WorksComputer Science
Neural Network-Assisted CLEAN for Efficient Multipath Parameter Estimation in MIMO Systems
Abstract
High-resolution multipath estimation in wideband multiple-input multiple-output (MIMO) channels requires repeated searches over angle of arrival, angle of departure, and propagation distance. This paper presents neural network-assisted CLEAN (NN-CLEAN), which replaces the multidimensional search at each extraction step with a multi-head classifier while retaining least-squares (LS) estimation of the complex path gain, parametric reconstruction, and residual cancellation. The same classifier is applied to successive residuals, allowing the extraction depth to vary without changing its output structure. With the number of iterations prescribed by the reference multipath-component (MPC) count, off-grid synthetic tests show that NN-CLEAN slightly exceeds grid-search CLEAN (GS-CLEAN) in path accuracy for one and two MPCs, although its accuracy declines more rapidly at larger path counts. For complete three-MPC extraction at a batch size of eight, NN-CLEAN achieves an approximately speedup and a -fold reduction in peak GPU memory relative to GS-CLEAN. Applied without retraining to two-MPC line-of-sight (LOS) and non-line-of-sight (NLOS) Sionna ray-tracing channels within a coded MIMO link, NN-CLEAN substantially reduces channel-reconstruction mean-squared error (MSE) and bit error rate (BER) relative to interpolated LS channel estimation. GS-CLEAN retains an accuracy advantage on these channels, indicating the estimation cost associated with the computational savings.
Introduction
Accurate multipath channel characterization is essential to wideband multiple-input multiple-output (MIMO) communication, channel sounding, localization, and integrated sensing and communication (ISAC). Measurements across multiple antennas and frequencies enable individual multipath components (MPCs) to be distinguished through their spatial and frequency-dependent signatures. Estimating physically meaningful parameters, including the angle of arrival (AoA), angle of departure (AoD), propagation distance or delay, and complex path gain, is therefore important for channel reconstruction, beamforming, spatial multiplexing, and sensing. Reliable parameter extraction remains difficult in complex propagation environments, particularly at low signal-to-noise ratio (SNR), where noise and inter-path interference obscure weak components and reduce the separability of closely spaced paths [1, 2].
Conventional high-resolution parameter estimation is predominantly model based. Subspace methods such as MUSIC [3, 4] and ESPRIT [5, 6] exploit the structure of the signal and noise subspaces to estimate propagation parameters without an exhaustive likelihood search. Their performance, however, depends on reliable model-order selection and subspace separation. Correlated or coherent MPCs can reduce the effective signal-subspace rank, while low-SNR observations can cause subspace leakage or subspace swaps, resulting in substantial estimation degradation [7, 8].
Maximum-likelihood estimation (MLE) provides an alternative model-based formulation by identifying the propagation parameters that best explain the observed channel under an assumed physical model. Although MLE does not rely on the same subspace-separation conditions, joint estimation becomes computationally prohibitive when several parameters and MPCs must be optimized simultaneously. CLEAN [9] reduces the joint problem to a sequence of dominant-path estimation and subtraction steps. Each iteration, however, generally requires a multidimensional parameter search, causing the runtime and memory requirements to grow with the grid resolution, number of search dimensions, and channel-observation size [10, 11]. This scaling limits the applicability of grid-search CLEAN in systems requiring high-throughput or low-latency channel processing.
Deep learning (DL) has been investigated as a means of reducing the computational cost of multipath parameter estimation [12, 13, 14, 15, 16, 17, 18]. Existing approaches can be broadly divided into direct DL estimators and DL-assisted model-based methods. A direct DL estimator uses a neural network (NN) to map the received signal or channel observation to the desired propagation parameters in one forward pass. Although inference can be performed efficiently, simultaneous estimation of a complete multipath channel remains difficult. A fixed output structure must identify the active paths, preserve the association among the parameters belonging to each path, and remain valid when the path count differs from the training condition [18]. Classification-based estimators are further limited by their finite output grids and can degrade when the true parameters lie between the predefined grid points [17].
DL-assisted model-based methods instead replace selected computational operations within an established signal-processing algorithm. This approach retains the physical channel model and structured parameter updates while using learned inference to reduce the cost of the most demanding operation. Such a formulation is well suited to CLEAN, whose reconstruction and successive-cancellation steps are analytically defined but whose repeated multidimensional searches are computationally intensive. To the best of our knowledge, neural inference has not previously been integrated directly into the iterative CLEAN procedure for joint AoA, AoD, and propagation-distance estimation.
This work proposes NN-assisted CLEAN (NN-CLEAN), a hybrid multipath parameter estimator in which a multi-head classifier replaces the exhaustive parameter search at each CLEAN iteration. The remaining operations are model based: the complex path gain is obtained through least-squares (LS) projection, the estimated MPC is reconstructed from the parametric channel model, and the reconstructed response is subtracted from the current residual. The updated residual is then supplied to the same single-MPC classifier for the next extraction. Reusing one classifier across iterations permits different extraction depths without changing the NN output dimension.
The replacement of grid search with NN inference also changes the computational scaling under concurrent processing. Grid-search CLEAN must evaluate the complete candidate space for every channel realization, whereas NN-CLEAN can process multiple residual channels using batched tensor operations on parallel hardware. The computational benefit is therefore expected to become most significant when several channel realizations are processed simultaneously, as may occur in centralized MIMO and ISAC systems serving multiple communication links or sensing targets within the same processing interval.
NN-CLEAN is evaluated using both statistically synthesized channels and geometry-dependent channels generated with the NVIDIA Sionna ray-tracing (RT) package. The synthetic experiments examine off-grid estimation over different path counts, SNRs, antenna dimensions, and numbers of frequency samples. Runtime and peak memory are measured as the concurrent processing batch increases. The RT channels are incorporated into a coded MIMO communication link, where path accuracy, channel-reconstruction mean-squared error (MSE), and bit-error rate (BER) quantify how parameter-estimation errors propagate to downstream equalization and decoding.
A preliminary conference paper introduced the core NN-CLEAN concept under a controlled synthetic configuration [19]. This journal article extends the evaluation to off-grid channels with different path counts and SNRs, adds conditional parameter-precision analysis, and examines observation dimensions and search-grid resolution. GPU profiling distinguishes single-path search from complete multipath extraction and measures runtime and peak memory under increasing batch workloads. Two-MPC LOS and NLOS Sionna RT channels further assess reconstruction and BER within a coded MIMO link, while three-MPC LOS ablations examine label smoothing, angular conditioning, distance-loss weighting, and the training curriculum.
The main contributions of this work are summarized as follows:
NN-Assisted CLEAN Framework: We develop an NN-CLEAN estimator that replaces the exhaustive AoA–AoD–distance search within each CLEAN iteration with a multi-head neural classifier. The proposed method retains LS gain estimation, parametric channel reconstruction, and successive residual cancellation, providing a structured alternative to both exhaustive grid search and fully data-driven one-shot estimation.
Sequential Extraction with Variable Depth: The same single-MPC classifier is applied repeatedly to the evolving channel residual. This formulation supports a variable extraction depth without changing the network output dimension and reduces the cross-path association burden encountered by one-shot multipath estimators. Its robustness is evaluated on off-grid synthetic channels across different path counts and SNRs.
Computational and System-Level Validation: The proposed method is evaluated in terms of path accuracy, conditional parameter precision, channel-reconstruction fidelity, runtime, and peak memory. Geometry-dependent Sionna RT channels are further incorporated into a coded MIMO link, where channel-reconstruction MSE and BER are used to quantify the effect of parameter-estimation errors on downstream communication performance.
System Model and CLEAN Algorithm
Multipath Channel Model
Consider a quasi-static double-directional radio channel between an -element transmitter (Tx) uniform linear array (ULA) and an -element receiver (Rx) ULA, as illustrated in Figure 1. The channel comprises resolvable multipath components (MPCs), including a possible line-of-sight (LOS) component and reflected or scattered paths. Each MPC is characterized by its angle of arrival (AoA), angle of departure (AoD), propagation distance, and complex path gain.

Figure 1. Channel Model Visualization
The geometric parameters of an MPC are collected in
where , , and denote the azimuth AoA, azimuth AoD, and propagation distance, respectively.
Let , , denote the th frequency point. Following the double-directional channel formulation in [9], let and denote the calibrated vertically polarized beam-pattern responses of the Tx and Rx arrays, respectively. The spatial-frequency response associated with a candidate MPC is then defined as
where is the speed of light and .
The channel transfer function between the Tx and Rx arrays at frequency is modeled as the superposition of the individual MPC responses:
where and denote the geometric parameter vector and complex path gain of the th MPC, respectively. The transpose establishes the receive-by-transmit channel orientation, such that . The matrix represents AWGN with independent and identically distributed complex Gaussian entries.
MLE-Based MPC Extraction Using CLEAN
Under the AWGN assumption, multipath parameter estimation can be formulated as a deterministic maximum-likelihood estimation (MLE) problem. For a known number of MPCs, maximizing the likelihood is equivalent to minimizing the reconstruction error between the observed channel and the noise-free parametric component of (3). The joint estimator is therefore given by
where denotes the Frobenius norm. Directly solving (4) requires simultaneous optimization over the geometric parameters and complex gains of all MPCs, resulting in a high-dimensional nonlinear problem whose complexity increases rapidly with the path count.
CLEAN provides a practical alternative by decomposing the joint problem into a sequence of single-MPC estimation steps [9], [20]. At each iteration, the dominant component in the current residual is estimated, reconstructed, and subtracted before the next component is extracted. In this work, the CLEAN implementation that identifies each dominant MPC through an exhaustive discretized search over AoA, AoD, and distance is referred to as grid-search CLEAN (GS-CLEAN).
Let denote the residual channel at extraction iteration . The residual is initialized using the observed channel as
In GS-CLEAN, the AoA, AoD, and distance dimensions are discretized over , , and , containing , , and candidate values, respectively. The resulting joint search space is
Under the normalized array responses adopted in this work, the energy of the parametric response is constant across the search grid. Consequently, maximizing the concentrated single-path likelihood reduces to evaluating the spatial–frequency correlation between each candidate response and the current residual. The dominant MPC parameter vector at iteration is estimated as
where denotes the Frobenius inner product. After estimating , the corresponding complex path gain is obtained through least-squares (LS) projection:
The estimated MPC is then reconstructed and subtracted from the current residual according to
The parameter search, gain estimation, and residual subtraction in (7)–(9) are repeated for subsequent MPCs. Extraction may terminate after a prescribed number of components or when an external stopping criterion is satisfied, such as when the estimated path gain or residual energy falls below a predefined threshold.
Computational Bottleneck of GS-CLEAN
The computational cost of GS-CLEAN is dominated by the exhaustive parameter evaluation in (7). For a single candidate , evaluating the spatial–frequency correlation over observations of an channel requires approximately
operations. Because the AoA–AoD–distance search space contains candidate combinations, the resulting per-iteration complexity scales as
For comparable grid sizes of approximately points per parameter dimension, this dependence reduces to . Because the exhaustive search is repeated after each MPC subtraction, the total computational burden also increases approximately linearly with the number of extracted components for a fixed search space and observation dimension. Repeated computation can be reduced by precomputing candidate responses or array-manifold dictionaries. This strategy, however, introduces a memory requirement that grows with both the number of candidate combinations and the spatial–frequency observation dimensions. Increasing the grid resolution, antenna count, or number of frequency samples therefore increases the computational workload, the required dictionary storage, or both.
The limitation becomes more pronounced when additional propagation parameters are incorporated. The present formulation considers a three-dimensional search over azimuth AoA, azimuth AoD, and distance. Including Tx and Rx elevation angles would expand the search space to five dimensions, resulting in approximately complexity for comparable grid resolutions. Introducing Doppler for a time-varying channel adds another search dimension and further increases the number of candidate hypotheses. Consequently, exhaustive GS-CLEAN becomes progressively more computationally and memory intensive as the parameter space and observation dimensions increase.
Proposed NN-CLEAN Algorithm
NN-Assisted Iterative MPC Extraction
To address the multidimensional search bottleneck identified in Section II, we propose NN-CLEAN, which integrates neural network (NN) inference into the iterative CLEAN framework. As illustrated in Figure 2, the exhaustive parameter search in (7) is replaced by a direct mapping from the current residual channel to the geometric parameters of the dominant remaining MPC. The model-based path-gain estimation, parametric reconstruction, and residual-subtraction operations of conventional CLEAN are retained.

Figure 2. Overview of the proposed NN-CLEAN framework.
At the th iteration, the residual channel observations are supplied directly to the NN, which produces the dominant MPC parameter estimate . Given this estimate, the corresponding complex path gain is computed using the LS projection in (8). The estimated MPC is then reconstructed and removed from the residual according to (9), after which is supplied to the same network for the next extraction iteration. NN-CLEAN therefore replaces only the computationally intensive grid-search operation while preserving the physical reconstruction and successive-cancellation structure of CLEAN.
An iterative single-MPC architecture is adopted instead of a one-shot network that estimates all MPCs simultaneously. A one-shot formulation must maintain the association among the AoA, AoD, and distance estimates belonging to each MPC, and this association becomes increasingly difficult as the path count increases. Moreover, a one-shot estimator generally requires a predefined output structure for the complete multipath set, making it less suitable for variable path counts. In contrast, NN-CLEAN repeatedly applies the same single-MPC estimator to the evolving residual, allowing the extraction depth to be prescribed or controlled by an external CLEAN stopping criterion without changing the NN output dimension.
Proposed NN Architecture
The NN architecture used for MPC parameter estimation is illustrated in Figure 3. A shared feature-extraction backbone processes the residual channel and learns a common representation of its spatial–frequency characteristics. The extracted representation is then provided to parameter-specific classification heads for AoA, AoD, and distance estimation.

Figure 3. Multi-head residual classification architecture used in NN-CLEAN.
The shared backbone comprises pooling and residual blocks that progressively extract the relevant channel features while reducing the intermediate feature dimension. The resulting representation is supplied to the AoA and AoD classification branches. The distance branch additionally employs a fully connected (FC) gating mechanism conditioned on the angular outputs, allowing information from the estimated angular dimensions to contribute to distance classification. Rather than
treating every element of the Cartesian search space as a separate class, the network produces three unnormalized class-logit vectors at extraction iteration :
Each logit vector assigns a classification score to the candidate grid points of its corresponding parameter dimension. The estimated class index for each output head is selected directly from the logits according to
The resulting estimate follows the same parameterization as the estimate produced by GS-CLEAN and can therefore be used directly in the path-gain estimation and residual-subtraction operations in (8) and (9). The multi-head formulation also avoids constructing a single classifier over all joint grid combinations. Instead, the network produces a total of output logits.
Computational Characteristics
The principal computational distinction between GS-CLEAN and NN-CLEAN lies in the estimation of . At every extraction iteration, GS-CLEAN evaluates joint parameter combinations, whereas NN-CLEAN obtains the three parameter estimates through a single forward pass with factorized classification outputs. NN-CLEAN consequently avoids explicit evaluation and storage of the complete multidimensional candidate space.
NN inference primarily comprises tensor operations that can be efficiently parallelized across multiple channel realizations on modern accelerators. The resulting runtime and memory characteristics are evaluated experimentally in § 6 across different processing batch sizes and system configurations.
Simulation and Evaluation Methodology
The evaluation uses the known number of MPCs retained in the reference channel, , to prescribe the extraction depth of both NN-CLEAN and GS-CLEAN. Only the path count is supplied; the geometric parameters and complex gains are estimated from the channel observation. Unequal reference and estimated path counts would require additional association rules for unmatched components, while reconstruction MSE and BER would combine model-order and parameter-estimation errors. Using a common extraction depth therefore focuses the comparison on parameter estimation and successive cancellation, with incorrect path selections and their resulting reconstruction errors still included.
When the path count is unknown, the extraction depth must be determined from residuals containing physical MPCs, observation noise, and cancellation artifacts. The appropriate stopping criterion depends on SNR, relative path powers, and estimator-dependent residual statistics, making a fixed threshold difficult to apply across operating conditions. Adaptive stopping criteria and their effects on missed and spurious detections, reconstruction quality, BER, and computational cost are left for future work.
Synthetic MIMO Configuration
The proposed NN-CLEAN framework is first evaluated using synthesized wideband MIMO channels following the parametric model in Section II. Unless otherwise specified, a MIMO configuration with is considered over a GHz bandwidth centered at GHz. The channel is sampled at uniformly spaced frequency points. Both the Tx and Rx employ ULAs with half-wavelength inter-element spacing evaluated at and ideal vertically polarized omnidirectional elements with negligible mutual coupling. The common system and search-space parameters are summarized in Table 1.
| Parameter | Setting |
| Center frequency, | 2.5 GHz |
| Bandwidth, | 1 GHz |
| Subcarriers, | 512 |
| Tx/Rx elements, , | 4, 4 |
| AoA/AoD, , | , resolution |
| Distance, | m, 0.05 m resolution |
Table 1. Common Simulation Parameters
For an -element ULA, the frequency-dependent steering vector at azimuth angle and frequency is
For grid-based estimation and NN classification, the parameter space is discretized as follows:
AoD () and AoA (): with spacing, giving .
Distance (): m with 0.05 m spacing, giving .
The complex path gains are generated independently according to . Let denote the noise-free component of the synthetic channel in (3), such that . The synthetic channel-observation SNR is defined as
Training and testing use different parameter distributions. During training, the AoA, AoD, and distance values are restricted to the predefined grids, corresponding to an on-grid configuration. During synthetic evaluation, these parameters are sampled continuously over the same physical ranges, resulting in an off-grid test condition. Consequently, both NN-CLEAN and GS-CLEAN are evaluated in the presence of grid mismatch rather than only at parameter values represented exactly by the search dictionary.
NN Training Procedure
Training samples are synthesized on the fly during optimization. For each sample, the path count is selected from , and the corresponding AoA, AoD, and distance parameters are drawn from their respective on-grid sets. The synthetic channel-observation SNR is independently sampled from dB. The network parameters are optimized using Adam with a learning rate of and a batch size of 128.
Because NN-CLEAN extracts MPCs sequentially, the ground-truth paths are ordered in descending magnitude of @page 5/13 starts_mid=1 ends_mid=1 be used directly in the path-gain estimation and residual-subtraction operations in (8) and (9). The multi-head formulation also avoids constructing a single classifier over all joint grid combinations. Instead, the network produces a total of output logits.
Computational Characteristics
The principal computational distinction between GS-CLEAN and NN-CLEAN lies in the estimation of . At every extraction iteration, GS-CLEAN evaluates joint parameter combinations, whereas NN-CLEAN obtains the three parameter estimates through a single forward pass with factorized classification outputs. NN-CLEAN consequently avoids explicit evaluation and storage of the complete multidimensional candidate space.
NN inference primarily comprises tensor operations that can be efficiently parallelized across multiple channel realizations on modern accelerators. The resulting runtime and memory characteristics are evaluated experimentally in § 6 across different processing batch sizes and system configurations.
Simulation and Evaluation Methodology
The evaluation uses the known number of MPCs retained in the reference channel, , to prescribe the extraction depth of both NN-CLEAN and GS-CLEAN. Only the path count is supplied; the geometric parameters and complex gains are estimated from the channel observation. Unequal reference and estimated path counts would require additional association rules for unmatched components, while reconstruction MSE and BER would combine model-order and parameter-estimation errors. Using a common extraction depth therefore focuses the comparison on parameter estimation and successive cancellation, with incorrect path selections and their resulting reconstruction errors still included.
When the path count is unknown, the extraction depth must be determined from residuals containing physical MPCs, observation noise, and cancellation artifacts. The appropriate stopping criterion depends on SNR, relative path powers, and estimator-dependent residual statistics, making a fixed threshold difficult to apply across operating conditions. Adaptive stopping criteria and their effects on missed and spurious detections, reconstruction quality, BER, and computational cost are left for future work.
Synthetic MIMO Configuration
The proposed NN-CLEAN framework is first evaluated using synthesized wideband MIMO channels following the parametric model in Section II. Unless otherwise specified, a MIMO configuration with is considered over a GHz bandwidth centered at GHz. The channel is sampled at uniformly spaced frequency points. Both the Tx and Rx employ ULAs with half-wavelength inter-element spacing evaluated at and ideal vertically polarized omnidirectional elements with negligible mutual coupling. The common system and search-space parameters are summarized in Table 1.
| Parameter | Setting |
| Center frequency, | 2.5 GHz |
| Bandwidth, | 1 GHz |
| Subcarriers, | 512 |
| Tx/Rx elements, , | 4, 4 |
| AoA/AoD, , | , resolution |
| Distance, | m, 0.05 m resolution |
Table 1 (PDF p. 5). Common Simulation Parameters
For an -element ULA, the frequency-dependent steering vector at azimuth angle and frequency is
For grid-based estimation and NN classification, the parameter space is discretized as follows:
AoD () and AoA (): with spacing, giving .
Distance (): m with 0.05 m spacing, giving .
The complex path gains are generated independently according to . Let denote the noise-free component of the synthetic channel in (3), such that . The synthetic channel-observation SNR is defined as
Training and testing use different parameter distributions. During training, the AoA, AoD, and distance values are restricted to the predefined grids, corresponding to an on-grid configuration. During synthetic evaluation, these parameters are sampled continuously over the same physical ranges, resulting in an off-grid test condition. Consequently, both NN-CLEAN and GS-CLEAN are evaluated in the presence of grid mismatch rather than only at parameter values represented exactly by the search dictionary.
NN Training Procedure
Training samples are synthesized on the fly during optimization. For each sample, the path count is selected from , and the corresponding AoA, AoD, and distance parameters are drawn from their respective on-grid sets. The synthetic channel-observation SNR is independently sampled from dB. The network parameters are optimized using Adam with a learning rate of and a batch size of 128.
Because NN-CLEAN extracts MPCs sequentially, the ground-truth paths are ordered in descending magnitude of their complex gains. The target at extraction iteration therefore corresponds to the th strongest MPC. Gaussian label smoothing (GLS) [21] is applied to account for the correlation among neighboring parameter bins. For parameter head , the GLS target at class index is
where is the ground-truth class index of the th extracted MPC, is the number of classes associated with head , and is the smoothing parameter.
Algorithm 1 Three-Phase Training Procedure for NN-CLEAN
Initialize: Model weights and Adam optimizer with lr
for each training step to do
Sample dB
Sample
Generate
Sample
Synthesize and add AWGN to obtain
Sort the MPCs in descending order of
Generate the GLS targets using (16)
if then
// Phase 1: Dominant-path estimation
Evaluate using the Phase 1 case of (19)
else if then
// Phase 2: Teacher-forced extraction
Extract paths using ground-truth residual subtraction
Evaluate using the Phase 2 case of (19)
else
// Phase 3: Autoregressive extraction
Extract paths using predicted parameters and LS gain estimates
Evaluate using the Phase 3 case of (19)
end if
Backpropagate and update
end for
At each extraction iteration, the network operates on a different residual channel. A separate classification loss is therefore evaluated for each residual state encountered during sequential extraction. Let denote the GLS target vector and denote the corresponding unnormalized logit vector. The categorical cross-entropy (CCE) for parameter head is
The loss associated with extraction iteration is defined as
A weight of 3 is applied exclusively to the distance loss because a distance error produces a frequency-dependent phase mismatch in the reconstructed MPC. The resulting coherent residual can propagate through subsequent CLEAN iterations and impair the extraction of weaker components.
The final training loss depends on the curriculum phase:
Training comprises optimization steps divided into three phases. Phase 1 uses the first 50% steps to train dominant-path estimation. Phase 2 uses the next 25% steps to introduce sequential extraction with ground-truth residual subtraction. Phase 3 uses the final 25% steps to generate the residual sequence from the predicted parameters and their LS gain estimates. Thus, Phases 2 and 3 use the same loss aggregation in (19), but differ in how the residual channels are formed. During Phase 3, the hard parameter decisions and residual updates are treated as stop-gradient operations.
Evaluation Protocol
Baselines and Estimation Metrics
The primary synthetic evaluation is conducted entirely under off-grid testing, with the ground-truth AoA, AoD, and distance drawn continuously within the ranges in Table 1. NN-CLEAN is compared with GS-CLEAN [9], 3D-MUSIC [4], 3D-ESPRIT [6], and a standalone one-shot NN estimator. GS-CLEAN serves as the exhaustive grid-search reference, while the subspace methods are evaluated only for path counts supported by the available signal-subspace dimension.
The one-shot baseline uses the same network architecture, parameter grids, loss weights, and optimization settings as NN-CLEAN. A single model is trained with varying MPC counts from . For each parameter head, the training target is formed by summing Gaussian kernels centered at the ground-truth bins of all MPCs and normalizing the resulting vector. Training uses the original multipath observation throughout, without the sequential residual updates or three-phase curriculum of NN-CLEAN. After a single forward pass, the highest-scoring bins are selected independently from each head using the known reference path count. Ideal association among the selected AoA, AoD, and distance values is assumed during evaluation.
For NN-CLEAN and GS-CLEAN, the reference MPCs are ranked in descending order of their complex-gain magnitudes. The MPC estimated at extraction iteration is compared with the th reference MPC in this ordering. Unless otherwise specified, a paired MPC is counted as correct only when its AoA, AoD, and distance estimates all lie within corresponding grid intervals of the ground truth. Accuracy is defined as
where and denote the numbers of correctly estimated and ground-truth MPCs, respectively.
The precision of successfully associated MPCs is further characterized using the angular and distance multipath-component distance (MCD). With the angles expressed in radians, the angular MCD is
and the normalized distance MCD is
For either metric, the root-mean-square error (RMSE) over the successfully associated MPCs is
The fidelity of the synthetic channel reconstruction is additionally characterized by the energy-captured ratio
where the Frobenius norms are evaluated over the complete spatial–frequency channel. Using the noise-free reference prevents AWGN energy from being interpreted as recoverable multipath energy.
2) Runtime and Memory Profiling: The computational comparison between NN-CLEAN and GS-CLEAN uses synthesized channels under the baseline , configuration on an NVIDIA GeForce RTX 4070 GPU with 12 GB of device memory. Runtime and peak GPU memory are measured as the processing batch size increases from 1 to 8 for two workloads. The one-MPC workload measures the parameter-search call from the observed channel to the estimator outputs, using NN inference for NN-CLEAN and exhaustive AoA–AoD–distance search for GS-CLEAN. The three-MPC workload measures the end-to-end extraction call, including parameter search, LS gain estimation, path reconstruction, and residual updates across three sequential iterations. Untimed
| Search grid | Angular resolution | Distance resolution |
| Very coarse | 0.20 m | |
| Coarse | 0.10 m | |
| Baseline | 0.05 m | |
| Fine | 0.025 m | |
| Very fine | 0.0125 m |
Table 2. Search Grids Used in the Resolution Sweep
warm-up runs are performed before measurement to exclude graph tracing and one-time initialization overhead.
Batch size specifies the number of independent channel realizations processed concurrently within the same estimation call, and the reported runtime covers the complete batch. Such a workload may arise in a centralized MIMO or ISAC processor handling observations from multiple users or vehicles within the same processing interval. Each sweep keeps the path count fixed at one or three MPCs per realization, so increasing batch size changes the concurrent workload while preserving the extraction depth.
3) Observation-Dimension and Search-Grid Resolution Sweeps: The dependence of path accuracy on the observation dimensions and search-grid resolution is evaluated using off-grid channels at dB, with AoA, AoD, and distance sampled continuously in each sweep. The antenna sweep varies from 4 to 8 with , and the subcarrier sweep varies with ; both use the baseline search grid. The search-grid resolution sweep holds and and uses the five grids in Table II. The remaining system parameters follow Table I.
A separate NN-CLEAN model is trained for each configuration following Algorithm 1. The antenna and subcarrier sweeps change its input dimension, whereas the search-grid resolution sweep changes only the sizes of its AoA, AoD, and distance classification outputs and their directly connected layers. The remaining architecture and training procedure are unchanged. For the corresponding GS-CLEAN evaluations, the array and frequency responses are regenerated for each observation dimension, and the search dictionary is constructed at each grid resolution. Both methods are evaluated using the path accuracy defined in (20). The antenna and subcarrier sweeps use the -grid tolerance of the baseline search grid. This same physical tolerance, for AoA and AoD and m for distance, is retained across all five resolution settings so that the permitted parameter error does not vary with the search grid.
Sionna Ray-Tracing Communication-Link Configuration
A separate system-level evaluation is conducted using the ray-tracing capabilities of the NVIDIA Sionna RT package [22]. The simple_street_canyon_with_cars scene is used to evaluate the proposed method under geometry-dependent propagation within an end-to-end communication link. The carrier frequency, antenna configuration, and parameter-search ranges remain consistent with Table 1.
Separate LOS and NLOS test sets are generated in the same ray-tracing scene. The Tx and Rx are randomly positioned at valid roadway locations, and placements intersecting scene objects are discarded. Realizations with an unobstructed direct path are assigned to the LOS set; those without a direct path are assigned to the NLOS set. Sionna RT provides the AoA, AoD, propagation delay , and complex coefficient of each resolved MPC, with propagation distance obtained as . After removing components outside the adopted estimation ranges and redundant paths, the main LOS and NLOS evaluations retain paths per realization. A separate three-MPC LOS test set is generated for the ablation study in Section IV-D1. The trained NN-CLEAN model and GS-CLEAN search grid remain fixed across these evaluations.
Each operating point in the LOS and NLOS sweeps is evaluated over 10 independent Monte Carlo runs, with 1,000 test samples per propagation condition in each run. The methods use the same channel realizations within each run, and the reported path accuracy, channel-reconstruction MSE, and BER are averaged over the 10 runs.
The orthogonal frequency-division multiplexing (OFDM) resource grid contains OFDM symbols and subcarriers, with a cyclic-prefix length of 64 samples. OFDM symbol 7 is allocated to orthogonal MIMO pilots. The information bits are encoded using a rate-1/2 5G low-density parity-check (LDPC) code and mapped to quadrature phase-shift keying (QPSK) symbols. The communication evaluation covers dB, with the AWGN variance determined from the modulation order, coding rate, and cyclic-prefix overhead.
For each test sample, Sionna RT generates the noiseless frequency-domain channel . Let and denote the transmitted and received resource grids, respectively. Their th subcarrier slices are denoted by and . Under the quasi-static channel assumption, the received signal on each subcarrier is
where is the channel matrix at frequency , and represents AWGN.
The complete ray-tracing, channel-estimation, reconstruction, and communication-receiver chain is illustrated in Figure 4. The pilot-bearing entries of , denoted by , are processed by the Least Squares (LS) channel estimator. Nearest-neighbor interpolation produces the full-grid estimate , which serves as the common input to NN-CLEAN and GS-CLEAN. The parameter–gain pairs estimated by each method are then used to reconstruct according to the channel expansion in (3), with the additive-noise term omitted.

Figure 4. Sionna ray-tracing and communication-link evaluation framework.
The reconstructed channel and received resource grid are supplied to the downstream receiver. The linear minimum mean-square error (LMMSE) equalizer produces estimates of the transmitted QPSK symbols, which are subsequently processed by the QPSK demapper and 5G LDPC decoder to obtain the estimated information bits .
The synthetic evaluation is parameterized by , whereas the communication-link evaluation uses , with denoting the energy per information bit. For the adopted OFDM configuration, pilot normalization, and noise scaling, the measured pilot-level channel-observation SNR is approximately 2 dB below in both LOS and NLOS conditions. Thus, dB corresponds to approximately 5 dB at the pilots. This correspondence provides a reference between the noise levels; the LS channel estimation and interpolation used to form the CLEAN input produce a different error structure from the directly perturbed synthetic channel observations.
RT path accuracy is evaluated separately for the LOS and NLOS datasets using the same extraction-order association and -grid tolerance adopted for the synthetic evaluation, with (22) used to compute the reported accuracy. All estimated MPCs are retained for channel reconstruction, irrespective of whether they satisfy the path-accuracy tolerance. The reconstructed channel is compared with the noiseless RT reference using
Communication performance is characterized by the BER obtained by comparing the decoded information bits with the transmitted information bits at each operating point. LS channel estimation with nearest-neighbor interpolation and perfect channel state information (CSI) are included as the conventional and idealized receiver references, respectively.
1) Ablation Study: The full NN-CLEAN model trained according to Algorithm 1 is compared with four ablation variants. In w/o-GLS, the targets in (16) are replaced by one-hot targets at the same ground-truth class indices, while the CCE in (17) is retained. In w/o-gating, the distance head receives the shared backbone features through a branch parallel to the AoA and AoD heads, without angular conditioning. The uniform-weight variant changes the distance CCE coefficient in (18) from 3 to 1, assigning equal weight to all three parameter losses. The direct-Phase3 variant uses the Phase 3 residual updates and loss in (19) from the first training step, omitting Phases 1 and 2. Except for the stated modification, each variant retains the full model’s architecture and training settings described in Section IV-B, including the same total number of optimizer updates. All models are trained separately and evaluated on the same three-MPC LOS RT test samples using the receiver configuration described above. Path accuracy, channel-reconstruction MSE, and BER are compared at dB.
Results and Discussion
Baseline Multipath Estimation Performance
Under the off-grid synthetic evaluation, Figure 5 compares NN-CLEAN with GS-CLEAN [9], 3D-MUSIC [4], 3D-ESPRIT [6], and the one-shot NN as the path count increases. For , NN-CLEAN achieves the highest path accuracy, slightly exceeding GS-CLEAN while maintaining a clear advantage over the subspace and one-shot baselines. In these sparse cases, the learned estimator identifies the relevant region of the three-dimensional AoA–AoD–distance space without explicitly evaluating every candidate at each CLEAN iteration. For , the accuracy of both CLEAN-based methods decreases, with GS-CLEAN retaining higher accuracy. Because the reference parameters are off-grid, reconstructing a path with a discrete dictionary atom generally leaves structured residual energy after subtraction. These errors affect subsequent extraction steps and make weaker MPCs more difficult to recover. At a fixed aggregate , increasing the path count also introduces greater inter-path interference and can reduce the individual SNR of weaker components. NN-CLEAN additionally depends on learned parameter decisions, and an inaccurate early estimate can distort the residual supplied

Figure 5. Off-grid MPC estimation accuracy versus path count under the baseline synthetic configuration at dB.
to later iterations. Its larger accuracy reduction is consistent with greater sensitivity to these increasingly difficult residual observations. The one-shot NN degrades more rapidly, even with ideal cross-MPC association. It predicts every MPC from the original superposition, whereas NN-CLEAN updates the observation after each estimated path is removed. Successive cancellation reduces the contribution of previously extracted components before later predictions. The comparison therefore supports retaining the CLEAN recursion under the evaluated baseline configuration.
Parameter Precision and Reconstruction Fidelity
Path accuracy determines whether an MPC satisfies the prescribed detection tolerance but does not characterize the residual error of a successfully associated path. Figure 6 therefore presents the empirical cumulative distribution functions (CDFs) of angular MCD (Figure 6a), distance MCD (Figure 6b), and energy-captured ratio (Figure 6c) for the two-path case. The MCD distributions are conditional on successful path association and therefore measure estimation precision rather than unconditional path-recovery performance.

Figure 6. CDFs of angular MCD, distance MCD, and energy-captured ratio for over the synthetic SNR conditions.
As decreases, the angular- and distance-error distributions broaden, while the energy-captured ratio deviates further from its ideal value. GS-CLEAN retains the highest conditional precision because it evaluates the complete candidate grid explicitly. Nevertheless, the NN-CLEAN distributions remain close to the GS-CLEAN reference across the evaluated SNR range, indicating that the additional localization error is limited once NN-CLEAN identifies the correct MPC neighborhood.
Both methods remain subject to the same off-grid quantization floor because the continuously generated paths are reconstructed using discrete dictionary atoms. Angular mismatch perturbs the spatial array response, whereas distance mismatch produces a coherent phase error across the observed bandwidth. The latter persists over all subcarriers and can propagate through successive residual updates. Although NN-CLEAN introduces an additional classification error, the closely aligned MCD and energy-captured distributions show that it preserves most of the dominant geometric and energetic channel structure.
Runtime and Memory Efficiency
Figure 7 compares the runtime and peak GPU memory of the one-MPC parameter-search stage as the processing batch size, , increases from 1 to 8. At , NN-CLEAN requires 0.82 ms, compared to 1.76 ms for GS-CLEAN. GS-CLEAN runtime and memory increase approximately linearly over the batch sweep, with both quantities roughly doubling from to 8. NN-CLEAN runtime remains below 1 ms throughout the sweep, while its memory requirement grows much more slowly. At , NN-CLEAN takes 0.92 ms, compared to 25.24 ms for GS-CLEAN, corresponding to an approximately 28× speedup. Peak GPU memory is 0.04 GB for NN-CLEAN and 2.16 GB for GS-CLEAN, a reduction by a factor of approximately 53. The computational advantage therefore becomes substantially larger as more channels are processed within the same estimation call.

Figure 7. Runtime and Memory Comparison of NN-CLEAN and GS-CLEAN for one-MPC
Table III compares the one-MPC search times with the end-to-end runtime and peak memory of complete three-MPC extraction. GS-CLEAN takes approximately 3× its one-MPC search time, consistent with dictionary evaluation dominating the workload and being repeated at each extraction step. NN-CLEAN requires 1.91 ms at , approximately 2.3×
| Method | One-MPC search (ms) | Three-MPC end-to-end runtime (ms) | Three-MPC peak memory (GB) | |
| 1 | GS-CLEAN | 1.76 | 5.40 | 0.28 |
| 1 | NN-CLEAN | 0.82 | 1.91 | 0.02 |
| 4 | GS-CLEAN | 12.31 | 36.68 | 1.08 |
| 4 | NN-CLEAN | 0.73 | 2.14 | 0.03 |
| 8 | GS-CLEAN | 25.24 | 75.16 | 2.16 |
| 8 | NN-CLEAN | 0.92 | 2.80 | 0.04 |
Table III. Runtime and Memory Comparison of NN-CLEAN and GS-CLEAN with three-MPC
its one-MPC search time. This lower ratio is consistent with fixed call and synchronization overhead forming a larger fraction of the short single-path measurement and being amortized within the three-path estimation call. The ratio approaches 3× as the batch workload increases. At , complete NN-CLEAN extraction achieves an approximately 27× speedup and a 52-fold reduction in peak GPU memory compared to GS-CLEAN. For both methods, peak memory remains close to the corresponding one-MPC requirement because the processing workspace is reused between successive extraction steps.
Impact of Observation Dimensions and Search-Grid Resolution
Figure 8 compares the effects of observation dimensions and search-grid resolution on three-MPC off-grid estimation at dB. In Figure 8(a), increasing from 4 to 8 gives GS-CLEAN a modest accuracy improvement, while NN-CLEAN shows no consistent increase. The number of frequency observations has a stronger effect, as shown in Figure 8(b). Both methods improve with , and NN-CLEAN obtains most of its gain below the baseline . Additional frequency observations help constrain the distance-dependent phase progression used in parameter estimation and cancellation. Because the bandwidth is fixed, increasing provides denser sampling and supports noise averaging without changing the nominal delay resolution. The gains become smaller beyond , which captures most of the improvement observed in this sweep.

Figure 8. Accuracy under variations in observation dimension and search grid resolution for at dB.
Figure 8(c) compares search grids using the same physical accuracy tolerances. NN-CLEAN slightly exceeds GS-CLEAN on the coarse grid, while both methods achieve their highest measured accuracy at the baseline resolution. Finer grids provide no consistent improvement for GS-CLEAN and reduce NN-CLEAN accuracy. For NN-CLEAN, the classification outputs and their directly connected layers expand with grid refinement, but the shared backbone, hidden feature dimensions, and optimization-step count remain fixed. The model must therefore distinguish more closely spaced classes using the same feature representation. In addition, the fixed GLS width in bins corresponds to a narrower physical smoothing interval on finer grids. These training and architecture choices may limit the benefit of reducing the grid spacing.
Ray-Tracing Generalization and Communication Performance
The Sionna RT experiments provide a geometry-based validation of NN-CLEAN and evaluate how parameter-estimation errors affect channel reconstruction and downstream communication performance. Unlike the independently sampled synthetic parameters, the RT paths are determined jointly by the scene geometry and propagation interactions. The following results therefore evaluate both the transfer of the learned search to a different channel distribution and the effect of the resulting estimation errors on the decoded communication link.
Ray-Tracing MPC Estimation Accuracy
Figure 9 compares path accuracy for the two-MPC LOS and NLOS test sets. Accuracy generally increases with and levels off toward the upper end of the sweep. GS-CLEAN maintains higher accuracy in both propagation conditions, while NN-CLEAN saturates earlier, particularly for NLOS channels. Both estimators achieve higher accuracy in LOS compared to NLOS under the same retained path count, search grid, and evaluation tolerance.

Figure 9. Accuracy versus for two-MPC LOS and NLOS RT channels.
The RT result differs from the synthetic two-MPC baseline, where NN-CLEAN slightly exceeds GS-CLEAN. RT paths have geometry-dependent parameter relationships, and their channel observations are obtained through pilot-based LS channel estimation and interpolation. Applying the trained NN-CLEAN model without retraining therefore exposes it to changes in both path statistics and input-error structure. GS-CLEAN continues to evaluate the physical dictionary directly. The observed accuracy gap captures the combined change in channel distribution and observation pipeline; the pilot-level
SNR correspondence does not establish identical CLEAN input conditions.
Ray-Tracing Channel Reconstruction
Figure 10 compares channel-reconstruction MSE against the noiseless two-MPC RT reference. Both CLEAN-based methods achieve substantially lower MSE compared to LS channel estimation with nearest-neighbor interpolation. The parametric fit constrains the reconstructed channel to the spatial–frequency responses of the estimated paths, reducing the error of the interpolated observation. GS-CLEAN achieves the lowest MSE in both propagation conditions. Over the tested range, the maximum absolute difference between NN-CLEAN and GS-CLEAN remains below for both LOS and NLOS channels.

Figure 10. Channel-reconstruction MSE versus for two-MPC LOS and NLOS RT channels.
All estimated MPCs are included in reconstruction, weighted by their complex gains. An estimate outside the path-accuracy tolerance can still approximate part of the reference response, and its effect on MSE depends on its power and the resulting spatial–frequency mismatch. At higher , both reconstruction curves vary only weakly as pilot noise decreases. This behavior is consistent with persistent errors from grid discretization and the interpolated LS input. These errors affect the fitted parameters and gains and can propagate through successive cancellation, limiting further improvement under the fixed estimation pipeline.
End-to-End Communication Performance: Figure 11 compares BER for the four receiver cases in two-MPC LOS and NLOS channels. Perfect CSI provides the lowest BER, followed by GS-CLEAN, NN-CLEAN, and LS channel estimation with nearest-neighbor interpolation. Both CLEAN-based reconstructions substantially improve BER compared to LS in both propagation conditions. LOS generally provides lower BER compared to NLOS across the four receivers, apart from a small GS-CLEAN reversal at the upper end of the sweep. The LOS advantage also appears with perfect CSI, showing that the LOS–NLOS difference reflects the communication characteristics of the channels as well as channel-estimation error.

Figure 11. BER versus for two-MPC LOS and NLOS RT channels.
Over the entire tested range, the maximum absolute BER difference between NN-CLEAN and GS-CLEAN is approximately 1.05% for LOS and 0.69% for NLOS. At
dB, the absolute NN-CLEAN–perfect-CSI gap is approximately 0.4% in both conditions. NN-CLEAN BER levels off toward the upper end of the sweep, consistent with its residual channel-reconstruction error. Its reconstruction nevertheless retains a substantial BER advantage over interpolated LS, while the gap to GS-CLEAN quantifies the link-performance penalty associated with the computational savings.
Ablation Study: Table IV compares the full model and four variants on three-MPC LOS RT channels at dB. The w/o-GLS variant shows the largest deterioration across all three metrics. This result supports distributing supervision across neighboring parameter bins, whose corresponding channel responses are correlated. Removing angular conditioning in w/o-gating produces a smaller loss in path accuracy, together with higher reconstruction MSE and BER.
| Variant | Accuracy (%) | MSE | BER (%) |
| Full NN-CLEAN | 75.13 | 0.79 | |
| w/o-GLS | 71.24 | 1.40 | |
| w/o-gating | 74.53 | 0.89 | |
| uniform-weight | 75.12 | 1.11 | |
| direct-Phase3 | 74.17 | 1.01 |
Table IV. NN-CLEAN Ablation Results for Three-MPC LOS RT Channels at dB
The uniform-weight variant retains nearly the same path accuracy as the full model but has higher MSE and BER. The larger distance-loss weight therefore benefits reconstruction and link performance even when aggregate path accuracy changes little. This outcome is consistent with distance errors introducing frequency-dependent phase mismatch in the reconstructed paths and subsequent residuals. The direct-Phase3 variant also degrades all three metrics, supporting the gradual introduction of sequential extraction before training with prediction-dependent residuals. This comparison assesses the complete curriculum and does not isolate the individual contributions of Phases 1 and 2.
Conclusion
This paper presented NN-CLEAN, which replaces the exhaustive parameter search within CLEAN with a multi-head neural classifier while retaining LS gain estimation, parametric reconstruction, and residual subtraction. Reusing the same single-MPC network permits different extraction depths without changing its output structure. Off-grid synthetic experiments show that NN-CLEAN is most competitive with GS-CLEAN in sparse channels, while GS-CLEAN retains higher accuracy as the path count increases. The observation-dimension and grid-resolution sweeps also show that larger antenna configurations and finer output grids do not uniformly improve the learned estimator.
For complete three-MPC extraction at a batch size of eight, NN-CLEAN achieves an approximately 27× speedup and a 52-fold reduction in peak GPU memory compared to GS-CLEAN. The two-MPC RT evaluation shows a reconstruction and BER penalty relative to GS-CLEAN in both LOS and NLOS, while both CLEAN-based methods substantially improve performance over interpolated LS. At dB, the absolute BER gap between NN-CLEAN and perfect CSI is approximately 0.4% in both conditions. Three-MPC LOS ablations support label smoothing, angular conditioning, distance-loss weighting, and staged training. The results quantify the tradeoff between batched computational efficiency and estimation performance. Adaptive stopping for unknown path counts remains a direction for future work.
Acknowledgements
This work was supported in part by CoCoSys, one of seven centers in JUMP 2.0, a Semiconductor Research Corporation (SRC) program sponsored by DARPA.
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 journal on selected areas in communications, vol. 40, no. 6, pp. 1728–1767, 2022.DOI
- [2]A. Sams and B. Smida, “On the reliability of estimation bounds in low-snr bistatic isac,” in GLOBECOM 2025-2025 IEEE Global Communications Conference. IEEE, 2025, pp. 1440–1445.
- [3]R. Schmidt, “Multiple emitter location and signal parameter estimation,” IEEE transactions on antennas and propagation, vol. 34, no. 3, pp. 276–280, 1986.DOI
- [4]M. C. Vanderveen, C. B. Papadias, and A. Paulraj, “Joint angle and delay estimation via der for multipath signals arriving at an antenna array,” IEEE Communications letters, vol. 1, no. 1, pp. 12–14, 2002.
- [5]R. Roy and T. Kailath, “Esprit-estimation of signal parameters via rotational invariance techniques,” IEEE Transactions on acoustics, speech, and signal processing, vol. 37, no. 7, pp. 984–995, 2002.
- [6]M. C. Vanderveen, A.-J. Van der Veen, and A. Paulraj, “Estimation of multipath parameters in wireless communications,” IEEE transactions on signal processing, vol. 46, no. 3, pp. 682–690, 2002.
- [7]A. A. Hussain, N. Tayem, and A.-H. Soliman, “Low complexity doa estimation of multiple coherent sources using a single direct data snapshot,” IEEE Access, vol. 12, pp. 2371–2388, 2023.DOI
- [8]K. Xu, M. Xing, R. Zhang, M. Sha, W. Nie, Y. Quan et al., “High-accuracy doa estimation algorithm at low snr through exploiting a supervised index,” IEEE Transactions on Aerospace and Electronic Systems, vol. 58, no. 4, pp. 3658–3665, 2022.DOI
- [9]V. Kristem, S. Sangodoyin, C. U. Bas, M. Käske, J. Lee, C. Schneider, G. Sommerkorn, C. J. Zhang, R. S. Thomä, and A. F. Molisch, “3d mimo outdoor-to-indoor propagation channel measurement,” IEEE Transactions on Wireless Communications, vol. 16, no. 7, pp. 4600–4613, 2017.arxiv.org/abs/1703.10607
- [10]L. Sun, C. Deng, J.-S. Hsu, and S. Sangodoyin, “Guided random search algorithm for offline high resolution terahertz radar sensing,” IEEE Transactions on Terahertz Science and Technology, 2026.DOI
- [11]F. Athley, “Threshold region performance of maximum likelihood direction of arrival estimators,” IEEE Transactions on Signal Processing, vol. 53, no. 4, pp. 1359–1373, 2005.DOI
- [12]T. Luo, P. Chen, Z. Yang, L. Zheng, Y. Zhang, and J. Liu, “A robust deep learning-based doa super-resolution estimation method with single snapshot,” IEEE Transactions on Vehicular Technology, 2025.DOI
- [13]H. Huang, J. Yang, H. Huang, Y. Song, and G. Gui, “Deep learning for super-resolution channel estimation and doa estimation based massive mimo system,” IEEE Transactions on Vehicular Technology, vol. 67, no. 9, pp. 8549–8560, 2018.DOI
- [14]S. Zheng, Z. Yang, W. Shen, L. Zhang, J. Zhu, Z. Zhao, and X. Yang, “Deep learning-based doa estimation,” IEEE Transactions on Cognitive Communications and Networking, vol. 10, no. 3, pp. 819–835, 2024.DOI
- [15]Q. Xu, Y. Gao, R. Zhang, J. Kong, and C. Yuen, “Deep learning-based doa estimation via grid search approach for unknown and variable number of signals,” IEEE Transactions on Cognitive Communications and Networking, 2025.
- [16]A. M. Elbir, “Deepmusic: Multiple signal classification via deep learning,” IEEE Sensors Letters, vol. 4, no. 4, pp. 1–4, 2020.DOI
- [17]S. Zheng, Z. Yang, W. Shen, L. Zhang, J. Zhu, Z. Zhao, and X. Yang, “Deep learning-based doa estimation,” IEEE Transactions on Cognitive Communications and Networking, vol. 10, no. 3, pp. 819–835, 2024.DOI
- [18]S. Xu, Z. Wang, W. Zhang, and Z. He, “End-to-end regression neural network for coherent doa estimation with dual-branch outputs,” IEEE Sensors Journal, vol. 24, no. 3, pp. 4047–4056, 2023.
- [19]C. Deng, L. Sun, J. Lee, and A. Raychowdhury, “Neural network-assisted clean for channel modeling in low-snr regimes,” arXiv preprint arXiv:2607.27450, 2026.arxiv.org/abs/2607.27450
- [20]J. A. Högbom, “Aperture synthesis with a non-regular distribution of interferometer baselines,” Astronomy and Astrophysics Supplement, Vol. 15, p. 417, vol. 15, p. 417, 1974.
- [21]B. Shaybet, V. Tourbabin, and B. Rafaely, “Srp-phat-net: A reliability-driven dnn for reverberant speaker localization,” arXiv preprint arXiv:2510.22682, 2025.arxiv.org/abs/2510.22682
- [22]J. Hoydis, S. Cammerer, F. Ait Aoudia, M. Nimier-David, L. Maggi, G. Marcus, A. Vem, and A. Keller, “Sionna,” 2022, https://nvlabs.github.io/sionna/.arxiv.org/abs/2203.11854