Attention-Based SINR Estimation in User-Centric Non-Terrestrial Networks

This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. Attention-Based SINR Estimation in User -Centric Non-T errestrial Networks Bruno De Filippo ∗ , Alessandro Guidotti ∗ † , Alessandro V anelli-Coralli ∗ ∗ Department of Electrical, Electronic, and Information Engineering (DEI), Univ . of Bologna, Bologna, Italy † National Inter-Uni versity Consortium for T elecommunications (CNIT), Bologna, Italy { bruno.defilippo, a.guidotti, alessandro.vanelli } @unibo.it Abstract —The signal-to-interference-plus-noise ratio (SINR) is central to performance optimization in user -centric beamforming for satellite-based non-terrestrial networks (NTNs). Its assess- ment either requir es the transmission of dedicated pilots or relies on computing the beamf orming matrix through minimum mean squared err or (MMSE)-based formulations bef orehand, a process that introduces significant computational overhead. In this paper , we propose a lo w-complexity SINR estimation framework that leverages multi-head self-attention (MHSA) to extract inter- user interference featur es directly fr om either channel state information or user location reports. The proposed dual MHSA (DMHSA) models evaluate the SINR of a scheduled user group without requiring explicit MMSE calculations. The architectur e achieves a computational complexity reduction by a factor of three in the CSI-based setting and by two orders of magnitude in the location-based configuration, the latter benefiting from the lower dimensionality of user reports. W e show that both DMHSA models maintain high estimation accuracy , with the root mean squared error typically below 1 dB with priority-queuing- based scheduled users. These results enable the integration of DMHSA-based estimators into scheduling procedures, allowing the evaluation of multiple candidate user groups and the selection of those offering the highest av erage SINR and capacity . Index T erms —Non-T errestrial Networks, User -Centric Beam- forming, Self-Attention, Deep Learning I . I N T RO D U C T I O N The signal-to-interference-plus-noise ratio (SINR) of a com- munication link plays a major role in the ef fectiveness of the data exchange. On the one hand, the achie vable rate in communication systems is limited by the SINR through the Shannon formula, effecti vely limiting the amount of bits that can be sent per time and bandwidth unit [1]. On the other hand, procedures in communication systems rely on SINR- related metrics to take decisions, e.g. , handovers from one cell to another can be carried out once the measured reference signal recei ved quality overcome a certain threshold [2]. Such proxy for link quality can typically be measured at the terminal using reference signals; howe ver , especially in the context of satellite-based non-terrestrial networks (NTNs), complications can be introduced during the reporting phase: geostationary- class satellites are characterized by a large propagation delay which can result in outdated reports, e.g. , in the case of user mobility , while non-geostationary-class satellites at lower orbit claim lo wer latency but introduce aging ef fects due to their inherent orbital movement. Focusing on user -centric beam- forming NTN systems, the assessment of the SINR strongly relies on the set of users scheduled during the same time slot. Indeed, once a group of users is scheduled and the beamform- ing matrix is computed, either by means of minimum mean squared error (MMSE) or location-based MMSE (LB-MMSE), the SINR achiev ed by each user can be e valuated (considering the theoretical clear sky channel formula based on the user’ s location in case of LB-MMSE). Howe ver , such ev aluation is based on the MMSE equation, which introduces a great computational o verhead [3]. Thanks to the function approx- imation capabilities of neural networks, the SINR achieved by a group of scheduled users in a user-centric beamforming NTN system can be assessed with reduced computational complexity by 1) extracting complex features from the users’ channel vectors, and 2) ev aluating the interactions between different users’ features. In this conte xt, we identified the popular attention mechanism as a potential candidate for extracting inter-user interference-related information. T o the best of our knowledge, previous works ha ve either obtained formulations of the SINR in cell-free MIMO systems based on the MMSE equation [3] [4] or focused on the distribution of the SINR [5]; on the opposite, approaches to SINR estimation based on deep learning (DL) are underexplored, especially in the NTN frame work. W e believe that the availability of low-comple xity SINR estimates can improve user scheduling algorithms by pro viding an effecti ve e valuation metric to select the best user group to be scheduled among a set of candidate groups. Based on these considerations, in this paper we: • Propose a low-comple xity self-attention-based SINR es- timator for user-centric beamforming; • Assess the performance of the model in a NTN consid- ering random and priority-based scheduling. I I . S Y S TE M M O D E L W e here consider a lo w Earth orbit (LEO) satellite providing cov erage to a set of N U E users in the service area in a user- centric fashion (Figure 1). The spaceborne node is equipped with a regenerativ e payload, with the on board gNodeB (gNB) distributed unit (DU) implementing up to the upper medium access control functions on board ( i.e. , functional split option 4), including user scheduling; furthermore, the satellite is equipped with a direct radiating array composed of N R radi- ating elements capable of generating up to N B simultaneously activ e beams [6]. At least one gatew ay , implementing the gNB centralized unit (CU), is always reachable by the satellite. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. Fig. 1: Considered system model. When in visibility , each user , equipped with a very-small- aperture terminal, periodically reports either its estimated CSI or location (latitude and longitude) to the gNB. The channel coefficient between the k -th user equipment (UE) and the n -th radiating element at time t can be computed as [3]: H ( t ) k,n = g ( tx,t ) k,n g ( rx,t ) k,n 4 π d ( t ) k λ q L ( t ) k κB k T k e − j 2 π λ d ( t ) k , (1) where g ( tx,t ) k,n and g ( rx,t ) k,n represent the transmitting and re- ceiving gain, respectiv ely , between the k -th UE and the n - th radiating element; d ( t ) k is the k -th user’ s slant range; L ( t ) k represents all stochastic losses ( e.g . , due to shadowing and the atmosphere); κ is Boltzmann’ s constant; B k and T k represent the k -th user’ s allocated bandwidth and equiv alent noise tem- perature, respectiv ely; and λ is the wav elength corresponding to the carrier frequency f c . When new reports are av ailable at the gNB, the user’ s estimated channel vectors (either reported in case of CSI reports or computed by means of (1) with L ( t ) k = 0 dB in case of location reports) are grouped together in the channel matrix ˆ H ( t ) U E ∈ C N U E × N R ; at ev ery scheduling instant, occurring with periodicity T sched , such channel matrix is then used to select one group of users for each time slot of duration T ( slot ) sched in the scheduling windo w . Thus, in the generic time slot t , a subset of N ( t ) U E ,sched ≤ N B users are scheduled for transmission, resulting in the estimated scheduled channel matrix ˆ H ( t ) ∈ C N ( t ) U E ,sched × N R . The corresponding beamform- ing matrix can be computed through the MMSE equation [3]: B ( t ) = ( ˆ H ( t ) ) H  ˆ H ( t ) ( ˆ H ( t ) ) H + α I N B  − 1 , (2) where α = N R /P av is the beamforming regularization factor , with P av representing the total transmission po wer at the spaceborne node, and I a representing the a × a identity matrix. The per-antenna constraint normalization ensures that all antenna elements transmit the same amount of power: ˜ B ( t ) = r P av N R diag 1 || B ( t ) 1 , : || , . . . , 1 || B ( t ) N R , : || ! B ( t ) , (3) with diag ( x 1 , . . . , x a ) being the a × a diagonal matrix having values x 1 , . . . , x a on the main diagonal. The SINR of user k at time slot t can then be ev aluated as: S I N R ( t ) k = S N R ( t ) k 1 + I N R ( t ) k = || H ( t ) k, : ˜ B ( t ) : ,k || 2 1 + P N ( t ) U E ,sched l =1 ,l  = k || H ( t ) k, : ˜ B ( t ) : ,l || 2 . (4) I I I . D UA L M U LT I - H E A D E D S E L F - A T T E N T I O N F O R S I N R E S T I M AT I O N A. DMHSA Model The proposed DL model for SINR estimation is a dual multi-headed self-attention (DMHSA) architecture (Figure 2). For the sake of conciseness, we recall only the core concepts relev ant to the model. Scaled dot-product attention (SDP A) maps queries, k eys, and v alues to output representations by assigning weights proportional to the dot-product between queries and keys. In self-attention, these vectors are linearly projected from the same input, and the multi-headed formu- lation (MHSA) applies several parallel projections to capture different relationships within the data. In the proposed model, the attention mask allows selecti ve control over which user interactions contribute to each attention head. Input r epr esentation: Dropping the temporal index for clarity , the model’ s input is a matrix F ∈ R N U E ,sched × δ containing δ features per scheduled user . Since N U E ,sched varies across samples, a binary mask m ( DM H S A ) = [ 1 1 × N U E ,sched , 0 1 × ( N B − N U E ,sched ) ] ∈ { 0 , 1 } 1 × N B is provided to ensure efficient parallelization. The network then computes ˆ Y ( DM H S A ) = f ( DM H S A ) ( F , m ( DM H S A ) ) , where each row F k, : ∈ R δ corresponds to the features of user k . For CSI-based beamforming, such features include: 1) The normalized phases of the k -th estimated CSI vector , Φ k,n = ∠ ˆ H k,n σ φ ∀ n = 1 , . . . , N R , with σ φ = π √ 3 , under the assumption ∠ ˆ H k,n ∼ U ( − π , π ) ; 2) The normalized mean squared magnitude of the CSI, ψ k =  1 N R P N R n =1    ˆ H k,n    2 − µ H  /σ H , with µ H and σ H being standardization parameters empirically esti- mated from a data sample; and 3) The ratio ρ = N U E ,sched / N B . The resulting feature vector has size δ ( C S I ) = N R + 2 : F (CSI) k, : = [ Φ k, 1 . . . , Φ k,N R , ψ k , ρ ] , (5) For location-based beamforming, the input features are the user coordinates in the ( u, v ) reference system, u k and v k , and the same ratio ρ , yielding δ ( GE O ) = 3 : F (GEO) k, : = [ u k , v k , ρ ] . (6) It is worth mentioning that the inclusion of the ratio ρ = N U E ,sched / N B provides the network with coarse information about the amount of interference to be expected. Embedding extr action: Each user’ s feature vector is pro- cessed independently by a stack of fully connected (FC) layers with N C neurons, layer normalization (LN), and leaky rectified linear unit acti vation. Position embedding (PE) is then added to add further information on the expected level of interference. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. Fig. 2: DMHSA SINR estimation model. Dual attention mechanism: The resulting embeddings are fed to two parallel MHSA+FC modules. Both operate on the set of scheduled users, but they dif fer in how user interactions are masked: • The SNR module uses a full attention mask M S = 1 N U E ,sched × N U E ,sched to enable all interactions. • The INR module uses M I = 1 N U E ,sched × N U E ,sched − I N U E ,sched , excluding self-interference to focus solely on inter-user contributions. Each module uses h S = h I = h = 4 heads, with projection dimensions chosen such that d k = d v = N C /h = 2 . The standardized SINR in dB is obtained by subtracting the output of the second module’ s FC layer with single output neuron from the first’ s: intuitiv ely , they correspond to an estimation of the terms (1 + I N R k ) and S N R k in dB in (4). B. Computational complexity Regardless of the choice between CSI-based beamforming and location-based beamforming, the SINR can be ev aluated through (4) (with the theoretical CSI being computed based on the users’ geographical location in case of location-based beamforming). When assessing the MMSE beamforming algo- rithm’ s complexity from (2), which is at the base of the SINR ev aluation, two main sources of complexity can be identified: • The matrix multiplication between the CSI matrix ˆ H ( t ) ∈ C N U E ,sched × N R and its conjugate transpose with com- plexity O  N 2 U E ,sched N R  ; and • The in verse operation on the N U E ,sched × N U E ,sched complex matrix ˆ H ( t )  ˆ H ( t )  H + α I N U E ,sched , with com- plexity O  N 3 U E ,sched  . Howe ver , as the number of scheduled users N U E ,sched is typically limited with respect to the number of radiating elements N R , i.e. , N U E ,sched ≪ N R , the MMSE complexity reduces to: O M M S E = O  N 2 U E ,sched N R  . (7) The computational complexity of the DMHSA model can be mainly attributed to FC and MHSA layers: • The first FC layer embeds δ features into a vector of size N C for each user , resulting in a computational complexity O ( δ N U E ,sched N C ) . • The second FC layers further refines the N C features for each user, resulting in a computational complexity O  N U E ,sched N 2 C  . • The MHSA layers apply linear projections to compute the queries, ke ys, and values matrices; the subsequent computations in the SDP A in volv e a matrix multiplication between queries and ke ys; finally , the heads outputs are concatenated to form a vector of N C = h · d k features and passed through one final FC layer . Thus, the overall computational complexity of the MHSA layers is [7]: O M H S A = O  N 2 U E ,sched N C + N U E ,sched N 2 C  . (8) • The last pair of FC layers generates one feature from a vector of length N C for each user , resulting in a computational complexity O ( N U E ,sched N C ) . Thus, the computational complexity of the DMHSA model is: O DM H S A = O ( δ N U E ,sched N C ) + O  N 2 U E ,sched N C + N U E ,sched N 2 C  . (9) In case of CSI-based beamforming, δ ( C S I ) = N R + 2 ; thus, the computational complexity becomes O ( N U E ,sched N C N R ) + O  N 2 U E ,sched N C + N U E ,sched N 2 C  . Considering that N U E ,sched ≪ N R , the complexity reduces to O  N U E ,sched N C N R + N U E ,sched N 2 C  . Furthermore, we choose N C = 8 so that N C ≪ N R ; thus, the o verall computational complexity for SINR estimation with the CSI-DMHSA model is: O ( C S I ) DM H S A = O ( N U E ,sched N C N R ) , (10) which, giv en N C < N U E ,sched (typically N U E ,sched = N B = 24 ), implies that the complexity is effecti vely reduced This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. T ABLE I: Computational complexity for SINR estimation. SINR estimation method Computational complexity MMSE-based O M M S E = O  N 2 U E ,sched N R  CSI-DMHSA O ( C S I ) DM H S A = O  N U E ,sched N C N R  GEO-DMHSA O ( GE O ) DM H S A = O  N 2 U E ,sched N C  Fig. 3: Computational comple xity as a function of N C with N U E ,sched = 24 (black dashed line corresponds to N C = 8 ). with respect to the ev aluation through the computation of the MMSE beamforming equation. In case of location-based beamforming, δ ( GE O ) = 3 ; thus, the computational complex- ity becomes O  N 2 U E ,sched N C + N U E ,sched N 2 C  . With the same consideration on N C as abov e, the complexity of the GEO-DMHSA model reduces to: O ( GE O ) DM H S A = O  N 2 U E ,sched N C  , (11) providing a great computational complexity reduction for SINR estimation. The results of the computational complexity assessment are summarized in T able I. Figure 3 reports the Big-O computational complexity achiev ed by the two DMHSA models (CSI-based and location- based in orange and yello w , respectively) and the MMSE baseline (blue line) as a function of the number of channels N C , i.e. , the size of the embeddings of each user’ s features vector; the black dashed line represents the value set in the trained models, i.e. , N C = 8 . The plot shows that the reduced features vector in the GEO-DMHSA model ensures that the complexity remains orders of magnitude lower than that of MMSE-based SINR estimation: at N C = 8 , the location-based algorithm achiev es a complexity of ≈ 4 . 6 · 10 3 against the ≈ 2 . 9 · 10 5 required by the benchmark. The CSI-DMHSA model also ensures a reduction in complexity to ≈ 9 . 8 · 10 4 ; howe ver , the large features vector makes so the complexity in the CSI-based case scales much faster than in the location- based case with respect to N C : while longer embeddings could improve the SINR estimation accuracy , the increased complexity may disfa vour the application of the CSI-DMHSA method. This is not the case with GEO-DMHSA, where even with embeddings of 24 elements the complexity is at one order of magnitude under the MMSE-based SINR estimation. Parameter V alue Subcarrier spacing ∆ f = 120 kHz Carrier frequency f c = 20 GHz User bandwidth B = 190 . 08 MHz Scenario Rural [8] Channel model 3GPP NTN system-level [9] [10] Number of radiating elements N R = 512 Power per radiating element P av,el = 65 mW Minimum user elevation angle 30° NTN node altitude 1000 km T raining batch size N batch = 8192 Maximum training epochs 15000 epochs L2 regularization factor ϵ R = 10 − 6 LR warm-up period T w = 40 epochs LR cosine annealing period T c = 100 epochs Minimum LR λ min = 10 − 4 Maximum LR λ min = 5 · 10 − 3 Early stopping patience T s = 4 cosine annealing cycles T ABLE II: Simulation parameters for DMHSA training. It should also be mentioned that the CSI-DMHSA and GEO- DMHSA models require 4.8k and 762 learnable parameters, respectiv ely , making them extremely compact. C. Model training T wo separate models with the reported DMHSA architecture hav e been trained using CSI reports as input data for one of them and location reports for the other . The models have been trained by simulating scheduling instances at random samples of a satellite pass, using the parameters reported in T able II. At each instance i , N U E ,sched, i ∼ U (8 , N B ) users are randomly selected for scheduling, their CSI vectors (or location v ectors) are pre-processed and collected as training input data sample F (Section III-A), and their SINR is assessed by means of the MMSE equation and stored as label ( i.e. , true SINR v alue) for the corresponding input data sample. Once the batch is filled with N batch samples, the input data X ( DM H S A ) ∈ R N B × δ × N batch and the corresponding padding masks M ( DM H S A ) ∈ R N B × N batch are fed to the model, and the corresponding SINR estimates ˆ Y ( DM H S A ) ∈ R N B × N batch are obtained; then, the MSE loss function L M S E is ev aluated between the SINR estimates at the output of the model and the corresponding ground truth labels SINR ∈ R N B × N batch : MSE u,i =  ˆ Y ( DM H S A ) u,i − SINR u,i  2 , (12) L M S E = P N batch i =1 P N B u =1 M ( DM H S A ) u,i ⊙ MSE u,i P N batch i =1 P N B u =1 M ( DM H S A ) u,i , (13) where ⊙ is the element-wise product. Through the backpropagation algorithm and the Adam opti- mizer [11], the trainable weights of the model are updated, and a ne w batch of data starts being generated. The training process employs L2 regularization, early stopping, learning rate (LR) warmup, and LR cosine annealing with warm restarts, with T able II reporting the parameters emplo yed to train both of the DMHSA models. A new batch of data is generated at This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. (a) CSI-DMHSA (b) GEO-DMHSA Fig. 4: SINR estimation error distribution for dif ferent N U E ,sched values (random scheduler). each epoch by means of simulation, and samples contained in pre vious batches are not reused in the following epochs: for this reason, a v alidation set is not generated, with early stopping being set based on the training set loss. I V . R E S U LT S A. Random scheduler The objecti ve of the first e valuation is to determine the SINR estimation error under general conditions: e.g. , training on ran- domly scheduled users ensures that the models observe both cases of optimal scheduling and ill-conditioned beamforming matrices. For this reason, the models were first ev aluated on a test set under the same conditions of the training set, i.e. , each sample contains a random number of randomly scheduled users’ reports obtained in random instants of the satellite pass, for a total of ≈ 4 · 10 7 SINR estimates. As for the training data, the simulation parameters used to generate test data are the same reported in T able II. Notably , user locations are sampled from a non-uniform distribution based on the Joint Research Center Global Human Settlement Layer data [12] according to the preprocessing procedure detailed in [8]. Figure 4 reports the histogram of the error distribution for both CSI-DMHSA and GEO-DMHSA, computed as E = ˆ Y ( DM H S A ) − SINR , with probability density function (PDF) normalization. The assessment is carried out by separating the test datasets based on the number of scheduled users. The distribution is ske wed in both cases, suggesting that the models cannot pro vide accurate estimates in certain conditions: this is the case of ill-conditioned beamforming matrices resulting from poor user scheduling ( i.e . , the random scheduler often selects users that are geographically close together , resulting in highly correlated CSI vectors). As expected, for both models, the root MSE (RMSE) increases with the number of users scheduled, and so does the frequenc y of high-error estimations. It should be noted that the high SINR error tails are caused by poor scheduling choices made by the random scheduler . Reli- able SINR estimations should thus be achiev able if candidate user groups are provided by a proper scheduling technique. Parameter V alue Scheduling slot duration T ( sched ) slot = 10 ms Scheduling periodicity T sched = 2 s Maximum residual visibility σ vis = 50 slots Unmet capacity factor σ cap = 2 Number of priority classes N P C = 2 Minimum capacity request C min = { 5 , 20 } Mbps Maximum capacity request C min = { 100 , 500 } Mbps T ABLE III: Configuration parameters for PQS. B. PQS scheduler In the second ev aluation step, the priority queue scheduler (PQS) presented in [8] is implemented. In this context, we assign the k -th user a traffic request C ( RE Q ) k ∈ [ C min , C max ] , k = 1 , . . . , N U E , based on the population density in the user’ s area. The PQS algorithm selects groups of N U E ,sched ≤ N B users to be served by 1) classifying the packets to be transmit- ted into N P C ≥ 2 priority classes based on the unmet capacity request and the remaining visibility of the corresponding user; 2) filling N P C priority-based queues with the corresponding packets; and 3) for each time slot in a scheduling period, randomly selecting N B packets from the non-empty queue with highest priority , ensuring that the selected users’ CSI vectors’ correlation (CSI-based scheduling) or the inter-user distance (geographic scheduling) are under a fixed threshold. By implementing the PQS algorithm, we consider the case in which a scheduler initially selects tentati ve groups of users to be serv ed during the same time slot for which the SINR should be assessed ( e.g . , to determine the expected rate for each user , or to select the group that achiev es the highest the sum-rate). The PQS scheduler parameters and the additional simulation parameters with respect to T able II are reported in T able III; the CSI and GEO scheduling constraint threshold was selected from the values reported in [8] based on the minimum and maximum requested capacity C min and C max . It should be noted that the models are assumed to be deployed as is, i.e. , without retraining on user reports with PQS scheduling. F or this reason, due to the different SINR statistics with respect to the training dataset, the obtained estimator is biased; such bias is estimated once and subtracted from the models’ output. Figure 5 reports the CDFs of the absolute error E abs =    ˆ Y ( DM H S A ) − SINR    achiev ed by the proposed models under the minimum and maximum capacity requests C min and C max , respectiv ely . The plot shows that both models are able to achieve a median absolute error under 0.7 dB for all cases. Overall, the performance of the DMHSA models is similar , with the location-based model providing slightly improved high percentiles and the CSI-based model excelling in the lo wer ones (except in the 5/100 case, where the opposite is true). The plot also shows that the cases with the lar gest C min are the ones pro viding the best absolute error: as reported in [8], a low C min value leads the PQS scheduler tends to schedule less users than the number of beams, thus resulting in consistently high SINR values, a case This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. Fig. 5: CDF of the SINR estimation absolute error with PQS scheduling for different capacity request ranges ( C min / C max ). CSI-DMHSA GEO-DMHSA C max [Mbps] 100 500 100 500 C min [Mbps] 5 0.96 dB 0.84 dB 1.18 dB 0.90 dB 20 0.71 dB 0.70 dB 0.68 dB 0.70 dB T ABLE IV: SINR estimation RMSE with PQS. that was not well represented in the training data considering the implementation of a random scheduler; on the opposite, C min = 20 Mbps results in a better aligned SINR distribution. Similarly , an increase of C max from 100 Mbps to 500 Mbps with C min = 5 Mbps leads to a significant improvement of the absolute error due to the more stringent constraint on the CSI vectors correlation and the inter-user distance. T able IV reports the SINR estimation RMSE achie ved by the DMHSA models under the considered capacity requests. Notably , the RMSE remains under the 1dB threshold (except for the 5/100 case with GEO-DMHSA). The two models perform similarly at high C min ; on the opposite, the CSI- DMHSA model has an advantage o ver its location-based counterpart at lo w C min . Clearly , the observations reported for Figure 5 are also reflected in the RMSE. V . C O N C L U S I O N S In this paper , we proposed a DL model to estimate the SINR of a group of users scheduled for transmission in a user-centric beamforming NTN system, reducing the compu- tational complexity with respect to the assessment of the SINR through the beamforming matrix while maintaining satisfying estimation accuracy . The considered architecture, based on the MHSA mechanism, achie ves Big-O complexity reduction of a factor 3 in the CSI-based case and of two orders of magnitude in the location-based case. This last result is a consequence of the lo wer dimensionality of the user reports for location- based user-centric beamforming systems, which a void scal- ing the computational comple xity by the number of antenna elements N R . Furthermore, both models achie ve satisfying SINR estimation performance considering PQS scheduling, with the RMSE not passing the 1dB threshold in most of the cases. With the estimation performance being ev aluated, the DMHSA models can be deployed for inference. Scheduling algorithms typically provide a single group of users to be served during the same time slot. While mechanisms, e.g. , the priority queue in the PQS algorithm, are in place to improv e the capacity offered in a time slot, the decision is often sub-optimal in this regard. Thus, the capacity to a group of potentially scheduled users may be ev aluated through the SINR by means of DMHSA-based estimation. Future w orks will in vestigate SINR-estimation-aided schedulers, including the impact of the estimation accuracy on the top-K selection mechanism, as a potential solution to impro ve the achiev ed capacity in user -centric NTNs. Furthermore, online learning framew orks will be in vestigated to ensure the adaptability of the model to dynamic channel conditions. V I . A C K N O W L E D G M E N T S This work has been funded by the 5G-ST ARDUST project, which receiv ed funding from the Smart Networks and Services Joint Undertaking (SNS JU) under the European Union’ s Horizon Europe research and innov ation programme under Grant Agreement No 101096573. The views expressed are those of the authors and do not necessarily represent the project. The Commission is not liable for any use that may be made of any of the information contained therein. R E F E R E N C E S [1] B. Ahmad, D. G. Riviello, B. De Filippo, A. Guidotti, and A. V anelli- Coralli, “Location-assisted Graph-based User Scheduling in Multi User MIMO LEO NTN Systems, ” W iley International Journal of Satellite Communications and Networking , 2025. [2] Y . I. Demir , M. S. J. Solaija, and H. Arslan, “On the Performance of Handover Mechanisms for Non-T errestrial Networks, ” in 2022 IEEE 95th V ehicular T echnology Conference: (VTC2022-Spring) , 2022, pp. 1–5. [3] A. Guidotti, A. V anelli-Coralli, and C. Amatetti, “Federated Cell-Free MIMO in Non-terrestrial Networks: Architectures and Performance, ” IEEE T ransactions on Aer ospace and Electr onic Systems , vol. 60, no. 3, pp. 3319–3347, 2024. [4] S. Rao, A. Ashikhmin, and H. Y ang, “”internet of things based on cell- free massive mimo”, ” in 2019 53r d Asilomar Conference on Signals, Systems, and Computers , 2019, pp. 1946–1950. [5] B. Chong, F . Guo, H. Lu, and L. Qin, “On the Distribution of SINR for Cell-Free Massiv e MIMO Systems, ” IEEE T ransactions on Communi- cations , vol. 73, no. 2, pp. 832–845, 2025. [6] 5G-ST ARDUST , “D3.2 - End-to-end ground and space integrated archi- tecture, ” 2024. [7] A. V aswani, N. Shazeer, N. P armar, J. Uszk oreit, L. Jones, A. N. Gomez, L. Kaiser , and I. Polosukhin, “Attention is All you Need, ” in Annual Conference on Neural Information Pr ocessing Systems 2017, December 4-9, 2017, Long Beach, CA, USA , 2017, pp. 5998–6008. [Online]. A vailable: https://proceedings.neurips.cc/paper/ 2017/hash/3f5ee243547dee91fbd053c1c4a845aa- Abstract.html [8] 5G-ST ARDUST , “D4.5 - Final report on the unified air interface, user- centric and beamforming solutions, ” 2025. [9] 3GPP , “TR 38.811 - Study on New Radio (NR) to support non-terrestrial networks (Release 15), ” 2020. [10] ——, “TR 38.821 - Solutions for NR to support non-terrestrial networks (NTN), ” 2021. [11] D. P . Kingma and J. Ba, “Adam: A Method for Stochastic Optimiza- tion, ” in 3r d International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference T rack Pr oceedings , Y . Bengio and Y . LeCun, Eds., 2015. [12] S. F . M. Schia vina and K. MacManus, “GHS-POP R2022A - GHS pop- ulation grid multitemporal (1975-2030), ” 2024, European Commission, Joint Research Centre (JRC).