A hybrid transformer–zero-shot learning framework with Muon optimization for intelligent channel estimation in MIMO wireless systems Authors: Wessam M Salama, Moustafa H Aly, Samah Alshathri, Walid El-Shafai Published: 2026 ABSTRACT AbstractFor MIMO wireless systems, accurate channel estimation is essential. However, traditional and current Deep Learning (DL) techniques have poor generalization to unknown situations and necessitate repeated retraining. For intelligent MIMO channel estimation, this study suggests a novel hybrid framework that combines Transformer designs, Zero-Shot Learning (ZSL), and the Muon optimizer. By projecting channel instances into a shared semantic-attribute space, the ZSL component allows precise inference under previously unknown SNR levels and fading conditions without retraining. The Muon optimizer offers better generalization and faster convergence than Adam, while the Transformer uses self-attention to capture intricate spatial-temporal connections. ZSL–Muon, Transformer–ZSL, and the whole Transformer–ZSL–Muon model are the three configurations that are assessed. The whole hybrid consistently outperforms LS, MMSE, CNN-, GRU-, and Transformer-only baselines in MSE over a broad SNR range, according to extensive simulations under quasi-static and time-varying fading. The proposed framework is rigorously evaluated over a wide SNR range and under both quasi-static block fading and time-varying fading channel models. Experimental results demonstrate that the hybrid Transformer–ZSL–Muon model consistently surpasses traditional estimators (e.g., LS and MMSE) and existing DL-based techniques in terms of Mean Squared Error (MSE), particularly in complex and unseen channel conditions. A remarkable finding reveals that increasing the number of antennas from 2 to 128 at an SNR of 30 dB results in a 94.44% MSE reduction, underscoring the potential of massive MIMO implementations. Furthermore, the proposed framework achieves approximately 93.75% MSE improvement over MMSE and 96.67% over LS under the same SNR conditions. These outcomes affirm the effectiveness of combining semantic-aware inference, attention-driven Transformer architectures, and adaptive optimization to deliver a scalable, intelligent, and resilient solution for next-generation MIMO wireless communication systems.Keywords: Zero-shot learning (ZSL), Transformer networks, Muon optimizer, Channel estimation, MIMO systems, DL, Mean squared error (MSE), Signal-to-noise ratio (SNR), Optimization algorithms, Semantic embeddingsSubject terms: Engineering, Mathematics and computing Introduction Accurate channel estimation is a cornerstone of wireless communication system optimization, as it directly enables the implementation of effective power allocation, modulation schemes, and channel coding strategies. Among the various technological advances in this domain, Multiple-Input Multiple-Output (MIMO) systems have emerged as a key enabler of high-throughput, reliable, and spectrally efficient communications. By exploiting multipath propagation through spatial diversity and multiplexing, MIMO systems, utilizing multiple antennas at both the transmitter and receiver ends, can significantly enhance overall system capacity, spectral efficiency, and link reliability1,2. These benefits are particularly vital in the context of dense user environments and data-intensive applications. Modern wireless communication standards such as Wi-Fi, Long-Term Evolution (LTE), and Fifth-Generation (5G) networks have widely adopted MIMO architectures due to their performance benefits. A fundamental requirement for the successful operation of these systems is the availability of accurate and timely Channel State Information (CSI). Precise CSI allows the system to adapt transmission parameters dynamically, enabling not only interference mitigation and improved robustness but also energy-efficient resource allocation across varying propagation conditions3–8. As wireless networks evolve toward 6G and beyond, the demand for intelligent and scalable channel estimation solutions becomes increasingly critical. Despite its critical importance, accurate channel estimation in MIMO systems remains a technically challenging task due to a variety of factors, including multipath fading, spatiotemporal correlations, suboptimal pilot design, complex antenna configurations, and inherent hardware limitations9–12. These challenges often manifest in the form of signal delays, amplitude fluctuations, and interference caused by overlapping multipath components. As a result, the task of acquiring precise and independent CSI becomes significantly more difficult in practical deployment scenarios, especially in highly dynamic or mobile environments. To address these challenges, a broad spectrum of channel estimation techniques has been developed. These include traditional pilot-based estimators, such as Least Squares (LS) and Minimum Mean Square Error (MMSE) methods1,13, which are valued for their analytical simplicity but are often limited in performance under complex or time-varying conditions. More sophisticated model-based approaches such as Maximum Likelihood Estimation (MLE) and Orthogonal Matching Pursuit (OMP)14,15, as well as feedback-oriented techniques like Channel State Information Feedback (CSIF), have also been explored to improve estimation accuracy. In recent years, DL-based approaches have gained substantial attention for their ability to model non-linear and complex channel behaviors without requiring explicit statistical modeling. Techniques employing Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, Graph Neural Networks (GNNs), and more recently, Transformer architectures, have demonstrated considerable promise in learning channel representations directly from raw or pilot-based inputs. These DL models are capable of performing end-to-end CSI reconstruction, offering enhanced adaptability and estimation accuracy, especially in large-scale MIMO settings and rapidly changing environments16–19. Research problem In MIMO wireless communication systems, accurate channel estimation plays a pivotal role in enabling the optimization of key transmission parameters such as power allocation, modulation strategies, and coding schemes. Without reliable CSI, system performance significantly degrades, particularly in terms of spectral efficiency, data throughput, and error rates. However, existing channel estimation techniques, including both classical model-based methods and contemporary DL approaches like CNNs and Gated Recurrent Units (GRUs), face substantial limitations in their ability to generalize across dynamically changing channel environments. These environments are often characterized by diverse fading profiles, varying signal-to-noise ratio (SNR) conditions, and different levels of user mobility. The core issue lies in the lack of adaptability and generalization in these traditional models. Most of them are trained in a supervised, scenario-specific manner, which necessitates frequent retraining or fine-tuning when encountering unseen channel conditions. This requirement poses a major barrier to real-time deployment, especially in rapidly evolving wireless environments such as those expected in 6G networks, vehicular communication systems, and IoT-enabled smart infrastructure. This study addresses the pressing challenge of designing a generalizable, robust, and adaptive channel estimation framework that can accurately infer CSI in unseen or untrained scenarios without incurring the cost of retraining. Specifically, the objective is to develop a model that can semantically interpret and project channel characteristics into a latent feature space, enabling zero-shot generalization to new channel conditions. At the same time, the model must maintain training stability, achieve fast convergence, and exhibit resilience across a wide SNR spectrum, making it suitable for scalable, intelligent wireless communication systems. Motivation The exponential growth in demand for reliable, high-speed, and ultra-low-latency wireless communication, driven by emerging technologies such as 6G networks, autonomous vehicles, smart cities, and the Internet of Things (IoT), has intensified the need for highly accurate, adaptive, and flexible channel estimation methods. These next-generation applications require systems that can operate seamlessly across a wide spectrum of environmental conditions, mobility scenarios, and interference levels. Accurate channel estimation is therefore a critical enabler of robust and intelligent communication infrastructure. While traditional model-based estimators and DL approaches, including CNNs and GRUs, have made strides in improving channel estimation performance, they are inherently constrained by several critical limitations. Most notably, these models rely heavily on large, diverse training datasets, are sensitive to variations in channel statistics, and lack the capacity to generalize effectively to unseen or evolving channel conditions. As wireless environments become increasingly heterogeneous and time-varying, retraining or fine-tuning models for each new scenario becomes both computationally expensive and operationally infeasible. To address these challenges, this research is motivated by three cutting-edge techniques that together promise to deliver a transformative solution. First, the Muon optimizer is introduced to improve training dynamics by accelerating convergence, enhancing generalization, and preventing overfitting through adaptive regularization and dynamic learning rate adjustment. Second, the integration of Zero-Shot Learning (ZSL) allows the model to transfer learned knowledge from known channel scenarios to previously unseen ones, thereby eliminating the need for retraining and enhancing scalability. Third, the incorporation of Transformer architectures introduces the ability to capture long-range dependencies and intricate temporal–spatial correlations within MIMO channels through self-attention mechanisms and context-aware feature extraction. The synergistic combination of these technologies, ZSL for semantic generalization, Transformer for sequence modeling, and Muon for optimization, forms a novel paradigm for intelligent channel estimation. This paradigm is not only computationally efficient and broadly applicable but also inherently robust to future wireless system variations, positioning it as a compelling solution for next-generation communication networks. Research objectives The primary objective of this study is to develop and evaluate a novel and generalizable channel estimation framework for MIMO wireless communication systems by exploring and comparing three core configurations: (1) Transformer integrated with ZSL and Muon optimizer, (2) Transformer with ZSL alone, and (3) ZSL combined with the Muon optimizer. The overarching aim is to construct a reliable, scalable, and semantically aware DL framework capable of accurate channel state prediction under a wide range of operational conditions, without the need for retraining when exposed to new channel environments. To achieve this, the study is guided by the following specific objectives: Design a semantic-aware DL model capable of reasoning under unseen channel conditions by projecting MIMO channel characteristics into a shared latent space that captures both physical signal properties and high-level semantic representations. Design a semantic-aware DL model capable of reasoning under unseen channel conditions by projecting MIMO channel characteristics into a shared latent space that captures both physical signal properties and high-level semantic representations. Integrate Transformer architectures with both the ZSL mechanism and Muon optimization strategy to enhance the model’s representational capacity, improve attention-driven feature extraction, and accelerate convergence during training. Integrate Transformer architectures with both the ZSL mechanism and Muon optimization strategy to enhance the model’s representational capacity, improve attention-driven feature extraction, and accelerate convergence during training. Eliminate the need for retraining across varying channel conditions, such as changes in SNR, fading types, or propagation patterns, by leveraging the generalization capabilities of the ZSL framework, enabling zero-shot inference. Eliminate the need for retraining across varying channel conditions, such as changes in SNR, fading types, or propagation patterns, by leveraging the generalization capabilities of the ZSL framework, enabling zero-shot inference. Conduct a comparative evaluation of the Muon and Adam optimizers to demonstrate the effectiveness of Muon in training stability, generalization across diverse channel scenarios, overfitting mitigation, and convergence speed. Conduct a comparative evaluation of the Muon and Adam optimizers to demonstrate the effectiveness of Muon in training stability, generalization across diverse channel scenarios, overfitting mitigation, and convergence speed. Utilize a comprehensive and realistic MIMO channel dataset, encompassing a wide range of SNR values, delay spreads, Doppler shifts, and fading scenarios (both quasi-static and time-varying), to conduct a rigorous assessment of the proposed framework’s performance and generalization ability. Utilize a comprehensive and realistic MIMO channel dataset, encompassing a wide range of SNR values, delay spreads, Doppler shifts, and fading scenarios (both quasi-static and time-varying), to conduct a rigorous assessment of the proposed framework’s performance and generalization ability. By fulfilling these objectives, this study aims to contribute a forward-looking, intelligent solution for channel estimation in rapidly evolving wireless environments, supporting the advancement of robust and adaptive communication technologies. Research contribution To address the limitations of traditional model-based estimators and existing DL approaches in channel estimation, this paper introduces a novel, intelligent framework for MIMO wireless communication systems. The proposed solution integrates Transformer architectures, ZSL, and the Muon optimizer to form a unified estimation paradigm that is both generalizable and efficient. The key research contributions are outlined as follows: The first application of ZSL to MIMO channel estimation, allowing precise CSI inference for mobility situations, fading types, and unseen SNR levels without retraining. The first application of ZSL to MIMO channel estimation, allowing precise CSI inference for mobility situations, fading types, and unseen SNR levels without retraining. A semantic-aware hybrid architecture is introduced that successfully models long-range spatial–temporal connections and transferable semantic representations by combining ZSL with Transformer-based attention mechanisms. A semantic-aware hybrid architecture is introduced that successfully models long-range spatial–temporal connections and transferable semantic representations by combining ZSL with Transformer-based attention mechanisms. Incorporation of the Muon optimizer, which performs better than Adam for non-stationary wireless data in terms of generalization, training stability, and convergence speed. Incorporation of the Muon optimizer, which performs better than Adam for non-stationary wireless data in terms of generalization, training stability, and convergence speed. The full Transformer–ZSL–Muon framework achieves state-of-the-art MSE performance better than LS and MMSE under both quasi-static and time-varying fading, with especially significant gains in massive MIMO configurations, according to a thorough evaluation and ablation study. The full Transformer–ZSL–Muon framework achieves state-of-the-art MSE performance better than LS and MMSE under both quasi-static and time-varying fading, with especially significant gains in massive MIMO configurations, according to a thorough evaluation and ablation study. While CNNs are good at extracting localized spatial features but struggle with long-range dependencies and need retraining for unseen conditions, and LSTMs are good at handling sequential temporal data but suffer from vanishing gradients, slow sequential processing, and poor parallelization. On the other hand, our hybrid approach uniquely combines Transformer’s global self-attention for semantic projection into a shared attribute space (enabling generalization to novel SNR levels, fading types, and mobility without retraining) as opposed to supervised CNN/LSTM techniques), and the Muon optimizer for dynamic learning rate/regularization adjustments that guarantee quicker convergence and better stability on noisy wireless data (lower final loss than Adam), with graceful scaling to massive MIMO and 24–31% MSE reduction over recent CNN baselines and 18–27% over LSTM baselines across 0–40 dB SNR under quasi-static/time-varying fading. The suggested Transformer–ZSL–Muon framework has several clear advantages over CNN- and LSTM/GRU-based channel estimation techniques. The Transformer uses global self-attention to capture arbitrary-range spatial and temporal dependencies in a fully parallelizable manner, achieving two to three times faster training and inference, in contrast to CNNs, which are constrained by small receptive fields, and LSTMs/GRUs, which process sequences sequentially and suffer from vanishing gradients over long sequences; The integrated ZSL component projects channel instances into a semantic attribute space (encoding SNR range, delay spread, Doppler shift), enabling accurate inference without any retraining; additionally, the Muon optimizer dynamically modifies learning rate and regularization, delivering faster convergence and greater stability on noisy wireless data than conventional Adam/SGD, resulting in significantly lower final loss; CNN- and LSTM-based methods require retraining or fine-tuning when facing unseen SNR levels, fading types, or mobility conditions. Paper structure The remainder of this paper is structured to systematically develop and validate the proposed framework. Section 2 reviews the state-of-the-art in channel estimation, including both traditional techniques and recent advancements in DL, semantic modeling, and optimization strategies. Section 3 outlines the system model used in this study, detailing the MIMO communication setup, mathematical formulation of the channel estimation process, and the integration of pilot symbols. In Sect. 4, the DL-based estimation framework is introduced, including the architecture of the Transformer–ZSL–Muon model, the process of dataset generation, semantic embedding techniques, and training procedures. Section 5 presents the simulation results and comprehensive performance evaluations under various SNR levels, fading conditions, and antenna configurations, comparing the proposed model with conventional and DL-based baselines. Finally, Sect. 6 concludes the paper by summarizing the key contributions, highlighting the effectiveness of the proposed method, discussing current limitations, and suggesting future research directions for intelligent and adaptive channel estimation in next-generation wireless systems. Related work Channel estimation has long been a central challenge in the design and optimization of MIMO wireless communication systems. Over the past two decades, research in this area has evolved from traditional model-based estimation techniques to more recent data-driven DL frameworks that aim to enhance estimation accuracy and robustness under complex, real-world conditions. This section provides a structured and critical review of the key contributions across four main domains: (1) classical statistical estimators such as LS and MMSE; (2) advanced DL architectures including CNNs, RNNs, and Transformer-based models; (3) optimization techniques used to improve training performance and model generalization; and (4) semantic learning approaches, with a particular focus on ZSL as a mechanism for generalization to previously unseen channel scenarios. The following subsections examine these developments, highlighting their respective strengths, limitations, and applicability to next-generation wireless systems. Traditional channel Estimation model Traditional channel estimation techniques, particularly the LS and MMSE estimators, have been extensively adopted in MIMO systems due to their analytical simplicity, low computational complexity, and ease of implementation20. The LS estimator is favored for its independence from channel statistics, while MMSE improves estimation accuracy by incorporating statistical information about the noise and channel covariance. Despite their popularity, these model-based methods exhibit several limitations that restrict their performance in modern communication environments. One of the primary drawbacks of these estimators is their vulnerability to noise, especially at low signal-to-noise ratios (SNRs). Moreover, MMSE relies heavily on accurate prior knowledge of the channel’s statistical properties, which are often either unavailable or outdated in real-world deployments. This dependency can result in significant performance degradation when channel conditions change unpredictably. Additionally, both LS and MMSE perform suboptimally in high-mobility scenarios, where channel conditions vary rapidly over time, or in frequency-selective fading environments, which are typical in broadband communications. These limitations render traditional estimators less suitable for next-generation wireless systems, such as 5G and beyond, which require robust, adaptive, and low-latency solutions capable of operating under dynamic and heterogeneous network conditions. Channel estimation based on DL models Recent advancements in DL have opened new avenues for addressing the limitations of traditional channel estimation techniques in MIMO systems. DL-based models are capable of learning complex, non-linear mappings from input pilot signals to corresponding channel coefficients, offering greater adaptability to real-world conditions without relying on rigid statistical assumptions. These models can effectively leverage both spatial and temporal correlations inherent in wireless channels, thereby enhancing estimation performance in diverse propagation environments. For instance, the study presented in Ref21. proposed a DL-assisted channel estimation method for large-scale MIMO systems, integrating depth image priors for denoising alongside conventional LS estimation. This hybrid approach aimed to mitigate multi-cell interference while improving estimation accuracy under constrained conditions. CNNs have been widely adopted for exploiting spatial structures in the channel matrix and have shown promising results in scenarios with relatively static or slowly varying channel conditions. These networks excel at learning localized features and have been particularly effective when applied to image-like representations of CSI. In contrast, Recurrent Neural Networks (RNNs), specifically LSTM networks and GRUs, have been employed to model temporal dependencies in time-varying fading channels. These architectures are capable of capturing the sequential nature of channel variations, making them well-suited for mobile or rapidly changing wireless environments. However, while DL models exhibit superior estimation accuracy compared to classical methods, their performance often depends on the availability of large, diverse training datasets and extensive offline training. Moreover, most of these models still face challenges in generalizing to previously unseen conditions, such as novel SNR ranges or fading types, which limits their real-time applicability in dynamic wireless systems Table 1. Table 1MIMO CE classified DL architectures.CategoryReferencesStrengthsLimitationsRelevance to our workCNN-based27,28Low complexity and effective spatial denoisingInadequate long-range temporal modeling; undetected fading requires retrainingOur transformer performs 24–31% MSE better than the baseline for ablation.RNN/LSTM-based29,30Uses sequences to handle time-varying channelsGradients disappearing; sequential → slow trainingFigure 5’s comparison emphasizes the need for parallelizable substitutes like ours.Transformer-based31,32MIMO correlations via global self-attentionHigh parameters; restricted extrapolation to unknown SNRLays the groundwork; we add ZSL for zero-shotHybrid33,34Combines CS/GAN and DL for resilienceNo semantic transfer; still under supervisionConfirms the hybrid theme; our ZSL-Muon adds an invisible modification MIMO CE classified DL architectures. Optimization techniques for DL-based channel estimation While DL models have shown remarkable potential for channel estimation in MIMO systems, their effectiveness heavily depends on the choice of optimization algorithm during training. Selecting a suitable optimizer is critical for ensuring fast convergence, model stability, and robust generalization, particularly when working with high-dimensional and noisy wireless signal data. This challenge becomes even more significant when dealing with temporal signal representations, where the model must extract meaningful features while being resilient to a wide range of interference and distortion sources. Traditional optimization algorithms, such as Stochastic Gradient Descent (SGD) and Adam, are widely used due to their simplicity and computational efficiency. However, they are often prone to getting trapped in local minima and exhibit sensitivity to hyperparameter tuning, especially learning rate selection. These limitations can lead to slow convergence, unstable training, and poor generalization performance, particularly in dynamic and non-stationary wireless environments. To address these challenges, several advanced optimizers have been proposed in recent years, including Lookahead, AdaBelief, and the recently introduced Muon optimizer22. These methods aim to enhance training efficiency by incorporating features such as adaptive momentum, gradient smoothing, and dynamic regularization mechanisms. Among them, the Muon optimizer stands out for its ability to accelerate convergence and mitigate overfitting in complex DL models, particularly when training data is sparse or highly variable, as is often the case in wireless channel estimation Table 2. Table 2DL-CE classified optimizers.CategoryReferencesTechniqueBenefits in CEGaps addressed by MuonStandard 35,36 Fixed momentum/decayEasy to use and quick for convex lossesInability to adjust to non-stationary noise Muon decreases loss by 30% (Fig. 8).Adaptive momentum 37 Gradient smoothingA better way to avoid local minimaNot enough for a small number of pilots Muon increases perception of curvatureBio-inspired/meta[39]Population-based searchHyperparameter optimization on a global scaleMuon is computationally heavy, yet it is lightweight and converges two times faster. DL-CE classified optimizers. A comparative summary of how the proposed framework addresses key gaps in the literature is presented in Table 3. As shown, most existing DL-based estimators require retraining when exposed to new channel conditions, lack robust semantic representation mechanisms, and suffer from convergence issues. The proposed model overcomes these limitations by incorporating ZSL for generalization, semantic embeddings for transferable knowledge representation, and the Muon optimizer to ensure efficient and stable training. Table 3Comparison of gaps in related work and how the proposed framework addresses them.Literature gapProposed framework contributionMost DL models require retraining for new channel conditionsIntroduces ZSL to handle previously unseen SNR levels and fading types without retrainingCNN/GRU models generalize poorly in non-stationary environmentsEmploys semantic embeddings to enable efficient knowledge transfer across diverse scenariosDeep models suffer from overfitting and slow convergenceUtilizes the Muon optimizer to enhance robustness, generalization, and training speedLack of unified models integrating semantic reasoning and channel estimationProposes a novel ZSL-based semantic framework tailored for MIMO channel estimation Comparison of gaps in related work and how the proposed framework addresses them. System model This section establishes the foundational signal model and system architecture for the MIMO wireless communication framework under investigation. We begin by presenting the mathematical formulation of the received signal in a flat-fading MIMO environment, emphasizing the role of the channel matrix H, which characterizes the complex propagation paths between transmit and receive antennas. A detailed pilot-based formulation is introduced to enable accurate channel estimation during the training phase. Classical estimation techniques, including the LS and MMSE estimators, are reviewed for benchmarking purposes, highlighting their strengths and limitations under practical conditions. Furthermore, the signal flow through the transmitter, channel, and receiver is depicted through a comprehensive system diagram, illustrating key components such as modulation, pilot insertion, cyclic prefix handling, and equalization. The integration of DL models, specifically the Transformer-ZSL-Muon variants, into the receiver chain is emphasized as a significant enhancement over traditional techniques, offering improved adaptability, robustness, and estimation performance across diverse channel conditions. Mathematical representation of the MIMO channel In this study, we consider a narrowband flat-fading MIMO wireless communication system, where the transmitter and receiver are equipped with ​ and ​ antennas, respectively. The received signal over a block of T symbol periods can be mathematically modeled as follows23: Here: denotes the received signal matrix, collected at the receiver over T time slots. denotes the received signal matrix, collected at the receiver over T time slots. represents the transmitted signal matrix containing pilot or training symbols. represents the transmitted signal matrix containing pilot or training symbols. ​ is the MIMO channel matrix, capturing the complex channel coefficients between each transmit and receive antenna pair. ​ is the MIMO channel matrix, capturing the complex channel coefficients between each transmit and receive antenna pair. denotes the additive white Gaussian noise (AWGN) matrix, whose entries are modeled as independent and identically distributed complex Gaussian random variables with zero mean and variance σ2, i.e., CN(0, σ2). denotes the additive white Gaussian noise (AWGN) matrix, whose entries are modeled as independent and identically distributed complex Gaussian random variables with zero mean and variance σ2, i.e., CN(0, σ2). This canonical MIMO signal model is widely used to describe the input-output relationship under flat fading conditions, where each transmitted symbol experiences a constant channel coefficient over the entire duration of a single transmission block. Accurate estimation of the channel matrix is essential for enabling coherent detection, equalization, and overall optimization of transmission parameters in such systems. Characterization of the channel matrix The core of the MIMO system’s behavior is governed by the channel matrix ​​, which encapsulates the complex-valued fading coefficients between each pair of transmit and receive antennas. Specifically, each element hij∈H represents the channel gain from the jth transmit antenna to the ith receive antenna, capturing both the amplitude attenuation and phase shift induced by the wireless propagation environment. To accurately estimate the channel matrix during the training phase, a sequence of known pilot symbols is transmitted. Let ​denote the pilot signal matrix, where Tp is the number of pilot time slots. These pilot signals are carefully designed and transmitted prior to data transmission to facilitate reliable channel estimation at the receiver. The received pilot signal at the receiver, affected by the channel and noise, is modeled as: where: is the received signal matrix corresponding to the pilot phase, is the received signal matrix corresponding to the pilot phase, ​ is the known pilot matrix, ​ is the known pilot matrix, ​ represents the AWGN matrix during pilot transmission, whose entries are modeled as CN(0, σ2). ​ represents the AWGN matrix during pilot transmission, whose entries are modeled as CN(0, σ2). This pilot-based formulation forms the basis for subsequent estimation of the unknown channel matrix H, which is essential for coherent demodulation and equalization at the receiver. In the context of pilot-based MIMO systems, several conventional techniques have been widely adopted to estimate the channel matrix H. Among the most prominent are the LS and MMSE estimators, which offer a balance between computational efficiency and estimation accuracy. LS Estimator: The LS estimator seeks to minimize the squared error between the received pilot signal and the predicted signal. It does not require prior knowledge of the channel or noise statistics and is computed as: LS Estimator: The LS estimator seeks to minimize the squared error between the received pilot signal and the predicted signal. It does not require prior knowledge of the channel or noise statistics and is computed as: where is the Moore–Penrose pseudoinverse of the pilot matrix . While LS is simple and easy to implement, it is known to perform poorly under low SNR conditions and does not exploit statistical information about the channel or noise. MMSE Estimator: To enhance estimation performance, particularly under noisy environments, the MMSE estimator incorporates statistical knowledge of both the channel and the noise. It aims to minimize the mean square error between the true and estimated channel matrices and is defined as: MMSE Estimator: To enhance estimation performance, particularly under noisy environments, the MMSE estimator incorporates statistical knowledge of both the channel and the noise. It aims to minimize the mean square error between the true and estimated channel matrices and is defined as: where is the cross-covariance matrix between the true channel H and the received pilot signal ,while is the autocovariance matrix of ​. Although MMSE generally outperforms LS in terms of estimation accuracy, especially at lower SNRs, it requires prior knowledge of the channel and noise covariance statistics, which may not always be available or easy to estimate in practical deployments. The complete end-to-end signal transmission process in a typical MIMO wireless communication system is illustrated in Fig. 1. This process is divided into three primary stages: transmitter processing, channel transmission, and receiver processing. Each stage plays a critical role in ensuring the accurate and efficient delivery of information over complex wireless environments. The MIMO communication system signal transmission procedure. At the transmitter side, the communication pipeline begins with the generation of data symbols, which are modulated using a suitable modulation scheme (e.g., QAM or PSK). The modulated symbols are then converted from serial-to-parallel (S/P) format to match the multiple spatial streams supported by the MIMO configuration. Pilot symbols are subsequently inserted into the data stream to enable reliable channel estimation at the receiver. Following this, the signal undergoes Inverse Fast Fourier Transform (IFFT) to convert the frequency-domain data into the time domain, making it suitable for transmission over physical channels. A cyclic prefix (CP) is added to combat inter-symbol interference (ISI), and the parallel data streams are converted back to serial form (P/S) before being transmitted through multiple antennas. During channel transmission, the signal is affected by multipath fading, additive noise, and interference. These phenomena introduce distortions, which must be mitigated at the receiver to ensure correct data recovery. On the receiver side, the incoming signal streams undergo cyclic prefix removal, followed by Fast Fourier Transform (FFT) to return to the frequency domain. The signals are then reorganized via S/P conversion, making them ready for processing by the DL-based channel estimation and equalization models. In the proposed framework, three variants are explored: Transformer-ZSL-Muon, Transformer-ZSL, and ZSL-Muon, all designed to enhance robustness under challenging conditions such as noise and fading. These models enable accurate channel estimation and equalization even in dynamically changing wireless environments. After equalization, the signals are converted back to parallel-to-serial (P/S) form and passed through a demodulation stage to recover the original transmitted symbols. The integration of DL models into the receiver chain marks a significant advancement in MIMO system design, offering enhanced adaptability, resilience, and performance compared to traditional model-based techniques. The hybrid DL model This section introduces a novel hybrid DL framework that combines Transformer architectures, ZSL, and the Muon optimizer to enhance channel estimation in MIMO wireless communication systems. The model is designed to tackle the limitations of traditional estimators and conventional DL techniques, specifically their inability to generalize across unseen channel conditions without retraining. The section begins with a detailed data generation strategy, which simulates realistic wireless environments under quasi-static and time-varying fading conditions, while embedding physical and semantic features to support ZSL-based generalization. The channel estimation process is then formalized through a learnable DL model that minimizes the MSE between the true and predicted channel coefficients across both static and dynamic fading regimes. The ZSL framework is introduced to enable inference under novel channel scenarios by projecting observed features into a shared semantic space using a lightweight encoder. Semantic embeddings are used to facilitate robust estimation even when the channel conditions differ from those seen during training. This is further reinforced through a semantic consistency loss, dropout regularization, and structured latent space constraints. Next, the Muon optimizer is described, offering adaptive learning rate control and regularization during training. It dynamically adjusts hyperparameters such as learning rate and weight decay based on the loss landscape, outperforming standard optimizers like Adam in convergence speed and generalization. The section also details the integration of the Transformer-ZSL-Muon framework, illustrating its training flow, semantic feature extraction, and generalization behavior using a dataset of 150,000 samples. In addition, variations of the model, including ZSL with Muon optimizer and Transformer-ZSL using Adam optimizer, are examined to highlight the role of each component. Tables and figures (e.g., Tables 3 and 6, and Fig. 2) are provided to summarize training configurations and simulation setups. Table 6Average training time comparison of the proposed models under identical hardware and dataset conditions.ModelConsuming timeTransformer-ZSL~ 7.5–7.7 hTransformer-ZSL with Muon~ 7.9–8.1 hZSL with Muon optimizer only~ 7.2–7.4 h Average training time comparison of the proposed models under identical hardware and dataset conditions. The proposed Transformer–ZSL–Muon framework and its ablation versions overall training and inference flowchart. Forward pass is indicated by solid arrows; backpropagation and parameter updates (only during training) are indicated by dashed red arrows. In essence, this section demonstrates how the synergistic combination of Transformer-based feature encoding, semantic reasoning via ZSL, and adaptive optimization through Muon results in a scalable, robust, and generalizable channel estimation framework suitable for future intelligent MIMO systems. Data generation and description To rigorously evaluate the proposed hybrid DL model, a synthetic dataset is constructed under two representative wireless channel scenarios: quasi-static block fading and time-varying multipath fading. These scenarios are designed to emulate realistic channel dynamics encountered in practical MIMO systems. In both cases, the transmission environment is modeled with AWGN, which introduces controlled levels of stochastic interference across the communication channel. For each antenna stream, the simulation generates detailed records of pilot signal patterns, corresponding channel responses, and input-output symbol mappings. These components are essential for supervised training and evaluation in conventional DL settings. However, in this framework, we adopt a ZSL paradigm that removes the need for exhaustive label-specific supervision. Instead of relying on traditional class labels, the input features are encoded into high-level semantic attribute representations, which describe the underlying physical and statistical characteristics of the channel, such as SNR ranges, fading types, or temporal behavior. These semantic descriptors serve as the basis for transferring knowledge from seen conditions to unseen channel environments, thereby facilitating robust and generalizable channel estimation. To support the training and evaluation of the proposed channel estimation framework, a carefully designed dataset is constructed, with representative samples provided in Table 4. Each row in the dataset corresponds to a unique communication instance within a simulated MIMO environment, capturing both the physical characteristics of the wireless channel and the semantic attributes required for ZSL. Table 4Samples of the utilized dataset.Sample IDTx antennaRx antennaPilot smbolsSNR (dB)Delay spread (ns)Doppler shift (Hz)Semantic attributesChannel response (Real)Channel response (Imag)111[0.5, −0.3, 1.0]105030[0.2, 0.1, 0.4, 0.8]0.623−0.115212[1.0, −0.7, 0.9]154025[0.3, 0.2, 0.5, 0.7]0.5980.063321[−0.2, 0.8, −1.0]206010[0.6, 0.4, 0.1, 0.9]−0.7020.42422[0.3, −0.1, 0.5]255535[0.4, 0.5, 0.2, 0.6]0.741−0.309511[−0.9, 0.2, 0.7]303520[0.7, 0.6, 0.3, 0.5]0.8120.158 Samples of the utilized dataset. The dataset includes key input parameters such as the number of transmit (Tx) and receive (Rx) antennas, pilot symbol vectors, SNR in decibels, delay spread in nanoseconds, and Doppler shift in hertz. These features reflect the inherent complexity and variability of wireless channels and are essential for building robust estimation models. A central innovation in this dataset design is the inclusion of semantic attribute vectors, represented as 4-dimensional real-valued descriptors. These semantic embeddings are constructed to abstract high-level characteristics of the wireless environment, such as fading type, user mobility, and noise behavior. By enabling the model to generalize from these semantic descriptions rather than relying solely on instance-specific labels, the dataset facilitates effective implementation of the ZSL framework. This design allows the model to transfer knowledge to unseen scenarios, a key requirement for intelligent and adaptive channel estimation in dynamic settings. In addition, the channel response is represented as a complex-valued output, with both real and imaginary components provided. This accurately reflects the nature of MIMO channel coefficients, which carry both amplitude and phase information essential for coherent detection. The diversity of the dataset, particularly in terms of antenna configurations, SNR values, delay spreads, and Doppler effects, ensures that the model is exposed to a wide range of conditions during training. Such comprehensive coverage is vital for assessing the generalization capacity of the Transformer-ZSL-Muon framework, especially under previously unseen and time-varying channel scenarios. A very difficult and realistic generalization scenario that goes well beyond basic interpolation is purposefully created by the updated zero-shot approach shown in Table 5. Only five discrete SNR points (5, 10, 15, 20, and 25 dB), four distinct Doppler values, four typical delay-spread profiles, and moderate mobility (≤ 120 km/h) are presented to the model during training. Twelve completely unseen SNR levels (including extreme low-SNR 0–2 dB and high-SNR 35–40 dB), previously unseen Doppler spreads (up to 300 Hz), four completely new propagation environments (such as high-speed rail and tunnel scenarios never observed during training), and extreme velocities of 300–500 km/h must all be estimated during the test. Furthermore, extreme combinations of these characteristics (such as 0 dB SNR + 300 Hz Doppler + tunnel profile) that are physically realistic in 6G deployments but never seen during training are included in the unseen test set. This rigorous division guarantees that any respectable performance on the unseen partition can only be attained by true semantic knowledge transfer made possible by the ZSL embedding, not by interpolation or memorization. The suggested ZSL-based models maintain robust performance (≤ 3 dB degradation), definitively demonstrating the efficacy of the semantic zero-shot mechanism in truly unseen and harsh channel environments. Table 5Generation of datasets and zero-shot protocol.AspectOriginal settingSNR range (training)0–40 dB continuousSNR range (testing – unseen)0–40 dBDoppler spread (training)10–300 HzDoppler spread (testing)Same rangeDelay spread profiles8 standard profilesMobility scenariosMixedTotal seen samples~ 150 kTotal unseen samples~ 50 k Generation of datasets and zero-shot protocol. Channel estimation under quasi-static and time-varying conditions In MIMO wireless communication systems, accurate channel estimation is closely tied to the temporal behavior of the wireless channel. When the receiver exhibits minimal mobility, the channel coherence time, the time duration over which the channel response remains relatively constant, tends to be significantly longer than the codeword duration. This scenario corresponds to the quasi-static block fading condition, in which the channel can be assumed constant over each transmission block but varies independently across successive blocks. In such low-mobility settings, the effects of Doppler frequency shifts and rapid time variations in the channel response become negligible, allowing for simplified estimation strategies. However, even when relative motion between transmitter and receiver is minimal, the wireless channel may still experience multipath propagation effects due to static reflectors, diffractors, and scatterers in the environment. These factors give rise to delay spreads and frequency-selective fading within each block. Moreover, phenomena such as shadow fading, caused by physical obstructions in the signal path, and spatial fading, due to the constructive and destructive interference of multipath components, may still affect signal quality, even in quasi-static conditions. These impairments underscore the necessity of employing robust channel estimation models capable of handling both stable and variable conditions. In the considered MIMO framework, it is assumed that the channel remains constant across a given symbol block but varies independently between blocks, aligning with the block fading model. This assumption simplifies the modeling of pilot-based training and makes it feasible to integrate the proposed Transformer-ZSL-Muon model for adaptive estimation, particularly under slowly varying or predictable channel dynamics. In the context of MIMO systems operating under quasi-static block fading conditions, the core challenge lies in accurately estimating the channel coefficients, denoted as h(r), for each receive antenna. The objective is to develop a DL-based channel estimator that minimizes the MSE between the true channel h and its predicted estimate , while also ensuring the effective design of pilot signals that support robust learning under limited feedback. Formally, the channel estimation task can be formulated as the following optimization problem: where L denotes the loss function, and ϕ represents the learnable parameters of the estimation model. The predicted channel output is defined as a function of the input pilot signals s and the DL model W, such that: where W refers to the DL architecture, e.g., Transformer-ZSL-Muon, used to approximate the mapping from the pilot input space to the channel output domain. While quasi-static fading assumes stable channel conditions over the duration of each block, this assumption no longer holds in high-mobility environments, where the channel coherence time becomes shorter than the codeword duration. In such scenarios, the channel coefficients h(r) fluctuate continuously within a transmission block due to Doppler shifts, leading to time-varying fading characteristics. This dynamic behavior imposes additional constraints on channel observation and sampling rates, often resulting in limited and noisy training samples. Consequently, precise channel estimation in high-mobility scenarios requires not only robust DL architectures but also intelligent strategies for resource-constrained observation and sampling optimization. The proposed hybrid model addresses these challenges by employing semantic-aware learning and adaptive optimization techniques, enabling the system to maintain high estimation accuracy even under limited-sample, time-varying conditions. In the time-varying fading scenario, channel estimation becomes more complex due to the temporal correlation of channel fluctuations within and across symbol durations. Specifically, the estimation of the current channel state h(r) relies on feedback information collected over a temporal window, typically from the symbol time to the current rth symbol. This temporal dependence allows the model to capture the evolving channel dynamics. In the context of MIMO systems, the estimation of the full channel matrix H similarly considers historical observations, typically from the to the (n−1)th symbol time, to account for intra-block correlation and fading continuity. To handle unseen or dynamically changing channel conditions, the proposed framework employs a semantic-aware decoder, guided by the ZSL module. This decoder predicts the full channel matrix based not solely on observed signals but also on the semantic representation of the environment, enabling generalization to previously unencountered fading types, SNR levels, or antenna configurations. Even in the absence of direct training on specific channel states, the ZSL component allows the model to infer latent structures and provide accurate channel estimates using attribute-based descriptors. During the training phase, the framework is optimized using the Muon optimizer, which plays a critical role in ensuring efficient convergence, robust generalization, and resistance to overfitting. The training objective remains the minimization of MSE between the true channel matrix H and its predicted counterpart . The Muon optimizer’s use of adaptive regularization and dynamic scheduling techniques ensures that the model navigates the loss landscape effectively, achieving both fast convergence and stable learning across diverse channel conditions. As a result, the hybrid Transformer-ZSL-Muon framework produces a scalable and resilient channel estimation model, suitable for both quasi-static and high-mobility environments. ZSL framework The core of the proposed estimation model is built upon a ZSL framework, which enables the system to generalize beyond seen training conditions by leveraging semantic embeddings. Rather than relying exclusively on signal-label pairs from known channel conditions, the ZSL-based architecture is designed to learn transferable mappings from pilot symbol patterns to channel coefficients. This is achieved by encoding pilot-aided received signals into a semantic vector space that represents both observed and unobserved channel states. The architecture adopts a hybrid encoder–decoder design, where the encoder is a nonlinear projection network that transforms noise-corrupted pilot sequences into compact latent representations, and the decoder uses these latent vectors, conditioned on semantic attributes, to generate channel estimates even for previously unseen scenarios (e.g., unobserved SNR values or fading profiles). This enables the system to operate in dynamically evolving environments without requiring retraining or additional supervision. The semantically aware encoder is a lightweight yet expressive neural network, typically composed of convolutional layers for extracting spatial features, pooling layers to reduce dimensionality, and fully connected (dense) layers that map features into a low-dimensional shared latent space. This latent representation, denoted as , captures the essential structure and characteristics of the MIMO channel. Crucially, the encoder does not learn raw signal values directly; instead, it focuses on robust, semantically rich feature representations that encode underlying physical properties of the wireless environment, such as fading behavior, delay profiles, and SNR levels. Semantic embeddings, vector representations that numerically encode different channel attributes, play a critical role in this framework. These embeddings, denoted by sk​, may be generated through two methods: (i) supervised learning using labeled channel conditions, or (ii) manual definition based on domain expertise. Each embedding corresponds to a distinct wireless scenario, such as a specific SNR range or fading type. During inference, the model compares the latent representation z of the observed channel against the precomputed semantic embeddings using a similarity metric (e.g., cosine similarity or Euclidean distance). The channel estimate is then derived by selecting the semantic embedding that best matches the current latent representation, allowing the decoder to generate accurate predictions under previously unseen conditions. This ZSL-enabled mechanism introduces a semantic reasoning capability into the channel estimation process, significantly enhancing the model’s ability to operate under sparse supervision and generalize across a wide spectrum of MIMO environments. Muon optimizer and training strategy To ensure efficient and robust training of the proposed hybrid channel estimation framework, the Muon optimizer is employed as a superior alternative to conventional optimization techniques such as Adam. Unlike static optimizers, the Muon algorithm integrates motion dynamics into the weight update process by modeling both momentum and frictional forces, thereby providing enhanced navigation through the non-convex loss surfaces typically encountered in ZSL training. This dynamic adjustment capability results in faster convergence, improved stability, and superior generalization performance, especially in highly heterogeneous MIMO channel environments. An important aspect of the training procedure involves the incorporation of a semantic consistency loss, which preserves the structural alignment between seen and unseen channel classes in the shared latent space. This loss function encourages the model to learn feature representations that maintain intra-class compactness and inter-class separability, enabling more accurate channel estimation in new scenarios. To further bolster generalization, dropout is applied at various layers of the encoder–decoder architecture, reducing the risk of overfitting by randomly deactivating neurons during training. Additionally, structural constraints are enforced on the latent space using cosine similarity metrics and margin-based loss functions, which help to refine the representation geometry by maximizing inter-class distances and minimizing intra-class variance. The Muon optimizer exhibits a high degree of adaptivity throughout the training process. It dynamically adjusts the learning rate based on real-time feedback from the training loss landscape. Specifically, the optimizer reduces the learning rate in steep regions or near convergence optima to prevent overshooting, while increasing it in flatter regions to escape plateaus and accelerate convergence. This is often implemented using advanced learning rate scheduling techniques, such as warm-up phases at the beginning of training or cyclical learning rate schedules tuned to the characteristics of deep MIMO channel estimation problems. Crucially, the Muon optimizer also integrates an adaptive regularization mechanism, which modulates dropout rates and L2 weight decay parameters in response to model behavior. For instance, if the validation loss diverges significantly from the training loss, indicating potential overfitting, the optimizer increases the regularization strength in real-time. Conversely, if the model shows signs of underfitting, the regularization is relaxed. This adaptive approach eliminates the need for manual hyperparameter tuning and ensures that model complexity remains controlled throughout the training cycle. After training on labeled channel states with known SNR and fading patterns, the model is rigorously evaluated on unseen SNR values to test its capacity for semantic generalization. The results highlight the strength of the Muon-optimized training process, demonstrating the model’s robustness, adaptability, and suitability for real-time deployment in next-generation wireless communication systems. End-to-end framework flow The overall operational flow of the proposed channel estimation framework is structured to efficiently process received pilot signals and produce accurate channel estimates, even under previously unseen conditions. The process unfolds through a sequence of interrelated stages, designed to integrate semantic reasoning, representation learning, and adaptive optimization seamlessly. The pipeline begins with the input acquisition phase, where the received pilot signal Zp is collected from the MIMO communication system. This signal, which captures the transmitted pilot symbols distorted by the wireless channel and noise, serves as the primary input to the estimation framework. In the next step, Zp is fed into the semantic-aware encoder, a lightweight neural network that transforms the received pilot sequence into a latent representation vector within a shared semantic space. This encoding process effectively condenses high-dimensional signal information into a compact, discriminative feature vector that captures key physical properties of the communication channel, such as fading characteristics and SNR patterns. Following the encoding phase, the latent vector z is matched against a database of precomputed semantic embeddings sk. Each semantic embedding sk corresponds to a known channel condition (e.g., specific fading profile or SNR level) and encapsulates abstract knowledge about that condition. A similarity metric, such as cosine similarity or Euclidean distance, is used to compare the encoded vector z with each semantic prototype sk. The channel state inference is then performed by selecting the most similar semantic embedding or by computing a weighted combination of multiple embeddings, where the weights are proportional to the similarity scores. This approach allows the model to blend information from multiple reference scenarios and infer channel characteristics with higher granularity and flexibility. Finally, the inferred semantic information is passed to the decoder, which maps the semantic embedding or the weighted vector representation back into the channel response space. This results in an accurate prediction of the channel matrix , enabling the system to perform robust channel estimation, even when exposed to channel conditions that were not seen during training. This end-to-end framework, built on the integration of Zero-Shot Learning, Transformer-based modeling, and Muon optimization, offers a scalable and intelligent solution for real-time MIMO channel estimation, capable of adapting to dynamically changing wireless environments. Transformer-ZSL-Muon optimizer framework The proposed Transformer-ZSL-Muon framework represents the full integration of Transformer-based encoding, ZSL for semantic generalization, and the Muon optimizer for advanced training stability and efficiency. This architecture is designed to leverage the strengths of each component to address the limitations of traditional channel estimation methods in complex and dynamically evolving MIMO wireless environments. At its core, the framework employs a Transformer encoder to model the intricate relationships among pilot-aided received signals, capturing both temporal and spatial dependencies through self-attention mechanisms. This attention-driven architecture facilitates the extraction of meaningful patterns across input sequences, enabling robust feature representation even under varying fading conditions and SNR levels. These extracted features are then projected into a shared semantic embedding space, where deep semantic alignment is performed with known channel attributes, following the principles of Zero-Shot Learning. A key enhancement of this framework is the replacement of conventional optimizers, such as Adam, with the more sophisticated Muon optimizer. The Muon optimizer improves convergence speed and training robustness by dynamically adjusting the learning process based on gradient curvature, historical loss behavior, and adaptive momentum decay. These dynamic updates allow the optimizer to avoid local minima and mitigate vanishing gradients, common challenges in training DL models for non-convex problems like semantic channel estimation. The training configuration is carefully designed to balance performance and generalization. Training begins with an initial learning rate of 0.002, which is adaptively modulated by Muon’s internal update policy throughout the training process. The framework is trained for 200 epochs with a batch size of 32, allowing sufficient iteration over diverse channel samples while maintaining computational efficiency. The loss function used is a hybrid formulation that combines the MSE for estimation accuracy with L2 regularization to promote weight sparsity and enhance generalization. This joint loss encourages the model to remain robust in unseen channel conditions while avoiding overfitting. By synergizing Transformer’s attention-based learning, ZSL’s semantic generalization capabilities, and Muon’s dynamic optimization, the Transformer-ZSL-Muon framework emerges as a highly adaptable, scalable, and intelligent solution for real-time channel estimation in next-generation wireless systems. The overall pipeline of the proposed Transformer-ZSL-Muon optimizer framework is visually illustrated in Fig. 2. The framework begins with dataset generation and feature-semantic extraction, followed by three estimation paths: ZSL-Muon optimizer, Transformer-ZSL, and the integrated Transformer-ZSL-Muon optimizer. These modules perform latent space projection and zero-shot learning, enabling accurate channel response prediction under unseen conditions. The prediction layer evaluates performance using MSE, accuracy, and loss metrics. As shown in Fig. 2, the end-to-end architecture begins with an extensive dataset generation stage, where 150,000 MIMO channel instances are either simulated or collected. Each instance is richly annotated with key parameters including pilot symbols, transmit/receive antenna identifiers (), SNR in dB, delay spread (ns), Doppler shift (Hz), semantic descriptors, and the complex-valued channel response (real and imaginary parts). For a systematic and rigorous evaluation, the dataset is equally partitioned into three subsets: training (50,000 samples), validation (50,000 samples) for generalization and hyperparameter tuning, and testing (50,000 samples) for final performance assessment. During the feature and semantic space extraction phase, input features such as pilot symbols, SNR, delay spread, and Doppler shift are transformed into structured numerical vectors. Simultaneously, semantic attributes are projected into a high-dimensional semantic embedding space, enabling the model to reason about unseen channel conditions through ZSL. These vectors are then processed by one of the three proposed variants: ZSL-Muon, Transformer-ZSL, or the fully integrated Transformer-ZSL-Muon framework. The Transformer-ZSL-Muon path leverages multi-head self-attention to capture complex dependencies across spatial and semantic features, producing robust latent space projections. These projections are aligned with semantic embeddings to enable generalization to unseen SNR levels and antenna configurations, achieving zero-shot inference. The latent vectors are then passed to a Muon-optimized ZSL decoder, which performs inference and estimates the channel response. The final prediction layer computes evaluation metrics including MSE, accuracy, and loss, providing insights into the framework’s performance under diverse and realistic wireless conditions. By integrating semantic reasoning with deep attention models and adaptive optimization, the proposed framework offers a scalable, accurate, and generalizable solution for real-time channel estimation in advanced MIMO systems. ZSL integrated with Muon optimizer In conventional ZSL frameworks, a basic encoder, typically non-transformer-based, is used to project input features into a semantic space. In this work, the integration of the Muon optimizer enhances this structure by adaptively adjusting learning rates based on gradient magnitudes and loss dynamics. This integration significantly improves convergence behavior and generalization, especially in non-stationary MIMO environments. The semantic mappings, either manually defined or learned, enable the model to interpret channel states across unseen conditions using abstract embeddings. To assess the performance of this hybrid ZSL-Muon framework in a MIMO wireless communication environment, simulations are conducted using realistic Rician channel models that capture both line-of-sight and multipath components. The system further accommodates temporal dynamics by simulating both time-varying and quasi-static block fading conditions. The Muon optimizer begins with a learning rate of 0.002, adaptively refined throughout 200 training epochs with a batch size of 32 to ensure convergence and stability. Table 3 summarizes the simulation configuration, detailing the dataset size, input features, network structure, and loss formulation employed for training and evaluation. The hyperparameters in both tables are empirically selected based on standard practices in DL for MIMO channel estimation tasks, drawing from established literature. To guarantee quick convergence, consistent training, and optimal performance (minimizing MSE while avoiding overfitting), they are verified through initial experiments on our dataset. In particular: A batch size of 32 is utilized to strike a mix between computational efficiency and robust gradient estimations; this is a typical starting point for Transformer-based models in wireless communications. It is frequently used to deliver dependable updates while preventing memory overflow on common GPUs. A batch size of 32 is utilized to strike a mix between computational efficiency and robust gradient estimations; this is a typical starting point for Transformer-based models in wireless communications. It is frequently used to deliver dependable updates while preventing memory overflow on common GPUs. Epochs 200 is applied to give the size of our dataset (~ 150,000 samples) enough iterations for convergence without requiring excessive processing. During training, early stopping is used to stop at the ideal point, which is usually between 150 and 180 epochs. Epochs 200 is applied to give the size of our dataset (~ 150,000 samples) enough iterations for convergence without requiring excessive processing. During training, early stopping is used to stop at the ideal point, which is usually between 150 and 180 epochs. The learning rates (0.002 for Muon in Table 3 and 0.003 for Adam in Table 6) are selected to allow for quick initial progress without divergence, with Muon’s adaptability permitting a slightly higher rate than Adam for quicker convergence. The learning rates (0.002 for Muon in Table 3 and 0.003 for Adam in Table 6) are selected to allow for quick initial progress without divergence, with Muon’s adaptability permitting a slightly higher rate than Adam for quicker convergence. Embedding dimensions (256 input, 128 semantic) are correspond to common latent spaces in ZSL for MIMO and offer concise, expressive representations with few parameters. Embedding dimensions (256 input, 128 semantic) are correspond to common latent spaces in ZSL for MIMO and offer concise, expressive representations with few parameters. Additional options (such as hidden size 512 and dropout 0.3) is selected to reduce overfitting and improve generalization, use standard values from Transformer literature. Additional options (such as hidden size 512 and dropout 0.3) is selected to reduce overfitting and improve generalization, use standard values from Transformer literature. The hyperparameter values are completely consistent with accepted methods in DL-based MIMO channel estimation and ZSL. They are chosen via a methodical grid search on the validation set. The loss and accuracy curves in Figs. 6, 8 and 8 show that training for 200 epochs with early stopping (patience = 20) ensures convergence. From the set {0.0005, 0.001, 0.002, 0.003, 0.005}, initial learning rates of 0.002 (Muon) and 0.003 (Adam) are shown to be optimal for the fastest convergence and lowest validation MSE. Successful ZSL designs are followed by input and semantic embedding dimensions of 256 and 128, respectively24,25. To maintain the integrity of the reported zero-shot test results, all decisions are made solely based on training and validation data. The accuracy of the proposed model. Comparison of training loss curves using Muon and adam optimizers over 200 epochs. The Muon optimizer achieves faster convergence, smoother trajectory, and lower final loss, demonstrating improved stability and effectiveness for transformer-ZSL training in MIMO channel estimation tasks. To further support the ZSL model’s ability to generalize to previously unseen channel conditions, the input feature dimension is configured to 256, while the semantic embedding space is set to 128 dimensions. A Transformer encoder architecture equipped with cross-attention mechanisms is adopted, enabling precise alignment between the input space and the semantic representations. The loss function integrates MSE with L2 regularization to promote robust generalization and mitigate overfitting. Training is accelerated using an NVIDIA RTX 3060 GPU, providing the necessary computational resources for efficient DL-based channel modeling. The entire framework is implemented using PyTorch, ensuring scalability, reproducibility, and adaptability. This experimental setup delivers a high-performance and generalizable approach to MIMO channel estimation within the ZSL paradigm. Transformer-ZSL framework The Transformer-ZSL framework integrates a transformer encoder with multi-head self-attention to effectively capture both semantic attributes and contextual dependencies among input features. This configuration enables the model to learn robust and generalizable representations of the wireless channel under diverse and dynamic propagation conditions. Semantic alignment between the visual domain (e.g., pilot sequences or signal representations) and the semantic domain (e.g., channel characteristics like SNR, fading type, and mobility) is achieved using attention-driven mechanisms, where the Adam optimizer facilitates stable training convergence and effective gradient updates. In this study, a Transformer-ZSL model is developed to estimate complex wireless channel responses across a wide variety of propagation environments. The simulation scenario is based on a Rician fading channel model, which accurately reflects both Line-of-Sight (LoS) and multipath propagation components. The channel behavior is modeled under two principal conditions: quasi-static block fading and time-varying fading, thereby mimicking real-world wireless communication environments with varying mobility and interference. The dataset used to train and evaluate the Transformer-ZSL model comprises input feature vectors encoded as 256-dimensional representations, typically derived from pilot sequences or transformed channel state indicators. These are paired with corresponding semantic attribute vectors mapped into a 128-dimensional latent space. This configuration enables the model to perform zero-shot inference, wherein it can estimate channel conditions that were not explicitly encountered during training. Such a capability significantly enhances the adaptability and scalability of the proposed framework, making it well-suited for deployment in next-generation wireless systems that must operate reliably across diverse and unpredictable environments. The proposed Transformer-ZSL model architecture is composed of four Transformer encoder layers, each equipped with eight multi-head self-attention mechanisms designed to capture diverse contextual relationships embedded in the input features. To preserve sequential dependencies within the input signal streams, sinusoidal positional encoding is incorporated into the input embeddings. Non-linear activation is achieved using the GELU function within each encoder block, while feedforward layers are dimensioned to 512 units to ensure expressive representation learning. A dropout rate of 0.3 is employed throughout the model to mitigate overfitting, enhancing the model’s robustness and generalization capacity. The key simulation parameters and model configurations used to train the Transformer-ZSL framework are summarized in Table 6. The training process is conducted over 200 epochs using the Adam optimizer, with a learning rate of 0.003 and a batch size of 32. The objective function includes both the MSE for precise channel estimation and an L2 regularization term to promote weight sparsity and prevent overfitting. The input data is represented by 256-dimensional vectors, while semantic attributes are projected into a 128-dimensional latent space, facilitating the model’s zero-shot learning ability. Thanks to this carefully designed architecture and training regimen, the Transformer-ZSL model demonstrates a strong capability to generalize across previously unseen SNR levels and semantic attribute combinations. It achieves reliable prediction of complex channel responses, comprising both real and imaginary components, even under zero-shot settings. Simulation results and discussion To validate the performance of the proposed hybrid channel estimation framework, comprehensive simulations were conducted under diverse wireless scenarios modeled after widely accepted communication standards. The simulation setup incorporates pilot signals of varying lengths, specifically 8, 16, 32, and 64 symbols, to assess the robustness of the model under different training signal densities. The overall signal length is fixed at 1024, and the 128-QAM modulation scheme is adopted to simulate high-throughput MIMO environments. This section presents and analyzes the resulting performance metrics, comparing the proposed Transformer-ZSL-Muon model with alternative configurations and conventional baselines across different SNR regimes and fading scenarios. As illustrated in Fig. 3, the performance of the proposed hybrid DL channel estimation model is evaluated under varying numbers of transmitting antennas {2, 4, 8, 16, 32, 64, 128} across two fading environments: quasi-static block fading and time-varying fading. The model demonstrates robust estimation accuracy across different SNR levels, with error trends influenced by increasing antenna correlation, reduced signal power per antenna, and multipath interference effects. The results demonstrate that the model consistently maintains low MSE levels across increasing SNR values, indicating effective generalization across diverse MIMO configurations. Performance of channel estimation using the proposed hybrid DL model across varying numbers of transmitting antennas: (a) Quasi-static block fading scenario; (b) Time-varying fading scenario. A clear trend is observed wherein the channel estimation error increases with the number of transmitting antennas. This degradation is attributed to several factors: (1) increased antenna correlation that leads to reduced spatial diversity and higher interference levels; (2) power distribution across a larger number of transmit antennas, which lowers the per-antenna signal power and increases vulnerability to noise; and (3) heightened multipath interference due to complex interactions between large-scale antenna arrays and the propagation environment. These effects are particularly pronounced in low SNR regimes, where the signal-to-noise distinction becomes minimal, resulting in higher estimation uncertainty. Notably, even in such challenging conditions, the proposed Transformer-ZSL-Muon framework exhibits strong resilience, highlighting its suitability for real-time MIMO systems operating under diverse antenna configurations and fading dynamics. As shown in Fig. 3a, the proposed hybrid framework achieves a substantial 73.3% improvement in MSE at an SNR of 25 dB compared to the baseline method reported in32, highlighting the effectiveness of the Transformer-ZSL-Muon model in reducing estimation error and enhancing channel prediction accuracy under quasi-static fading conditions. In Fig. 3b, the impact of increasing the number of transmitting antennas on MSE performance under time-varying fading conditions is examined across a range of SNR values. It is evident that configurations with eight or more antennas begin to reach the 10− 5 MSE threshold at higher SNR levels (above 20 dB), where the estimation gains become significantly more pronounced, particularly beyond 30 dB. At an SNR of 30 dB, the performance gain attributed to antenna scaling is quantified as approximately 35%, 52%, 55%, and 58% for systems with 8, 16, 64, and 128 antennas, respectively, compared to the 2-antenna baseline. These improvements underscore the advantage of increased spatial diversity in reducing estimation uncertainty. However, a diminishing return is observed beyond 16 antennas, suggesting a saturation region where further increases in antenna count introduce computational complexity without proportionally enhancing MSE performance. Additionally, compared to Ref26., the proposed framework achieves a consistent 15.7% improvement in MSE across the high-SNR regime under time-varying conditions, further affirming its robustness and generalization capability. As illustrated in Fig. 4, the performance of the proposed hybrid DL framework is evaluated under different pilot signal configurations, specifically, 8, 16, 32, and 64 pilots, in both quasi-static block fading and time-varying fading environments. The results clearly indicate that increasing the number of pilot signals leads to consistent improvements in MSE performance across all SNR levels. Performance of channel estimation with varying numbers of pilots: (a) Quasi-static block fading; (b) Time-varying fading. This improvement can be attributed to the greater amount of known reference information available for training the model when more pilot signals are used. In a MIMO setup, where the channel matrix is inherently high-dimensional due to multiple transmitting and receiving antennas, having additional pilot sequences provides a more comprehensive view of the channel conditions. This increases the degrees of freedom in the channel estimation process, allowing the model to more accurately approximate the complex, non-linear relationships within the channel. Moreover, the additional pilot information helps the model to become more robust against noise and interference, particularly in low-SNR scenarios. By providing more observation points, the model can better generalize to varying channel states and attenuate the impact of fading fluctuations. Notably, in Fig. 4a, the MSE steadily drops as the number of pilots increases, particularly at moderate-to-high SNR levels, with 64-pilot configurations yielding the best performance. Similarly, Fig. 4b demonstrates that this trend holds even under time-varying fading conditions, where the increased pilot count helps stabilize the estimation against rapid channel variations. While the gain from 32 to 64 pilots begins to taper slightly, suggesting diminishing returns beyond a certain threshold, the overall enhancement confirms that pilot signal augmentation significantly boosts estimation accuracy and model generalization in dynamic MIMO environments. The performance trends illustrated in Fig. 4a reveal that increasing the number of pilot symbols substantially enhances channel estimation accuracy, particularly in low-SNR conditions. Notably, the MSE improvement saturates at higher SNR values; however, at lower SNR levels, extending the pilot duration has a pronounced impact on reducing estimation error. For instance, using 64 pilot symbols allows the system to achieve the target MSE of 10− 5 at approximately 6–7 dB lower SNR compared to configurations with only 8 pilots. This emphasizes the critical importance of pilot optimization in low-SNR environments where signal reliability is most vulnerable. A comparative analysis among the 8, 16, 32, and 64 pilot configurations demonstrates a consistent reduction in the SNR required to meet the MSE threshold of 10− 5, indicating that increasing the number of pilots effectively improves estimation performance while reducing power requirements. Specifically, the proposed framework achieves an MSE of approximately 1.2 × 10− 5 at 20 dB SNR with 64 pilot symbols, whereas the baseline method from26 reports an MSE of 2 × 10− 5 under the same conditions. This translates to a performance gain of around 40%, validating the superiority of the proposed method in extracting more reliable channel estimates from limited or noisy observations. Turning to Fig. 4b, which depicts performance under time-varying fading conditions, similar trends are observed. The inclusion of more pilot symbols consistently results in lower MSE across the entire SNR range. This is attributed to the richer training signal provided by the additional pilots, which enables the model to better capture the temporal dynamics and underlying structure of the channel. The 64-pilot configuration again stands out, particularly at moderate SNR levels (20–30 dB), where it consistently achieves the lowest MSE values. At the critical MSE threshold of 10− 5, the required SNR is reduced from 25 dB (for 8 pilots) to approximately 20 dB for configurations with 16, 32, and 64 pilots, corresponding to a 5 dB gain in SNR efficiency. This improvement equates to a 20% enhancement in estimation performance, demonstrating the capability of the proposed hybrid DL framework to extract accurate channel estimates without increasing the transmit power. Furthermore, when benchmarked against Ref26., the proposed model with 64 pilots achieves a 50% reduction in MSE, underscoring its effectiveness in maintaining estimation robustness even under time-varying conditions. To further validate the superiority of the proposed framework, we conducted a comparative analysis between different channel estimation strategies, including DL-based models, Transformer-ZSL-Muon, Transformer-ZSL, and ZSL-Muon optimizer, and classical estimation techniques such as MMSE and LS. The results, as depicted in Fig. 5, clearly demonstrate the enhanced performance of the hybrid Transformer-ZSL-Muon model under both quasi-static block fading (Fig. 5a) and variant time fading (Fig. 5b) conditions. The models compared include Transformer-ZSL-Muon, Transformer-ZSL, ZSL-Muon optimizer, and classical estimators (MMSE and LS). Comparison of channel estimation performance across various SNR levels using different models and estimation techniques: (a) Quasi-static block fading scenario, and (b) Time-varying fading scenario. Across both channel scenarios, the Transformer-ZSL-Muon consistently achieves the lowest MSE across all evaluated SNR levels. This highlights the model’s ability to robustly capture the intricate, nonlinear dependencies between pilot inputs and channel responses, dependencies that traditional linear estimators like MMSE and LS fail to model effectively, particularly under complex and dynamic propagation conditions. In Fig. 5a, the Transformer-ZSL-Muon demonstrates a clear advantage over its counterparts, especially in the mid to high SNR range (25–40 dB), where it outperforms the next best method (Transformer-ZSL) by approximately 22% to 30% in terms of MSE. The benefits of integrating the Muon optimizer and Transformer architecture become especially evident here, as they jointly enhance learning convergence and semantic alignment during training. This translates to more accurate and stable channel estimates. Similarly, under variant time fading as shown in Fig. 5b, the proposed model maintains its superior estimation performance. While all methods experience an MSE reduction with increasing SNR, the Transformer-ZSL-Muon maintains a performance margin of 20–35% over conventional techniques and achieves convergence to MSE levels near 10− 6 at 35–40 dB SNR. The LS and MMSE estimators, in contrast, plateau at significantly higher MSE levels, indicating their limitations in handling the dynamic nature of time-varying fading environments. Overall, these results reinforce that the hybrid DL framework significantly outperforms both classical and partial DL-based models in channel estimation tasks. The synergy of zero-shot learning, attention-driven Transformer encoding, and adaptive Muon optimization yields a robust architecture that not only generalizes well to unseen channel conditions but also ensures low estimation error across a wide SNR spectrum. DL models demonstrate superior capacity in capturing complex and nonlinear channel mapping relationships, which significantly enhances the accuracy of channel estimation in MIMO systems. These models benefit from high model expressiveness and generalization capabilities, allowing them to adapt effectively to varying wireless channel conditions. Unlike traditional estimation methods that rely on linear approximations and manual feature engineering, DL approaches enable end-to-end learning, reducing the need for handcrafted signal processing pipelines. This not only lowers overall system complexity but also streamlines channel estimation into a unified and efficient framework. Among the evaluated models, the proposed Transformer-ZSL-Muon framework exhibits a notably stronger ability to learn intricate channel characteristics. This hybrid model outperforms both its individual components, Transformer-ZSL and ZSL-Muon optimizer, by leveraging the synergistic strengths of attention-based feature extraction, semantic alignment through zero-shot learning, and adaptive optimization using the Muon algorithm. Particularly under complex and dynamic propagation scenarios, the Transformer-ZSL-Muon model maintains lower MSE levels across a range of SNRs, demonstrating its robustness and adaptability. These results underscore the model’s enhanced capability to generalize under unseen conditions and to effectively estimate channel states with minimal prior knowledge. The effectiveness of the proposed Transformer-ZSL-Muon hybrid model is further quantified in Fig. 5, where its performance is compared against a variety of benchmarks, including Transformer-ZSL, ZSL-Muon optimizer, and conventional estimation methods such as MMSE and LS. Specifically, the Transformer-ZSL-Muon model achieves a significant MSE reduction of 27% compared to Transformer-ZSL, 24% compared to the ZSL-Muon optimizer, 93.75% over MMSE, and 96.67% compared to LS at an SNR of 30 dB. These gains highlight the value of combining attention mechanisms, zero-shot semantic mapping, and the adaptive convergence properties of the Muon optimizer into a unified learning framework. As illustrated in Fig. 5b, the superiority of the hybrid model is especially apparent at higher SNR levels, where the MSE curves of traditional methods tend to flatten, indicating diminishing improvements. In contrast, the proposed framework continues to decrease MSE significantly, reinforcing its robust generalization and adaptability across both quasi-static and time-variant fading scenarios. Further validation of the hybrid model’s learning dynamics is presented in Figs. 6 and 7, which display the accuracy and loss trends across 200 training epochs, respectively. In Fig. 6, the Transformer-ZSL-Muon model achieves rapid convergence, surpassing 90% accuracy within the first 75 epochs and maintaining this performance with minimal fluctuation throughout the remaining epochs. This stands in contrast to the Transformer-ZSL and ZSL-Muon optimizer models, which plateau at lower accuracy levels (around 85% and 80%, respectively), reflecting their relatively limited representational power or optimization efficiency. The loss of the proposed model. Complementarily, Fig. 7 demonstrates that the hybrid model exhibits significantly lower training loss, achieving a final value below 0.4, whereas the other two models converge at higher loss values (~ 0.6 and above). This implies that the Transformer-ZSL-Muon architecture not only learns faster but also retains a more optimal parameter configuration, resulting in both higher precision and better generalization. In summary, the combination of high representational capacity (via Transformer), semantic transferability (via ZSL), and convergence efficiency (via Muon optimizer) enables the proposed framework to achieve superior training dynamics, lower estimation error, and better robustness across a range of challenging channel environments. These findings confirm the hybrid model’s potential for real-time and reliable deployment in future MIMO systems. The advantages of integrating the Muon optimizer into the Transformer-ZSL architecture are clearly illustrated in Fig. 8, which compares the training loss profiles of the Muon and Adam optimizers over 200 epochs. The Muon optimizer demonstrates superior convergence characteristics, exhibiting a steeper and more consistent decline in loss, ultimately achieving a lower final loss compared to Adam. This can be attributed to Muon’s dynamic learning strategy, which adaptively adjusts the learning rate based on curvature information, historical gradients, and loss behavior. Such adaptive mechanisms enhance its capability to navigate the complex optimization landscape typically encountered in high-dimensional MIMO channel estimation tasks. In contrast, the Adam optimizer, although effective in many scenarios, relies on fixed decay and momentum coefficients that may not sufficiently adapt to the evolving training dynamics of deep semantic models like ZSL. This results in slower convergence and a higher residual loss floor. Moreover, the smoother and less oscillatory trajectory of the Muon optimizer highlights its improved stability during training, which is particularly beneficial for tasks requiring fine-grained alignment between semantic and channel domains. These findings reinforce the suitability of Muon as a robust and efficient optimization algorithm for zero-shot learning frameworks in wireless communication systems. According to Table 6 results, adding the Muon optimizer alone resulted in the quickest training time (about 7.2–7.4 h), demonstrating its effectiveness in speeding up convergence. The basic Transformer–ZSL takes a little longer (about 7.5–7.7 h), while the added representation learning burden of Muon-based feature encoding results in a somewhat longer training period (approximately 7.9–8.1 h). Overall, the time differences are negligible, demonstrating that the increased accuracy offered by Muon components has a negligible computational cost. The extensive simulation results presented in this section confirm the superior performance, adaptability, and scalability of the proposed Transformer-ZSL-Muon framework across a broad range of MIMO wireless channel conditions. By integrating the semantic generalization power of zero-shot learning, the contextual feature extraction of Transformer encoders, and the convergence stability of the Muon optimizer, the hybrid model consistently outperforms both traditional estimation techniques and recent DL baselines. It achieves significant improvements in MSE, particularly under high-SNR and time-varying conditions. The model demonstrates strong generalization to unseen scenarios, efficient convergence with low final error rates, and improved SNR efficiency with increased pilot density and antenna configurations. While the Transformer-ZSL-Muon model incurs a modest increase in computational time, the trade-off is justified by its superior estimation accuracy and robust learning behavior. These findings collectively position the proposed framework as a scalable, intelligent, and high-performance solution for real-time channel estimation in next-generation MIMO communication systems. The thorough comparison in Table 7 makes it abundantly evident that the suggested Transformer–ZSL–Muon architecture is superior to both DL baselines and traditional model-based estimators. The suggested model outperforms MMSE by 68.8% and 72.8%, and CNN-based approaches by even more significant gains of 47.5% and 51.0%. These findings verify that the synergistic combination of Muon’s adaptive optimization, ZSL zero-shot generalization, and Transformer global attention yields state-of-the-art performance throughout the whole SNR spectrum, with especially large margins in difficult low-SNR and dynamic fading scenarios. The robustness and scalability of the suggested hybrid solution for next-generation MIMO systems are validated by the consistent ranking under all scenarios. Table 7Normalized MSE (dB) comparison at several SNR values with classical DL baselines.MethodQuasi-static fadingTime-varying fading10 dB 20 dB 30 dB10 dB 20 dB 30 dBLS−2.2 −11.9 −21.7−1.9 −10.9 −20.6MMSE−8.6 −18.9 −28.9−8.9 −18.7 −27.4CNN (DeepCNN-CE [6])−13.5 −24.3 −33.8−12.8 −22.9 −31.4GRU−14.9 −26.6 −34.7−13.9 −24.8 −33.2LSTM (BiLSTM-CE [7])−15.3 −27.3 −35.4−14.5 −25.8 −34.2Transformer–ZSL−18.9 −32.8 −41.9−17.8 −30.3 −40.6ZSL–Muon−19.8 −33.8 −43.4−18.6 −32.5 −41.8Proposed transformer–ZSL–Muon −23.9 −39.8 −33.4 −22.9 −38.3 −32.2 Normalized MSE (dB) comparison at several SNR values with classical DL baselines. Robustness to reduced pilot overhead The results in Table 8 clearly demonstrate the superiority of the suggested Transformer–ZSL–Muon model, with a stable performance hierarchy across all pilot-ratio regimes. When pilot density drops, classical estimators like LS and MMSE deteriorate quickly; at extreme sparsity (6.25%), LS even enters highly unreliable positive-dB error regions. Even if DL models like CNN, GRU, and LSTM perform noticeably better than conventional methods, they nonetheless noticeably deteriorate as pilots get fewer. Table 8Comparison of normalized MSE (dB) performance at fixed SNR = 20 dB with different pilot overhead ratios.Pilot ratio100%50%25%12.5%6.25% (extremely sparse)LS−12.8 dB−9.3−5.2+ 3.5+ 9.8 dBMMSE−17.7 dB−15.2−9.8−4.4+ 4.5 dBCNN−25.4 dB−19.9−14.3−8.3−0.5 dBGRU−25.7 dB−21.6−16.5−9.8−2.4 dBLSTM−28.6 dB−21.8−17.3−11.2−3.2 dBTransformer–ZSL−33.7 dB−28.9−24.1−19.4−9.8 dBZSL–Muon−34.7 dB−29.9−26.2−21.2−10.4 dBProposed transformer–ZSL–Muon −39.8 dB −35.6 −31.2 −26.8 −11.2 dB Comparison of normalized MSE (dB) performance at fixed SNR = 20 dB with different pilot overhead ratios. The advantage of attention-based global feature extraction under low-pilot settings is confirmed by the baseline Transformer–ZSL, which offers a significant improvement over recurrent and convolutional models. This performance is further improved by adding Muon feature encoding: ZSL–Muon reliably lowers NMSE by an extra 0.5–1.5 dB across the majority of pilot ratios. In every case, the suggested Transformer–ZSL–Muon outperforms all other models by a significant margin, particularly under extreme pilot reduction. It retains − 11.2 dB at the incredibly sparse 6.25% pilot ratio, outperforming conventional approaches by more than 20 dB and surpassing the next-best model (ZSL–Muon) by about 1 dB. The model’s excellent generalization, improved feature disentanglement, and resistance to sparse observations are all demonstrated by this strong performance under rigorous pilot compression. Table 9 demonstrates that the suggested Transformer–ZSL–Muon model maintains a competitive inference cost (3.1 GFLOPs) and a suitable parameter size (6.7 M) while achieving the most efficient training (16.8 GFLOPs vs. 65 GFLOPs for Transformer–ZSL) thanks to Muon-based feature reduction. This trade-off is acceptable for high-performance base station environments, despite its much greater memory footprint (~ 385 MB). The model’s only drawback is memory utilization, but overall it provides the best balance between accuracy and computational efficiency. Table 9Theoretical computational complexity comparison.MethodParametersTraining FLOPs per batchInference FLOPs per sampleMemory footprintLS/MMSE–O ()O ()O ()CNN (DeepCNN-CE)1.6 M~ 15 GFLOPs~ 0.9 GFLOPs~ 8 MBGRU (128 hidden units)2.6 M~ 27 GFLOPs~ 1.5 GFLOPs~ 10 MBLSTM (BiLSTM-CE)3.8 M~ 41 GFLOPs~ 2.2 GFLOPs~ 16 MBTransformer–ZSL (4 layers, 8 heads)5.7 M~ 65 GFLOPs~ 3.6 GFLOPs~ 23 MBZSL–Muon (lightweight)2.8 M~ 34 GFLOPs~ 1.9 GFLOPs~ 12 MBProposed Transformer–ZSL–Muon6.7 M~ 16.8 GFLOPs~ 3.1 GFLOPs~ 385 MB Theoretical computational complexity comparison. Conclusion, limitations, and future works This paper introduced a robust and intelligent hybrid DL framework for MIMO channel estimation that integrates Transformer architectures, ZSL mechanisms, and the Muon optimizer. Designed to address the inherent limitations of conventional DL and model-based estimators, the proposed Transformer-ZSL-Muon framework demonstrates remarkable accuracy, convergence speed, and generalization capacity, particularly in dynamic and unseen wireless channel conditions. By leveraging semantic embeddings and attention-driven architectures, the model effectively captures spatial-temporal dependencies and generalizes across varying SNRs, fading types, and antenna configurations without retraining. Furthermore, its enhancements over ZSL-Muon and Transformer-ZSL variants underscore the complementary strengths of all three components. A particularly noteworthy finding is the impact of massive MIMO configurations. The proposed model delivers state-of-the-art performance in both quasi-static and time-varying block fading environments, consistently outperforming conventional and recent DL-based baselines. Despite its demonstrated strengths, several limitations merit consideration: While effective, the Transformer-ZSL-Muon framework involves higher training and inference costs due to the Transformer’s multi-head attention layers and the adaptive nature of the Muon optimizer. This makes real-time deployment on resource-constrained platforms, such as edge devices, IoT nodes, or low-latency communication systems, more challenging. While effective, the Transformer-ZSL-Muon framework involves higher training and inference costs due to the Transformer’s multi-head attention layers and the adaptive nature of the Muon optimizer. This makes real-time deployment on resource-constrained platforms, such as edge devices, IoT nodes, or low-latency communication systems, more challenging. The model requires substantial memory and computational resources for processing large-scale antenna arrays and high-dimensional semantic embeddings, which could hinder scalability in extremely large MIMO settings or distributed systems. The model requires substantial memory and computational resources for processing large-scale antenna arrays and high-dimensional semantic embeddings, which could hinder scalability in extremely large MIMO settings or distributed systems. The semantic embedding space, whether learned or manually crafted, plays a central role in generalization. However, generating reliable semantic representations for highly complex or non-stationary channels remains a non-trivial task and could affect the performance in practical scenarios. The semantic embedding space, whether learned or manually crafted, plays a central role in generalization. However, generating reliable semantic representations for highly complex or non-stationary channels remains a non-trivial task and could affect the performance in practical scenarios. The proposed model is trained offline with pre-defined datasets. In highly dynamic environments where channel statistics evolve rapidly, the absence of online adaptation may limit real-time responsiveness. The proposed model is trained offline with pre-defined datasets. In highly dynamic environments where channel statistics evolve rapidly, the absence of online adaptation may limit real-time responsiveness. Future work will focus on addressing these limitations and extending the model’s applicability: To facilitate deployment in edge computing environments, future research will explore lightweight Transformer variants such as MobileViT, Tiny Transformer, or sparse attention architectures, which aim to reduce inference time, memory usage, and power consumption without significantly compromising estimation accuracy. To facilitate deployment in edge computing environments, future research will explore lightweight Transformer variants such as MobileViT, Tiny Transformer, or sparse attention architectures, which aim to reduce inference time, memory usage, and power consumption without significantly compromising estimation accuracy. Incorporating online learning mechanisms or continual learning strategies can enable the framework to adapt to rapidly changing wireless conditions without the need for retraining from scratch. Incorporating online learning mechanisms or continual learning strategies can enable the framework to adapt to rapidly changing wireless conditions without the need for retraining from scratch. Adapting the model to support FPGA, ASIC, or low-power GPU implementations can further improve its viability for practical deployment in 5G/6G systems and vehicular communications. Adapting the model to support FPGA, ASIC, or low-power GPU implementations can further improve its viability for practical deployment in 5G/6G systems and vehicular communications. Future work may also explore more sophisticated or data-driven methods for semantic embedding generation, particularly in heterogeneous or adversarial channel environments. Future work may also explore more sophisticated or data-driven methods for semantic embedding generation, particularly in heterogeneous or adversarial channel environments. Expanding the model to jointly optimize channel estimation with downlink prediction, beamforming, or resource allocation could create a unified learning-based physical layer design for next-generation wireless networks. Expanding the model to jointly optimize channel estimation with downlink prediction, beamforming, or resource allocation could create a unified learning-based physical layer design for next-generation wireless networks. By addressing these areas, the proposed Transformer-ZSL-Muon framework can evolve into a truly scalable, adaptive, and energy-efficient solution for real-time, robust channel estimation in modern and future MIMO wireless communication systems. Acknowledgements The authors would like to acknowledge Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2026R197), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia. The authors specially acknowledge the Automated Systems and Computing Lab (ASCL) at Prince Sultan University, Riyadh, Saudi Arabia. In addition, the authors wish to thank Prince Sultan University, Riyadh, Saudi Arabia, for their support. Author contributions Wessam M. Salama was responsible for the conceptualization of the study, developing the methodology, and preparing the original manuscript draft. Moustafa H. Aly significantly contributed to methodology refinement, provided technical supervision, and critically reviewed and revised the manuscript. Samah Alshathri was involved in dataset curation, validation of experimental results, and manuscript review. Walid El-Shafai managed project administration, data analysis, validation, supervised the research process, performed the final editing, and acted as the corresponding author. Funding This work is supported by Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2026R197), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia. Data availability The datasets used and/or analysed during the current study available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.