted first, while iterative physics-based optimization is invoked selectively when additional feasibility recovery is required.
Scalability with E-MIMO Array Size
To evaluate scalability, we vary the number of BS antennas while keeping communication users and sensing targets fixed. Four square UPA configurations, 20 × 20, 24 × 24, 28 × 28, and 32 × 32, are considered, corresponding to , 576, 784, and 1024, respectively. For each configuration, a new dataset is generated and PATBO is retrained using the same network architecture and training procedure. The corresponding Rayleigh distances are 13.54, 19.84, 27.34, and 36.04 m, maintaining the considered 8–12 m operating region within the near-field regime.
Figure 6 shows the measured runtime as a function of . Direct PATBO exhibits nearly constant inference latency, with mean runtimes of 1.131, 1.137, 1.101, and 1.109 ms for , 576, 784, and 1024, respectively. This behavior is enabled by the structured representation, where the Transformer predicts only stream-power variables irrespective of the number of antenna elements. In contrast, cold-start SQP runtime increases from 0.132 s to 0.951 s and PATBO-warm-start SQP from 0.122 s to 0.864 s as increases from 400 to 1024. Similarly, the physics-constrained repair runtime increases from 0.269 s to 2.401 s.
Figure 6. Runtime scalability of the proposed PATBO framework versus the number of BS antenna elements , including direct PATBO inference, PATBO-warm-start SQP, cold-start SQP, and physics-constrained repair.
The reference-rate retention remains above 100% for all evaluated array sizes. Although direct joint feasibility varies with the array configuration, physics-constrained repair substantially improves feasibility, while the complete verify–repair–fallback framework achieves 100% final feasibility for all considered values of . Since repair is invoked only for samples failing the direct feasibility check, its increasing runtime does not affect direct PATBO inference latency. Overall, the results demonstrate favorable computational scalability of the proposed framework for large-aperture near-field E-MIMO-ISAC systems.
Ablation and Baseline Analysis
To isolate the contributions of the proposed architecture and training objective, token-preserving sensitivity-aware PATBO is compared with an MLP and four Transformer variants under the same dataset, test split, and structured beamformer reconstruction. Table VI summarizes their direct-inference performance before repair or optimizer fallback.
| Method | NMSE (dB) | Rate Ret. (%) | CRB Change (%) | Sensing (%) | Joint Feas. (%) |
| MLP | −10.7953 | 101.18 | +11.47 | 98.13 | 13.94 |
| Vanilla Transformer | −10.8379 | 101.14 | +12.05 | 99.71 | 10.06 |
| Pooled sensitivity-aware Transformer | −10.8144 | 101.28 | +11.98 | 99.57 | 11.64 |
| Pooled margin-only Transformer | −10.8379 | 101.07 | +11.94 | 99.71 | 9.20 |
| Pooled sensitivity+margin Transformer | −10.8368 | 100.93 | +11.83 | 99.71 | 8.48 |
| Token-preserving sensitivity-aware PATBO | −11.1670 | 100.98 | +9.94 | 99.86 | 19.54 |
Table VI. DIRECT-INFERENCE ABLATION AND BASELINE COMPARISON
The proposed token-preserving PATBO achieves the highest direct joint feasibility of 19.54%, compared with 13.94% for the MLP and 11.64% for the pooled sensitivity-aware Transformer, as highlighted in Figure 7. Relative to the latter, this represents a 7.90 percentage-point improvement. PATBO also achieves the lowest power-allocation NMSE of −11.1670 dB and reduces the average CRB degradation from 11.98% for the pooled sensitivity-aware model to 9.94%.
Figure 7. Direct joint-feasibility comparison among the MLP baseline, Transformer ablations, and the proposed token-preserving sensitivity-aware PATBO over the 696-sample test set.
The improvement is consistent with preserving the association between each contextualized entity token and its corresponding power variable. In contrast, the pooled Transformer variants compress the entity tokens into a single representation before generating the stream powers, thereby removing this explicit correspondence.
The loss-function ablation further shows that sensitivity-aware training is more effective than margin-based weighting for the considered dataset. Within the pooled architecture, sensitivity-aware, margin-only, and combined sensitivity-and-margin training achieve 11.64%, 9.20%, and 8.48% joint feasibility, respectively. Accordingly, the final PATBO model uses the CRB-sensitivity-aware objective with , while the safe CRB threshold is confined to reference-solution generation.
Although all learning models retain approximately 100% of the reference communication rate, their CRB feasibility differs substantially. This confirms that communication-rate retention alone is insufficient for architecture selection and motivates the verify–repair–fallback mechanism used by PATBO.
Conclusion
This paper proposed PATBO, a physics-aware Transformer framework for near-field E-MIMO-ISAC beamforming that
combines physically structured beam directions with learned stream-power allocation. The token-preserving architecture and CRB-sensitivity-aware training enable low-dimensional power prediction while accounting for localization-sensitive constraints. Numerical results show that physics-constrained repair increases joint feasibility from 19.54% for direct PATBO to 87.79%, while the complete verify–repair–fallback framework achieves 99.89% reference-rate retention and 100% feasibility on the retained test set. Direct PATBO inference requires only 1.11 ms per sample, corresponding to approximately 858× and 779× lower runtime than the evaluated cold-start and PATBO-warm-start SQP benchmarks, respectively. Moreover, the scalability results show that PATBO maintains approximately 1.1 ms inference latency across –1024, while the numerical optimization stages exhibit increasing runtime with array size. These results demonstrate the potential of combining structured learning with selective physics-based optimization for low-latency and scalable near-field E-MIMO-ISAC beamforming. Future work will investigate dynamic mobility, channel uncertainty, and real-time implementation.
References
- [1]F. Liu, Y. Cui, C. Masouros, J. Xu, T. X. Han, Y. C. Eldar, and S. Buzzi, “Integrated sensing and communications: Toward dual-functional wireless networks for 6G and beyond,” IEEE J. Sel. Areas Commun., vol. 40, no. 6, pp. 1728–1767, Jun. 2022.DOI
- [2]M. Ali, A. A. Badrudeen, S. Baek, Y. Kim, and S. Kim, “Joint beamforming design under target reflected interference in RIS-assisted bi-static ISAC system,” in Proc. Int. Conf. Artif. Intell. Inf. Commun. (ICAII C), Tokyo, Japan, 2026, pp. 1259–1264.DOI
- [3]Z. Wang, J. Zhang, H. Du, D. Niyato, S. Cui, B. Ai, M. Debbah, K. B. Letaief, and H. V. Poor, “A tutorial on extremely large-scale MIMO for 6G: Fundamentals, signal processing, and applications,” IEEE Commun. Surveys Tuts., vol. 26, no. 3, pp. 1560–1605, 3rd Quart. 2024.
- [4]H. Lu, Y. Zeng, C. You, Y. Han, J. Zhang, Z. Wang, Z. Dong, S. Jin, C.-X. Wang, T. Jiang, X. You, and R. Zhang, “A tutorial on near-field XL-MIMO communications toward 6G,” IEEE Commun. Surveys Tuts., vol. 26, no. 4, pp. 2213–2257, 4th Quart. 2024.DOI
- [5]C. Ouyang, Z. Wang, Y. Chen, X. Mu, and P. Zhu, “A primer on near-field communications for next-generation multiple access,” Proc. IEEE, vol. 112, no. 9, pp. 1527–1565, Sep. 2024.arxiv.org/abs/2408.00952
- [6]X. Li, Z. Dong, and Y. Zeng, “Multi-user modular XL-MIMO communications: Near-field beam focusing pattern and user grouping,” IEEE Trans. Wireless Commun., vol. 23, no. 10, pp. 13766–13781, Oct. 2024.arxiv.org/abs/2308.11289
- [7]H. Wang, Z. Xiao, and Y. Zeng, “Cramér–rao bounds for near-field sensing with extremely large-scale MIMO,” IEEE Trans. Signal Process., vol. 72, pp. 701–717, 2024.arxiv.org/abs/2303.05736
- [8]K. Qu, S. Guo, N. Saeed, and J. Ye, “Near-field integrated sensing and communication: Performance analysis and beamforming design,” IEEE Open J. Commun. Soc., vol. 5, pp. 6353–6366, Oct. 2024.arxiv.org/abs/2308.06455
- [9]A. Ray, S. Majhi, H. Rajput, and M. Di Renzo, “ISAC for near-field and extremely large-scale uniform planar array systems,” IEEE Trans. Commun., vol. 74, pp. 1497–1512, 2026.DOI
- [10]H. Li, Z. Wang, X. Mu, Z. Pan, and Y. Liu, “Near-field integrated sensing, positioning, and communication: A downlink and uplink framework,” IEEE J. Sel. Areas Commun., vol. 42, no. 9, pp. 2196–2212, Sep. 2024.arxiv.org/abs/2311.07722
- [11]B. Zhao, C. Ouyang, Y. Liu, X. Zhang, and H. V. Poor, “Modeling and analysis of near-field ISAC,” IEEE J. Sel. Topics Signal Process., vol. 18, no. 4, pp. 678–693, May 2024.DOI
- [12]X. Wang, W. Zhai, X. Wang, M. G. Amin, and K. Cai, “Wideband near-field integrated sensing and communication with sparse transceiver design,” IEEE J. Sel. Topics Signal Process., vol. 18, no. 4, pp. 662–677, May 2024.DOI
- [13]Y. Lin, Z. Liu, J. Zhang, F. Liu, X. Li, Q. Zhang, Z. Wei, S. Fan, and J. Yan, “Near-field integrated sensing and communication beamforming considering complexity,” IEEE Trans. Veh. Technol., vol. 74, no. 4, pp. 6021–6034, Apr. 2025.DOI
- [14]C. Meng, D. Ma, Z. Wang, Y. Liu, Z. Wei, and Z. Feng, “Near-field hybrid beamforming design for modular XL-MIMO ISAC systems,” IEEE Trans. Commun., vol. 73, no. 11, pp. 11840–11854, Nov. 2025.DOI
- [15]J. Ding, Z. Zhou, X. Shao, B. Jiao, and R. Zhang, “Movable antenna-aided near-field integrated sensing and communication,” IEEE Trans. Wireless Commun., vol. 25, pp. 493–508, 2026.DOI
- [16]Y. Kim, S.-H. Park, S. Yang, K.-K. Wong, L. Dai, and C.-B. Chae, “Near-field challenges in ultra-wideband ISAC beamforming strategies and system insights,” IEEE Commun. Mag., pp. 1–7, May 2026.
- [17]H. Lei, J. Zhang, Z. Wang, B. Ai, and E. Björnson, “Near-field user localization and channel estimation for XL-MIMO systems: Fundamentals, recent advances, and outlooks,” IEEE Wireless Commun., vol. 32, no. 4, pp. 190–198, Aug. 2025.arxiv.org/abs/2407.10147
- [18]F. Zeinali, S. Javadi, M. R. Mili, M. B. Mashhadi, A. Al-Dweik, and N. Al-Dhahir, “Beamforming design for near-field fim-aided isac systems: A drl approach,” IEEE Trans. Veh. Technol., vol. 75, no. 8, pp. 18467–18472, Aug. 2026.DOI
- [19]G. Jiang and C. Qi, “Near-field beam training based on deep learning for extremely large-scale MIMO,” IEEE Commun. Lett., vol. 27, no. 8, pp. 2063–2067, Aug. 2023.DOI
- [20]X. Shi, J. Wang, X. Wang, C. You, and J. Song, “Double-sided near-field XL-MIMO: Beamfocusing codeword selection and channel estimation,” IEEE Trans. Commun., vol. 73, no. 5, pp. 3441–3455, May 2025.DOI
- [21]Q. Yang, A. Guerra, F. Guidi, N. Shlezinger, H. Zhang, L.-X. Yang, D. Dardari, B. Wang, and Y. C. Eldar, “Beam focusing for near-field multi-user localization,” IEEE Trans. Veh. Technol., vol. 74, no. 8, pp. 12259–12273, Aug. 2025.
- [22]Y. Cui, H. Ding, Y. Ma, X. Li, H. Zhang, and Y. Fang, “Energy-efficient integrated sensing and communication in collaborative millimeter wave networks,” IEEE Trans. Wireless Commun., vol. 24, no. 3, pp. 2341–2357, Mar. 2025.DOI
- [23]Y. Zhao, Q. Wu, W. Chen, Y. Zeng, R. Liu, W. Mei, F. Hou, and S. Ma, “Multi-functional beamforming design for integrated sensing, communication, and computation,” IEEE Trans. Commun., vol. 73, no. 8, pp. 6322–6336, Aug. 2025.arxiv.org/abs/2407.01006
- [24]S. Zhang, W. Hao, G. Sun, C. Huang, Z. Zhu, X. Li, and C. Yuen, “Joint beamforming optimization for active STAR-RIS-assisted ISAC systems,” IEEE Trans. Wireless Commun., vol. 23, no. 11, pp. 15888–15902, Nov. 2024.DOI
- [25]H. Hua, J. Xu, and R. Zhang, “Near-field integrated sensing and communication with extremely large-scale antenna array,” IEEE Trans. Wireless Commun., vol. 24, no. 12, pp. 9962–9977, Dec. 2025.arxiv.org/abs/2407.17237
- [26]M. Sun, Y. Gong, R. Zhang, Y. Zhao, X. Jing, C. Yuen, and D. W. K. Ng, “Robust spherical wavefront beamforming for near-field ISAC with MMSE optimization: A distance-angle perspective,” IEEE Trans. Commun., vol. 74, pp. 6718–6733, 2026.DOI
- [27]W. Mao, Y. Lu, C.-Y. Chi, B. Ai, Z. Zhong, and Z. Ding, “Communication-sensing region for cell-free massive MIMO ISAC systems,” IEEE Trans. Wireless Commun., vol. 23, no. 9, pp. 12396–12411, Sep. 2024.arxiv.org/abs/2307.16259
- [28]U. Demirhan and A. Alkhateeb, “Cell-free ISAC MIMO systems: Joint sensing and communication beamforming,” IEEE Trans. Commun., vol. 73, no. 6, pp. 4454–4468, Jun. 2025.arxiv.org/abs/2301.11328
- [29]X. Xia, G. Zheng, Y. Zhu, J. Zhang, J. Wang, and A. P. Petropulu, “A deep learning framework for optimization of MISO downlink beamforming,” IEEE Trans. Commun., vol. 68, no. 3, pp. 1866–1880, Mar. 2020.DOI
- [30]J. Zhang, W. Xia, M. You, G. Zheng, S. Lambotharan, and K.-K. Wong, “Deep learning enabled optimization of downlink beamforming under per-antenna power constraints: Algorithms and experimental demonstration,” IEEE Trans. Wireless Commun., vol. 19, no. 6, pp. 3738–3752, Jun. 2020.
- [31]Y. Li, Y.-F. Liu, F. Xu, Q. Shi, and T.-H. Chang, “Learning to optimize QoS-constrained beamforming in multi-user systems: A penalty-dual framework,” IEEE Trans. Wireless Commun., vol. 23, no. 11, pp. 16123–16138, Nov. 2024.DOI
- [32]Y. Li and Y.-F. Liu, “HPE transformer: Learning to optimize multi-group multicast beamforming under nonconvex QoS constraints,” IEEE Trans. Commun., vol. 72, no. 9, pp. 5581–5594, Sep. 2024.arxiv.org/abs/2402.16081
- [33]J. Zhang, M. Liu, J. Tang, N. Zhao, D. Niyato, and X. Wang, “Joint design for RIS-aided ISAC via deep unfolding learning,” IEEE Trans. Cogn. Commun. Netw., vol. 11, no. 1, pp. 349–361, Feb. 2025.DOI
- [34]Q. Gao, R. Zhong, Y. Zou, H. Shin, and Y. Liu, “Deep learning-based beamforming optimization for isac systems: A low-complexity and transferable framework,” IEEE Trans. Wireless Commun., vol. 25, pp. 8754–8768, 2026.DOI
- [35]X. Zhang, W. Yuan, C. Liu, J. Wu, and D. W. K. Ng, “Predictive beamforming for vehicles with complex behaviors in ISAC systems: A deep learning approach,” IEEE J. Sel. Topics Signal Process., vol. 18, no. 5, pp. 828–841, Jul. 2024.DOI
- [36]X. Xu, J. Xu, C. Dong, C. Xing, N. Zhao, Q. Wu, and D. Niyato, “Learning-based predictive beamforming for secure ISAC via IRS,” IEEE Trans. Commun., vol. 73, no. 12, pp. 14304–14315, Dec. 2025.DOI
- [37]M. Zhang, R. Zhong, X. Mu, H. Shin, and Y. Liu, “AI-enabled near-field communications: User movement prediction and beam tracking,” IEEE Trans. Cogn. Commun. Netw., vol. 12, pp. 6247–6264, 2026.DOI
- [38]P. Jiang, M. Li, R. Liu, and Q. Liu, “Graph learning for cooperative cell-free ISAC systems: From optimization to estimation,” IEEE Trans. Wireless Commun., vol. 25, pp. 13992–14008, 2026.arxiv.org/abs/2507.06612
- [39]M. Cui and L. Dai, “Channel estimation for extremely large-scale MIMO: Far-field or near-field?” IEEE Trans. Commun., vol. 70, no. 4, pp. 2663–2677, Apr. 2022.arxiv.org/abs/2108.07581
- [40]C. Wu, C. You, Y. Liu, L. Chen, and S. Shi, “Two-stage hierarchical beam training for near-field communications,” IEEE Trans. Veh. Technol., vol. 73, no. 2, pp. 2032–2044, Feb. 2024.arxiv.org/abs/2302.12511
- [41]H. Hua, J. Xu, and Y. C. Eldar, “Near-field 3D localization via MIMO radar: Cramér–rao bound analysis and estimator design,” IEEE Trans. Signal Process., vol. 72, pp. 3879–3895, 2024.
- [42]A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017.arxiv.org/abs/1706.03762
- [43]C. B. Peel, B. M. Hochwald, and A. L. Swindlehurst, “A vector-partition technique for near-capacity multiuser communication—part i: Channel inversion and regularization,” IEEE Trans. Commun., vol. 53, no. 1, pp. 195–202, Jan. 2005.
- [44]J. Nocedal and S. J. Wright, Numerical Optimization, 2nd ed. New York, NY, USA: Springer, 2006.





