Edge AI enabled MIMO MC-CDMA for 6G optimizing spectrum and energy efficiency with SIC and deep reinforcement learning Authors: A. Vijay Published: 2026 ABSTRACT The rapid expansion of 6G wireless networks requires high SE techniques to improve spectral efficiency (SE), energy efficiency (EE), and interference. In this paper, an Edge AI MIMO MC-CDMA system with SIC and DRL is proposed for spectrum and energy efficiency. As for the difference from the traditional MIMO-OFDM and hybrid precoding, this paper is based on DRL for adaptive learning, which will show better performance for dense network. The proposed system is implemented using MATLAB to allow for real- time decision- making pertaining the streams’ networking parameters at the network edge. DRL makes it possible to ensure interference cancellation, while Edge AI avoids increased computational load and guarantees fast responses. Simulation outcome shows that the proposed model has a spectral efficiency of about 32.7 bits/s/Hz and energy efficiency of 14.8 bits/Joule and thereby surpass than MC-CDMA, MIMO-OFDM and the hybrid precoding-based MIMO systems. Furthermore, the SINR is enhanced to be 34 dB while BER is decreased to 10−5 to guarantee high transmission reliability. The deep learning-based mechanism effectively interferes the power allocation and enhances the stability of networking. The results fully support significantly optimizes power allocation, improving overall network stability and efficiency for the future 6G wireless network is highly scalable, capable of operating with minimum interference issues and very much energy efficient. Due to the fact that the proposed model can dynamically adjust to the conditions in the network, it conceptualises one approach to a smart future of wireless communication in the next generation applied to ultra-high densities of networks and users. Introduction The rapid evolution of wireless communication has resulted in next-generation networks with 6G following 5G1. To address increasing requirements for high data rates, ultra-low latency, and energy efficiency in communications, researchers are working on frontier technologies to enhance spectrum efficiency and power consumption2. Multiple-Input Multiple-Output (MIMO) with Multi-Carrier Code Division Multiple Access (MC-CDMA) is one of the most important enablers of 6G, as it enhances spectral efficiency by enabling several users to communicate in parallel on various frequency subcarriers3. But with increasing network density and complexity, interference, wastage of energy, and dynamic spectrum allocation are important issues. Integration of Edge AI—facilitating smart, real-time decision-making at the network edge can potentially revolutionize wireless communication by reducing latency and making optimal use of network resources4. To unlock the full potential of these innovations, new approaches need to be formulated in order to overcome interference, promote energy efficiency, and dynamically allocate the spectrum based on network conditions. In previous studies, the technology used in the 6G wireless communication mainly focused on MIMO and MC-CDMA as the means to improve capacity and reliability. Thus, interference mitigation techniques as Successive Interference Cancellation (SIC) have been proposed to enhance a signal detection and to decrease the bit error rates (BER)5. Both machine learning and optimization have been used in the improvement of the efficiency of spectrum and power6. However, such methods are based on heuristics and fail to address various and dynamically changing network systems. Furthermore, deep learning solutions are need for resource optimization; however, they are centralized causing high latency and high-power consumption7. One of the major challenging factors is the absence of effective deep reinforcement learning (DRL) frameworks that can learn resource management strategies for 6G networks in real-time8. Moreover, the employed SIC techniques although have reasonable performance have reliability problems in terms of computation complexity and convergence in large scale MIMO MC-CDMA systems and useful in ultra-dense networks. In order to deal with these challenges, this study suggested to employ an Edge AI for 6G MIMO MC-CDMA system along with Successive Interference Cancellation (SIC) and Deep Reinforcement Learning (DRL) to optimize spectrum and energy efficiency. Incorporating Edge AI will allow the system to make decisions without requiring the use of the cloud and increase the pace of decision-making by reducing the latency time9. The SIC algorithm is the incorporation of artificial intelligence interference cancellation which will augment the ability to detect other user’s signal and rule out interference from the same. Furthermore, DRL will be applied for dynamic spectrum and power control, since the network has an interaction with the environment in order to learn the best policy on how to use the spectrum and power efficiently. It is believed that the proposed method shall provide better spectral efficiency, consume lesser power and have better robust system than the existing techniques. Unlike the previous literature that used DRA with either MIMO or SIC alone, the presented scheme uses a combination of DRA, Edge AI, MIMO MC-CDMA, and SIC to optimize the system in real-time. It is the only mechanism that can do decentralized decision-making at the network edge and also enhance interference management, distribution of power, spectral efficiency, as well as energy efficiency in dense 6G wireless settings. Research motivation The rapid evolution of wireless communications to 6G networks demands greater spectrum efficiency, lower latency, and improved energy efficiency. Traditional MIMO and MC-CDMA systems support strong multi-user communication but suffer from interference, resource allocation, and scalability. To counter such vulnerabilities, integrating Edge AI with Deep Q-Network (DQN)-based Successive Interference Cancellation (SIC) can optimize power allocation, interference cancellation, and adaptive beamforming. This study will design an intelligent, adaptive MIMO MC-CDMA system with AI-based optimization for better spectrum utilization and energy efficiency in next-generation 6G networks. Problem statement The increased demand for high-speed, low-latency, and energy-efficient wireless communication has resulted in extensive research on advanced multiple access and signal processing techniques for next generation networks. All of these proposed computer-aided optimization techniques-from deep learning-based channel coding10, hybrid cognitive radio networks11, AI-powered resource allocation12 still have high computational overhead and scalability constraints. Resource allocation schemes for MIMO systems based on intelligent surface technologies13 and interference-aware massive access schemes are good cases that enhance spectral efficiency, but they need further fine-tuning for dynamic scenarios. For this purpose, the combination of Edge AI and MIMO-MC-CDMA exploiting deep reinforcement learning (DRL) and SIC could serve as a panacea for optimizing spectrum and energy efficiency while ensuring lower computational complexity, thus being absolutely a candidate for 6G networks14. Recent innovation and its limits Advances in 6G wireless networks have recently suggested the use of Massive MIMO, MC-CDMA, and AI-based resource allocation for enhancing communication efficiency. Techniques like Deep Reinforcement Learning (DRL) for power control, Successive Interference Cancellation (SIC) for interference cancellation, and Edge AI for real-time optimization have shown promising results. However, such advancements are hindered by limitations like intensive computational complexity, scalability issues in dense networks, and dynamic real-time traffic adjustments. Additionally, existing models usually do not have effective energy management policies, leading to increased power consumption. To overcome such limitations, the current work focuses on optimization, with the promise of intelligent, adaptive resource allocation for enhancing spectrum efficiency, energy efficiency, and reliability in 6G networks. Significance of the study This study is significant as it envisions an intelligent, AI-powered method to achieve maximum spectrum and energy efficiency for 6G wireless systems. By applying Deep Q-Network (DQN)-based Successive Interference Cancellation (SIC) with Massive MIMO and MC-CDMA, the envisaged system achieves real-time resource allocation, interference reduction, and power management efficiently. This enhances the network performance, reliability, and scalability, making it an ideal enabler for future autonomous wireless communication systems. The findings of this study will benefit network operators, industries, and researchers through an improved, adaptive, and sustainable approach for future 6G rollouts. Key contributions Implementation of Deep Q-Network (DQN)-based SIC to minimize multi-user interference and improve signal detection.A reinforcement learning-based power control mechanism that dynamically adjusts power levels to maximize energy efficiency.Enhanced throughput, reliability, and scalability for dense and dynamic network environments in 6G.A framework that ensures real-time adaptation to changing network conditions, making it suitable for next-generation wireless systems. Implementation of Deep Q-Network (DQN)-based SIC to minimize multi-user interference and improve signal detection. Implementation of Deep Q-Network (DQN)-based SIC to minimize multi-user interference and improve signal detection. A reinforcement learning-based power control mechanism that dynamically adjusts power levels to maximize energy efficiency. A reinforcement learning-based power control mechanism that dynamically adjusts power levels to maximize energy efficiency. Enhanced throughput, reliability, and scalability for dense and dynamic network environments in 6G. Enhanced throughput, reliability, and scalability for dense and dynamic network environments in 6G. A framework that ensures real-time adaptation to changing network conditions, making it suitable for next-generation wireless systems. A framework that ensures real-time adaptation to changing network conditions, making it suitable for next-generation wireless systems. Rest of the section The rest of the section is divided into the following sections: “Literature survey” section Related Works describe the existing technique briefly and mention their drawbacks in relation to the spectral and energy efficiency. In “Methodology of proposed MIMO MC-CDMA with SIC and deep reinforcement learning for 6G optimizing spectrum and energy efficiency” section, the details of the proposed Edge AI-enabled MIMO MC-CDMA system are explained that includes DRL for the resource allocation, type of interference cancellation methods and few parameters of the overall system model. “Results and discussions” section focuses on the assessment of the system performance in terms of SE, EE, SINR, and BER relatively to conventional models. The conclusion and future works are highlighted in "Conclusion and future work" section which also recommends better implementation with the real-time method. Literature survey Siddiqui et al.14 have discussed on the problems of escalating mobile data traffic, density wireless devices and the transformation of industrial radio monitoring. The study had an ambitious proposal to incorporate mMIMO with HetNet and to utilise ML, FL, DL, and ANN in enhancing network performance. The findings pointed out that, with AI solutions, the problems related to mMIMO system implementation include computational complexities and latency. However, the work was not without its limitations wherein, the computational complexity of optimising mMIMO using AI was the most conspicuous hindrance to network adaptability. The paper seeks to act as a reference to the B5G/6G researchers and operators in the development of the wireless communication systems. Almutairi and Krishna15 explored the drawbacks of using F-OFDM for the fifth-generation and beyond wireless communication technology. To this end, they introduced a new MCM approach named filtered orthogonal wavelet division multiplexing (F-OWDM) for substituting FT with Wavelet Transforms. Their research compared and simulated the F-OWDM with noisy and fading environment defined by additive white Gaussian noise and flat fading channels in aspects of PAPR and BER and different types of wavelets such as Haar wavelet, discrete Meyer wavelet, Bi-orthogonal, symlet and Daubechies wavelet. It has been established from the results that the complexity of F-OWDM is less than C-F-OFDM with low PAPR and BER. However, a major limitation is seen in the factor that practice,wavelet-based MCM is not easily possible because of its complexity. Katwe et al.13 proposed a new resource allocation based on MU-MIMO downlink system with ARIS for the URLLC enabling RSMA. Because the formulation was non-convex in nature, it optimizes beamforming design using successive convex approximation and ARIS positioning using PSO. It was evident from the result that fast convergence and higher throughput rate was achieved compared to other multiple access techniques. Nevertheless, optimization framework still remains an issue for the application in real-time use cases in URLLC. Chen et al.16 also studied the issue of coordinating AP and UE in cell-free mMIMO setting. To this end, the authors presented an Interference-aware massive access (IAMA) scheme, AP-UE association and pilot assignment parameters are optimized based on the large-scale interference. They came up with an interference-aware reward metric and designed two iterative algorithms to optimization of the association and the pilots. Additionally, higher signalling overhead of interference awareness for optimization may have negative effects on the network performance in large scale systems. Boukenadil et al.10 pursued a channel coding technique based on deep learning framework for LDS-CDMA and NOMA with an intent to mitigate errors, minimize BER and decrease computational overhead. The research showed that deep learning could be effectively applied to necessary efficiency in wireless communication, as well as in the development of 6G and next, in terms of adaptability, data rate, and latency. However, one limitation is that training the model requires large amounts of training data which is highly dependent on network condition. Innovation could be done to assess the benefit of integrating with massive MIMO as well as NOMA for further prosecution in the efficiency of the system. Jeon and Jin17 proposed random access systems that work with successive interference cancellation (SIC) for decoding multiple simultaneous transmissions by an access point. They used a Bayesian estimation technique to dynamically estimate the number of users currently active and optimize the transmission control accordingly. Using their approach, they achieved throughput up to 0.693 packets per slot which is equivalent to the theoretical limit proposed by Yu—Giannakis and maintained it at 85.7% even with severely restricted SIC capacity of just two. The most important drawback of such a system is that it relies on accurate estimation in real time, which could be definitely troubled by rapid changes that occur in the network conditions. Ahmad et al.18 also viewed the prospects of 5G wired towards connectivity needs and the degree of imperfection introduced by the applications of Successive Interference Cancellation in power-domain Non-Orthogonal Multiple Access. They thus suggested pulse shaping based on wavelet transform for the purpose of reducing multiuser interference and derived expression of the Symbol Error Rate (SER) considering a three-user case. Following this, a numerical comparison against NOMA using Fourier transform showed that the wavelet-based modulation scheme performed better. However, the implementation in real-time systems is still stringent due to the increased signal processing burden. Further performance gain may vary with changing channel conditions, which will also require more optimization. Whatever the constraints, their proposal for calculation may enhance reliability in multiuser scenarios. Kansal et al.19 consider hybrid massive multiple-input multiple-output (mMIMO) and orthogonal frequency division multiplexing (OFDM) techniques to deliver multimedia in the future 5G and 6G mobile networks. The authors have modeled the fast Fourier transform (FFT), fractional Fourier transform (FRFT), and discrete wavelet transform (DWT) under Rayleigh fading and M-PSK modulation conditions. Among these, FrFT and DWT have been found to improve image quality over an FFT-based system. However, there are issues related to increased complexity of computation and hardware requirements implying that the study is not real-time latency constrained, which is typically undesirable in immediate-use multimedia applications. However, robustness and efficiency improvements have been achieved in wireless image communication even with these demerits. Singhal and Patil11 introduced HCR-WSN, a Hybrid MIMO Cognitive Radio Wireless Sensor Network, which integrates underlay and interleave spectrum-sharing schemes to improve spectral efficiency. The system dynamically switches between Narrowband and MC-DS-CDMA while ensuring minimum interference to primary users (PU). In the absence of PU, secondary users (SU) optimize spectrum through interference alignment. Results increased throughput,however real-time transmission selection adds to system complexity. The impact of high mobility scenarios remains unexplored. The framework does provide excellent enhancement for spectral utilization and proves efficient in a dynamic wireless environment. Vijay and Umadevi20 studied the synergy of MC-CDMA and massive MIMO toward spectral efficiency gains and overcoming multi-user co-channel interference (CCI) in 5G networks. Based on this approach, they devised an optimal spectrum sharing (OSS) protocol and optimal linear precoding (OLP) in conjunction with swarm algorithms for improved system capacity and power allocation. Their MATLAB-based simulation outperformed all conventional approaches, but algorithmic complexity and dynamic network conditions are pertinent considerations in practical implementations. This study demonstrates the promise of MC-CDMA and massive MIMO integration,however, these techniques still need to be optimized for real-world applications. Venkateswarlu and Rao21 identified a mutation weight adaptive particle swarm optimization (MAPSO) approach to address problems on channel estimation of MIMO-OFDM. Their method consisted of four phases, including space-time block coding (STBC) and MC-CDMA for optimal channel evaluation. MAPSO tunes parameters dynamically and gives reduced bit error rates (BER) with improved performance of the whole system. However, the computational overhead will be an issue for real-time applications in this approach. No variations have been considered in the study with respect to different channel conditions or mobility scenarios. In spite of that limitations, their work outperformed the existing methods in terms of SER, BER, MSE, and NMSE. Azimi et al.12 developed the energy-efficient approach called deep reinforcement learning-assisted resource allocation (EE-DRL-RA) for slicing into 5G networks. Deep learning (DL) was used for long-term decisions, while deep reinforcement learning (DRL) was used for instant resource allocations. It optimizes power-slot and radio resource distribution while guaranteeing that slice isolation is maintained. From simulations, it was seen that the algorithm outperformed the existing methods in speed of convergence, energy efficiency, and acceptance rate of users. Nevertheless, the concerns that remain are regarding scalability to ultra-dense networks, adaptability for dynamic traffic, and computational complexity. Recognition is given to the potential of EE-DRL-RA, but an inability to generalize its models and impose further constraints to actual-time deployment environments in practical 5G settings was reported. Current literature on 5G systems states that the energy efficiency ranges of 2–10 bits/Joule are generally considered, which is why smart optimization methods should be utilized to enhance the performance. The overall findings of existing studies are given in Table 1. Table 1Summary of existing studies.AuthorsKey focusProposed methodAdvantagesLimitationsSiddiqui et al.14AI-driven optimization of mMIMO in B5G/6GAI techniques (ML, FL, DL, NN) for HetNet-mMIMOOptimized network performance, reduced latencyAI model interpretability and real-time adaptabilityAlmutairi and Krishna15F-OFDM vs. F-OWDM for 5GReplacing FT with Wavelet Transform in F-OFDMLower PAPR, better BER, no cyclic prefixCompatibility with existing OFDM-based infrastructureKatwe et al.13Resource allocation for MU-MIMO using ARISAlternating optimization (beamforming + PSO)Improved spectral efficiency in dead zonesHardware feasibility of ARIS deploymentChen et al.16AP-UE association in CF mMIMOInterference-Aware Massive Access (IAMA)Improved fairness and SEScalability in massive access scenariosBoukenadil et al.10Deep learning in LDS-CDMA and NOMANeural networks for channel codingLower BER in low SNRDeployment in real-time systemsJeon and Jin17Slotted random access with SICBayesian estimation for active usersHigher throughput under limited SICHandling dynamic user arrival ratesAhmad et al.18NOMA with wavelet transformsWavelet pulse shaping for SIC mitigationReduced multi-user interferenceImplementation in real-world NOMA networksKansal et al.19Hybrid mMIMO-OFDM for multimedia transmissionEvaluating FFT, FrFT, and DWTHigher PSNR, SSIM with FrFT/DWTReal-time processing overhead in multimedia transmissionSinghal and Patil11Hybrid MIMO CR for WSNUnderlay + Interweave CRHigher spectral efficiency in multi-user scenariosManaging interference with PU-SU coexistenceVijay and Umadevi20MC-CDMA with mMIMOOSS + OLP with Swarm AlgorithmHigher system capacity, reduced BERComplexity of spectrum sharing in dense networksVenkateswarlu and Rao21MIMO-OFDM channel estimationMAPSO algorithmLower SER, BER, MSEReal-time adaptation to varying channel conditionsAzimi et al.12Resource allocation in 5G slicingEE-DRL-RA (DL + DRL for RAN slicing)Energy efficiency, better isolationGeneralization of learning models in dynamic networks Table 1Summary of existing studies.AuthorsKey focusProposed methodAdvantagesLimitationsSiddiqui et al.14AI-driven optimization of mMIMO in B5G/6GAI techniques (ML, FL, DL, NN) for HetNet-mMIMOOptimized network performance, reduced latencyAI model interpretability and real-time adaptabilityAlmutairi and Krishna15F-OFDM vs. F-OWDM for 5GReplacing FT with Wavelet Transform in F-OFDMLower PAPR, better BER, no cyclic prefixCompatibility with existing OFDM-based infrastructureKatwe et al.13Resource allocation for MU-MIMO using ARISAlternating optimization (beamforming + PSO)Improved spectral efficiency in dead zonesHardware feasibility of ARIS deploymentChen et al.16AP-UE association in CF mMIMOInterference-Aware Massive Access (IAMA)Improved fairness and SEScalability in massive access scenariosBoukenadil et al.10Deep learning in LDS-CDMA and NOMANeural networks for channel codingLower BER in low SNRDeployment in real-time systemsJeon and Jin17Slotted random access with SICBayesian estimation for active usersHigher throughput under limited SICHandling dynamic user arrival ratesAhmad et al.18NOMA with wavelet transformsWavelet pulse shaping for SIC mitigationReduced multi-user interferenceImplementation in real-world NOMA networksKansal et al.19Hybrid mMIMO-OFDM for multimedia transmissionEvaluating FFT, FrFT, and DWTHigher PSNR, SSIM with FrFT/DWTReal-time processing overhead in multimedia transmissionSinghal and Patil11Hybrid MIMO CR for WSNUnderlay + Interweave CRHigher spectral efficiency in multi-user scenariosManaging interference with PU-SU coexistenceVijay and Umadevi20MC-CDMA with mMIMOOSS + OLP with Swarm AlgorithmHigher system capacity, reduced BERComplexity of spectrum sharing in dense networksVenkateswarlu and Rao21MIMO-OFDM channel estimationMAPSO algorithmLower SER, BER, MSEReal-time adaptation to varying channel conditionsAzimi et al.12Resource allocation in 5G slicingEE-DRL-RA (DL + DRL for RAN slicing)Energy efficiency, better isolationGeneralization of learning models in dynamic networks Summary of existing studies. Methodology of proposed MIMO MC-CDMA with SIC and deep reinforcement learning for 6G optimizing spectrum and energy efficiency The integrated methodology proposes the MIMO MC-CDMA with Edge AI optimization towards better spectral and energy efficiency in 6G wireless networks. The signal-first encodes user data by MC-CDMA spreading and OFDM modulation, followed by MIMO transmission at the base station. The transmitted signals encounter a frequency-selective fading channel, multi-user interference (MUI), and noise. The interference is suppressed to improve the SINR at the receiver using MIMO detection and Successive Interference Cancellation (SIC). Edge AI is also employed for adaptive beamforming and power control to achieve real-time resource allocation optimization based on Channel State Information (CSI). The system uses DRL for parameter updates in an iterative way, so as to minimize interference and improve reliability. In this way, latency is reduced, spectral efficiency increased, and interference cancellation made robust through this AI-based approach-a key distinguisher for next-generation 6G networks. The proposed system uses an edge AI board to create intelligent functionality that extends beyond traditional resource allocation functions. Additionally, the imported AI circuit can allow for adaptive beam forming by dynamic adjustments of beam forming weights via real time Channel State Information (CSI), therefore improving signal performance and reducing interference at the same time. Finally, the AI circuit can perform intelligent scheduling decisions by assigning higher priority to users with higher traffic demand (or lower latency requirement) in combination with real world conditions present in the network at the moment. Similarly, this AI circuit will provide a better solution for managing interference by using Depth Reinforcement Learning (DRL) to optimize the SIC ordering scheme. Finally, the provision of these capabilities from the edge will minimize the time to effective service, improve the time required to receive a response from another device, enhance Quality of Service (QoS) and increase the overall efficiency of the network using edge AI circuits in 6G communications environments. This architecture shown in Fig. 1. that employs AI to assist in processing data will provide an efficient and flexible structure to build 6G Wireless Networks using MIMO (Multiple Input Multiple Output). This type of communication supports MIMO User Data and channel information will be processed at a Base Station using MIMO, MC-CDMA (Multi code (Multi-Carrier) CDMA) and Edge AI Beam forming thereby improving the Quality of the Signal and decreasing interference on the network. Considering Flaky Signal Strength (Fading) and Channel Interference during transmission of Wireless signals, MIMO detection, Successive Interference Cancellation, and MC-CDMA Processing (Dispreading) would take place at the receiver site. Once all information is processed, Adaptive Power Optimization (APO) through Deep Reinforcement Learning (AI driven) would optimize spectral efficiency, energy efficiency and overall network communication reliability. Fig. 1Flow diagram for proposed system. Flow diagram for proposed system. System model of edge AI-enabled MIMO MC-CDMA for 6G networks with SIC and DRL-based optimization Multi-Carrier Code Division Multiple Access (MC-CDMA) is a new wireless communication technique that is based on the integration of Orthogonal Frequency Division Multiplexing (OFDM) with Code Division Multiple Access (CDMA) to deliver enhanced spectral efficiency and enhanced resistance to multipath fading22. MIMO technology added to it is called MIMO MC-CDMA, which is a new paradigm for high-capacity wireless communications in next-generation networks such as 5G and 6G. In MIMO MC-CDMA, the system exploits spatial diversity and improves signal quality with multiple antennas at the receiver and transmitter. MIMO technology enables multiple data streams to be sent simultaneously over the same frequency band, significantly improving channel capacity and spectral efficiency. Alternatively, MC-CDMA spreads each user’s data symbols across a number of subcarriers using individual spreading codes in order to achieve robust multi-user access while combating the effects of narrowband interference and deep fades. The received signal model is used to calculate the per-user SNR based on the desired signal component, multi-user interference and noise. The transmission process of signals in MIMO MC-CDMA involves several crucial steps: First, the data symbols of different users are spread across a wideband channel using orthogonal spreading codes such as Walsh-Hadamard. The spread symbols are subsequently mapped onto an array of subcarriers using OFDM, effectively transforming the wideband channel into a parallel array of narrowband channels with reduced inter-symbol interference (ISI). The MIMO configuration also enhances system performance by employing techniques such as spatial multiplexing, beamforming, and diversity schemes to improve the SINR and BER. Multi-user interference (MUI) is one of the basic issues in MIMO MC-CDMA and arises from overlaid transmissions in heavily loaded network environments. To combat this, Successive Interference Cancellation (SIC) is implemented at the receiver, cancelling and detecting iteratively the strongest interfering signals first, thereby enhancing the detection of weaker signals. MIMO MC-CDMA is one of the facilitators for future wireless systems, enabling enhanced spectral efficiency, multipath fading resistance, and adaptive interference control, making it highly appropriate for future 6G networks and beyond. The Edge AI module, in turn, controls this adaptive beamforming mechanism and continuously learns to adapt and update the transmission strategies depending on the current network conditions. The proposed system integrates Multiple-Input Multiple-Output (MIMO) Multi-Carrier Code Division Multiple Access (MC-CDMA) into a 6G wireless communication system optimized using Edge AI. It is depicted in Fig. 2. It is designed to provide both energy efficiency and spectrum efficiency with low latency and reliability, especially in high-density network scenarios. Fig. 2System model for edge AI-enabled DRL-based MIMO MC-CDMA communication framework with SIC. System model for edge AI-enabled DRL-based MIMO MC-CDMA communication framework with SIC. The multi-user MIMO MC-CDMA architecture with the transmitter, channel, and receiver components is also clearly defined which improves the system model. The major assumptions like channel conditions, noise model and user distribution are given. As well, the processes of Edge AI-based DRL optimization, resources allocation, and SIC-based interference cancellation are also included in such a way as to have a fully representative system. The system is an array base station (BS) with multiple antennas, enabling the support of numerous users to get data through frequency-selective fading channels. MIMO exploits spatial diversity, while MC-CDMA provides extension of multi-user access through OFDM and CDMA techniques in combination. The base station (BS) with multiple antennas serves multiple users, each receiving data over dynamically changing 6G wireless channels. The transmitted signal vector of the base station is given as in Eq. (1),1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X = \left[ {X_{1} ,X_{2} ,...,X_{K} } \right]^{T}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{K}$$\end{document} is the transmitted signal of MC-CDMA for user k, and K is the total number of users. The MC-CDMA transmission of every user is spread across multiple subcarriers using unique spreading codes. The transmitted signal for user \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k }^{th}$$\end{document} is given in Eq. (2),2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{k} = s_{k} C_{k} F$$\end{document}where, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${s}_{k}$$\end{document} is the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}^{th}$$\end{document} user data symbol, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C}_{k}$$\end{document} is the spreading code matrix for the user. F is the Fourier Transform matrix, transforming the spread signal to orthogonal subcarriers. The conversion between the transmitted signal model of the equations takes place as follows: The received signal is of the form (1) and (2) to the received signal formulation of the form in the equation. The effect of the wireless channel is what lies behind (3). The transmitted signal vector (comprising several users) is combined and sent to the MIMO channel matrix which causes fading and signal distortions. The channel coefficients multiplied each transmitted signal component. The signals of all the users are then overlaid on the receiver and additive noise is added. This leads to the expression of the received signal whereby the total signal is expressed, as the channel-affected transmitted signal added to noise. The received signal Y at the user’s end is given in Eq. (3),3Y = HX + Nwhere H is the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{r}{\times N}_{t}$$\end{document} MIMO channel matrix that represents fading coefficients between the receive and transmit antennas. N is the additive white Gaussian noise (AWGN) with mean zero and variance of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma }^{2}$$\end{document}. At the receiver, the received signal is subjected to MIMO detection and MC-CDMA de-spreading. The estimated symbol for user k is computed as in Eq. (4).4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{s}_{{k = W_{k} Y}}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}_{k}$$\end{document} is the receiver detection matrix, used to mitigate multi-user interference and improve signal recovery. The system utilizes the MIMO detection method to enable effective data separation, thereby enabling high-capacity communication even for high-density user conditions. All the variables are defined in a better way in the form of Eqs. (1–4). The transmitted signal can be denoted as X = [X 1, X 2, and so on, X n] 0, with K being the number of users. In the case of an individual user, the signal sent can be expressed as X 0 and X 1 = s 0 C 0 and s 1 C 1 respectively where s 0 and s 1 are the data symbols and C 0 and C 1 are the spreading codes. The signal that is received can be expressed as Y = HX + N where H is the channel matrix, Y is the signal received and N is noise. The estimated signal is calculated ŝ k = w kY with w k being the receiver weight to detect the signal. In order to simplify the analysis, several assumptions are used to develop the system model. The assumption made is that the channel adheres to the flat fading model and is constant within a transmission interval. Noise is assumed to be an additive white Gaussian noise (AWGN) where the mean of noise is zero and variance is constant. Assumptions are made that all users are co-synchronized and that they have equal power to start with. The quantity of transmit and receive antennas is fixed and the perfect channel state information (CSI) is considered at the receiver. Also, the DRL agent is based on discrete state and action space and can effectively learn and make decisions in a multi-user environment. The assumption is that the channel state information (CSI) is perfect at the receiver where the signal is properly detected and decoded. But in real life situations, pilot signals may be used to estimate CSI thereby creating errors in estimation. DRL-based optimization can preadapt the proposed model to such variations, which guarantees a strong performance in imperfect channel conditions. It is assumed that the users are perfectly synchronized in time and frequency to prevent inter-symbol and inter-carrier interference. The spreading codes, which are applied in MC-CDMA system, are also said to be orthogonal meaning the minimum amount of multi-user interference in the ideal conditions. Nonetheless, there are possibilities of a slight loss of orthogonality in real life situations because of channel impairments and the proposed DRL-SIC mechanism caters to this. Channel model is per-subcarrier in nature in order to model frequency-selective behavior within the MC-CDMA system. The subcarriers undergo separate flat fading channels, which are characterized as complex gain coefficients. In this way, the entire channel matrix is broken down to a number of subcarrier-specific channel responses. The signal at each subcarrier received is the product of the transmitted signal and the channel gain of the channel and the noise. The method would make analysis easier as it transforms a frequency selective channel into parallel flat fading sub channels and allow efficient signal detection and allocation of resources with the proposed DRL based framework. The use of Edge AI in the system enables real-time adaptive beamforming and power control to maximize both energy efficiency and spectrum usage. The beamforming weight vector ω is updated based on the estimated channel state information (CSI) dynamically, and this maximizes the signal-to-interference-plus-noise ratio (SINR). The optimal beamforming weight is delivered as given in Eq. (5).5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w^{*} = \arg \max_{w} \frac{{\left| { w^{H} H_{k} s_{k} } \right|^{2} }}{{\mathop \sum \nolimits_{j \ne k} \left| { w^{H} H_{j} s_{j} } \right|^{2} + \sigma^{2} }}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H}_{k}$$\end{document} is user \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k^{\prime } s$$\end{document} channel matrix, and w is the beamforming weight vector. The beamforming update in Eq. (5) modulates the weights of the transmissions in accordance with channel state information (CSI). The system in this process computes the ideal beamforming vectors to send the signal to the target user and reduce the interference to other users. The weights are dynamically updated as per the varying channel conditions. This increases signal strength, SINR and efficient transmission. The system is characterized by enhanced spectral efficiency and assured multi-user communication by continually changing the beamforming parameters. Through this adaptive optimization mechanism, the system learns and updates beamforming strategies continuously such that the transmitted signals cause minimal interference and attain maximum energy efficiency and throughput. Furthermore, Edge AI transfers the computational load from the central cloud to local edge nodes, and it significantly reduces latency, enabling rapid response times in real-time wireless applications. Consequently, the proposed Edge AI-enhanced MIMO MC-CDMA system is the foundation of achieving high-performance and low-latency wireless communication in 6G networks. Further optimisations, such as resource allocation and interference cancellation, will be elaborated upon in subsequent sections. Spectrum efficiency optimization with successive interference cancellation (SIC) In a MIMO MC-CDMA system, multiple users transmit signals in parallel over the same frequency band, resulting in multi-user interference (MUI). This interference seriously degrades the overall spectral efficiency of the system, impacting data rate performance and increasing the bit error rate (BER). At the receiver side, Successive Interference Cancellation (SIC) is employed to suppress these effects, identifying and cancelling out the stronger signals in sequence first and enhancing the reception of weaker users’ signals23. Through sequential cancellation of strong interfering signals, SIC improves signal-to-interference-plus-noise ratio (SINR), resulting in more effective use of the spectrum. The SIC procedure begins with the ordering of the incoming signals according to their power levels such that the most powerful signal is received and decoded first. It is shown in Fig. 3. For estimating this signal, the Minimum Mean Square Error (MMSE) detection method is employed, which optimally reduce noise for multi-user interference. The estimated signal is reconstructed based on the known channel response and then subtracted from the received signal vector. This recursive cancellation procedure continues, gradually refining the current signal terms and enabling accurate detection of the weaker signals. SIC-based signal estimation through MMSE is defined mathematically by Eq. (6).6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widehat{{x}_{1}}={\text{arg min}}_{{x}_{1}}|\left|{y}_{1}- {h}_{1}{x}_{1}|\right|$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widehat{{x}_{1}}$$\end{document} is the cancelled signal, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${y}_{ 1}$$\end{document} is the received signal, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h}_{1}$$\end{document} is the channel coefficient. The best signal is detected and reconstructed utilizing the known channel response and then subtracted from the received signal. Fig. 3Working of successive interference cancellation. Working of successive interference cancellation. After cancellation, the updated received signal is represented as Eq. (7).7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Y}^{\prime}=Y- {H}_{1}\widehat{{X}_{1}}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Y}^{\prime}$$\end{document} is the residual received signal after removing the strongest detected signal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widehat{{x}_{1}}.$$\end{document} This process is iteratively repeated, refining received signal as given in Eq. (8).8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Y}_{new}=y- H\widehat{X}$$\end{document} With a distinct division between desired/desired signals, interference and noise, the proposed formulation for the SINR of component signal will provide an unbiased evaluation of the system-level performance. A calculation of the useful power of the received signal will be completed by using the gain of the channel and the transmit power allocated for use, while interference will be generated from all users who are transmitting on the same frequency (known as co-channel multi-user transmissions) and the noise will be characterized as AWGN with a fixed level of variance. Spectral efficiency can be calculated using Shannon’s definition of capacity based on the calculated SINR, allowing for realistic prediction of performance. Equations (6–8) incorporate a term representing residual interference after SIC to account for practical imperfections caused by channel estimation error and noise. Using the DRL based resource allocation framework, the power control and detection order can be dynamically adjusted to minimize residual interference so as to improve SINR, spectral efficiency and overall system reliability in dense multi-user scenarios. This successive cancellation of strong interfering signals leads to a cumulative SINR enhancement for the remaining users. The SINR for user k after SIC is characterized by Eq. (9).9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${SINR}_{k}=\left(\frac{{P}_{k}{|{h}_{k|}}^{2}}{\sum_{j>k}{P}_{j}{|{h}_{j|}}^{2}+ {\sigma }^{2}}\right)$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{k}$$\end{document} is the transmission power, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h}_{k}$$\end{document} is channel gain and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma }^{2}$$\end{document} is the noise power. SIC reduces interference term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sum_{j>k}{P}_{j}{|{h}_{j|}}^{2}$$\end{document}, thereby enhancing SINR, minimizing Bit Error Rate (BER) and making data transmission efficient. The term interference in the Eq. (9) consists of inter-user and intra-user interference. The MC-CDMA system encounters inter-user interference due to two or more users using the same spectrum, and the overlaps in the signal of other users. Perfect orthogonality among the subcarriers or spreading codes may not be perfect and intra-user interference may take place. The overall interference is defined as the combination of all these, and noise. This detailed modelling enables the precise assessment of SINR and the DRL based optimization to be used to minimise the interference and provide a better overall system performance. From the point of spectral efficiency, the iterative interference cancelation in SIC significantly improves the data rate per unit bandwidth supported. Spectrum efficiency (SE) for SIC-based MIMO MC-CDMA system can be expressed as in Eq. (10).10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$SE= \sum_{k=1}^{K}{log}_{2}\left(1+\frac{{P}_{k}{|{h}_{k|}}^{2}}{\sum_{j>k}{P}_{j}{|{h}_{j|}}^{2}+ {\sigma }^{2}}\right)$$\end{document} Spectral efficiency is calculated as the sum-rate of all the users in the system. This is done by adding the individual user data rates, and the individual data rate is determined depending on the SINR of an individual user. Also, there is bandwidth normalization and the normal output is converted to bits/s/Hz so that it could be reasonably compared with different systems configurations. It is a formulation with a holistic view of system performance and enables the proposed system of DRL to efficiently optimize the total spectral usage. where K is the total users. With the sequential removal of interference by SIC, each user experiences an improved SINR and consequently logarithmic gains in spectrum efficiency. Under the 6G wireless network architecture, integration of SIC in Edge AI-driven MIMO MC-CDMA systems optimizes resource allocation and interference cancellation algorithms. Using deep learning models, SIC can dynamically adjust detection order and cancellation procedures in real-time according to real-world network environments. The intelligent, AI-driven SIC system optimizes spectral efficiency, reduces latency, and enhances communication reliability, making it a critical component of next-generation wireless systems. The flowchart of SIC in Edge AI-driven MIMO MC-CDMA is given in Fig. 4. Fig. 4Flowchart of successive interference cancellation algorithm. Flowchart of successive interference cancellation algorithm. Deep reinforcement learning (DRL) for energy efficiency and resource allocation In the future of 6G wireless networks, it is imperative to ensure proper resource allocation due to the very dynamic and changing nature of network environments. Traditional heuristic-based optimization methods, such as rule-based or convex optimization approaches, are likely to be hindered by difficulties in adjusting to real-time variations in channel conditions, user traffic, and energy demands. These conventional methods entail massive earlier experience of the network dynamics and may not work well in complex multi-user, multi-channel scenarios. To bridge the above disadvantages, Deep Reinforcement Learning (DRL) has emerged as a viable challenger by enabling intelligent and adaptive mechanism of resource allocations. DRL employs real-time learning and decision-making functionalities for dynamically allocating the power, spectrum, and scheduling resources with a promise of 6G MIMO MC-CDMA networks with low latency, high capacity, and low energy consumption. The architecture of Deep Reinforcement Learning is given in Fig. 5. Fig. 5Architecture of deep reinforcement learning. Architecture of deep reinforcement learning. The proposed DRL-based resource allocation scheme is formulated to achieve optimal spectrum efficiency (SE) and energy efficiency (EE) while minimizing multi-user interference (MUI)24. Unlike traditional static optimization methods, DRL leverages an iterative learning process in which an agent continuously updates and enhances its decision-making policy by directly interacting with the network environment. The DRL model acquires the optimal allocation strategy by trial and error through environment interaction and policy adjustment of decision-making. The DRL system includes the following key components, Inference delay per scheduling interval is the period taken by the DRL model to come up with a decision on how to allocate resources. This delay is reduced in the proposed system by running the model in the network edge. Generally, inference delay is exceedingly minimal relative to training time and is in the milliseconds range. This will make sure that decisions are made within each scheduling interval and yet will not interfere with the system performance. The low inference delay allows real-time adaptation, which makes the system effective in dynamical and delay-sensitive 6G wireless systems. State (St) is the network state it includes channel state information (CSI), energy consumption, user demand, and interference state. Action (At) represents the resource allocation decisions, e.g., power allocation, spectrum assignment, and transmission scheduling. Reward (Rt) tracks the performance of the system in energy efficiency (E) and spectrum utilization gain (SE). It is given in Eq. (11).11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{t}=\alpha . SE+ \beta . EE$$\end{document}where, α and β are weight factors for balancing spectrum efficiency and energy efficiency. The ratio of the sum of the maximum obtainable data rate to the sum of the power consumption of the system is referred to as the energy efficiency (EE). It is measured in bits per Joule, which is the efficiency of the system in data transmission of energy. The spectral efficiency provides the total data rate and the total power comprises the transmission power and circuit power consumption. This definition allows achieving a harmonious assessment of system performance taking into account the throughput and the use of energy. The suggested DRL-based scheme optimizes EE, in which the power allocation is changed dynamically according to the varying network conditions. Directly transferring the reward function to the network edge enhances the benefits of computation on the network edge. (11) through minimizing latency and energy usage. The reward is founded on both spectral efficiency and energy efficiency thus; a faster local processing can lead to making decisions faster and resource allocation is better. This helps to cut on the amount of power being wasted and enhances the performance of the systems. Subsequently, the two terms of efficiency in the reward function are positively influenced. Thus, Edge AI assists the DRL model with increasing rewards based on real-time optimization and minimizing time delays of dynamically changing network situations. SE (Spectrum Efficiency) is represented as Eq. (12).12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$SE= \sum_{k=1}^{k}{log}_{2}\left(1+\frac{{P}_{k}{|{h}_{k|}}^{2}}{\sum_{j>k}{P}_{j}{|{h}_{j|}}^{2}+ {\sigma }^{2}}\right)$$\end{document} EE (Energy Efficiency) is calculated by the ratio of total data rate and power consumption. It is given in Eq. (13).13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$EE= \frac{\sum_{k=1}^{K}{R}_{k}}{\sum_{k=1}^{K}{P}_{k}}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{k}$$\end{document} is the data rate and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P}_{k}$$\end{document} is user k’s transmission power. In the equation, the system efficiency is calculated on the whole system and not per-user basis as in the Eq. (13). It is determined by dividing the total sum-rate obtained by all the users and dividing by the total power consumption of the transmission and the circuit power. This system level definition will be a global network performance and efficiency measure. The proposed DRL-based framework guarantees the balanced allocation of resources and better overall network performance under different conditions of the users and the channels and is developed on the principles of energy efficiency on the system level. The proposed system has a direct connection to the formulation in terms of the improvement in its energy efficiency. The Edge AI module, which runs on DRL, is dynamically optimized and so could reduce the resource use and power allocation, as well as the unneeded transmission and processing power. The numerator to denominator ratio is greatly improved by reducing the power consumption and keeping or increasing the overall rate of data. This leads to increased efficiency of energy. Also, Edge AI allows quicker and local decision-making, which minimizes computation time and leads to better system-level energy efficiency. The total power consumption that is taken into account in this work: transmission power, circuit power, signal processing power, and Edge AI computation overhead. Hardware components including RF chains and baseband circuits are represented by circuit power, and modulation, detection, SIC among others by signal processing power. Also, the computational power of the Edge AI module, with decision-making based on DRL is added to the overall power consumption. All these components allow the proposed model to give realistic estimate of the energy efficiency and make correct assessment of the system performance in realistic 6G conditions. The Q-learning procedure updates the policy of the decision-making of the agent by updating the Q-values based on the rewards received. The Bellman equation updates the policy iteratively. It is given by Eq. (14).14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q\left( {S}_{t}, {A}_{t}\right)= Q\left( {S}_{t}, {A}_{t}\right)+\upeta [ {R}_{t}+ \gamma {max}_{A}Q\left({S}_{t+1,}A\right)-Q\left( {S}_{t}, {A}_{t}\right)]$$\end{document}where η is the learning rate. γ is the discount factor of rewards in the future, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${max}_{A}Q\left({S}_{t+1,}A\right)$$\end{document} is the maximum next state expected reward. The integration of the Eqs. (11–14) is based on the traditional reward-driven DRA approaches in that it simultaneously maximizes spectral efficiency, energy efficiency, and interference reduction in a single reward model. In comparison to the current solution that involves the single objective optimization, the proposed model balances dynamically various network parameters to ensure adaptive and stable operations in complex multi-user MIMO MC-CDMA environment. The representation of the state is channel conditions, SINR and user-specific parameters, which give the network status at the moment. Action space determines the power distribution and the timing of the decisions of various users. The reward function is developed according to maximizing the spectral efficiency, energy efficiency, and reducing the interference and power consumption. The policy is updated through the learning process with the help of observed rewards to reach optimal decisions within a long period. This formulation makes the system dynamic in various ways to the dynamism in network conditions and also achieve better performance. The convergence of the DRA model is calculated depending on the probability of the stabilization of the Q-values and the function of reward during training. In particular, convergence is attained when the difference between consecutive updates of Q-values is negligible, and the cumulative reward levels off in several training sessions. The Bellman update in equation Iteration (14) minimizes the error in the difference in time, so that the Q-values predicted tend towards the maximum values. Also, the efficiency and stability of learning is implemented by a given stopping criterion; a set of specified training episodes. DRA forward inference and back propagation steps are considered to include two major processes. When performing forward inference, the present state is entered into the neural network to give the Q-values of various actions. Depending on these values, the best course of action is taken. During backpropagation, the difference between predicted and target Q-values is determined and it is used to adjust network weights. This learning and re-learning process enhances the precision of decision-making, enabling resource allocation and adaptive optimisation under changing network dynamics, as shown in (Fig. 6). Fig. 6Flowchart of DRL for energy efficiency and resource allocation. Flowchart of DRL for energy efficiency and resource allocation. The dimension of state of a system is determined by the number of users and system parameters taken into account in the network. Particularly, the state vector contains the values of channel gain, the SINR, and the power allocation parameters of each user. A system having N users has state dimension as a concatenation of these features, with a size that is proportional to 3N. This representation records the state of the entire network at a particular step in time. With such a structured state, the DRL model can acquire effective strategies of allocating resources efficiently in a dynamically changing multi-user environment. Action space is measured to be the decisions on resource allocation to each user within the system. It contains the degrees of power distribution and scheduling activities of the users of each transmission period. In a system of N users, the action vector will have N elements with each element representing the amount of power an individual user will have. Depending on the type of DRL algorithm, the action space may be discrete levels or continuous values. This form of action representation allows the DRL agent to effectively optimize the management of interferences and performance of the system under dynamic network conditions. The neural network that is applied in the DRL framework is a fully connected Deep Q-Network made up of an input layer, several hidden layers, and an output layer. The size of the input layer is the same as the state vector dimension. To address the relationship of the complex system, two hidden layers are utilized with 128 and 64 neurons, respectively. The hidden layers use ReLU activation functions to add non-linearity, and the output layer uses a linear activation function to provide estimates of Q-values of each action. This architecture provides effective learning and correct decision making in resource allocation that is dynamic. The DRL model has the Adam optimizer that offers efficient and adaptive learning in complicated environments. It is the loss function of Mean Squared Error on the predicted Q-values and target Q-values of the Bellman equation. This loss term works to reduce the gap between the anticipated and actual rewards in training. An appropriate learning rate is chosen to guarantee a stable model convergence. Adam optimizer is used in conjunction with the MSE loss, which allows easy training of the DRL agent to optimal resources allocation and interference control. The past state, action and reward transitions are stored in an experience replay buffer which enhances learning stability and efficiency. The buffer size is chosen according to the need to trade between memory utilization and learning diversity. A large enough buffer also makes sure that the model trains on a large collection of past experiences that are not correlated with each other. In the given work, the buffer size is selected so as to keep representative samples of various network conditions without too much computation overhead. Such a design decision allows converging the DRM model steadily and enhances its generalization skills in dynamic wireless scenarios. The deep Q-network (DQN)-based SIC optimization algorithm is used to optimize energy and spectrum efficiency The combination of DQN with SIC very well helps in optimizing power allocation and spectrum efficiency in 6G MIMO MC-CDMA networks25. Different from conventional resource allocation methods which utilize heuristic or static ways for their operations, the DQN-based optimization works on an adaptive scheme that can apply real-time tracking of dynamic network conditions in ensuring available resources are used efficiently. Using deep reinforcement learning (DRL) principles, the system continuously learns from real-time observations of channel variations, user demands, and interference patterns, allowing informed decision-making within power control and spectrum assignment. One of the primary benefits of this method is improved spectrum efficiency, gained by dynamic interference cancellation with SIC. For multi-user scenarios, intense interfering signals significantly reduce the Signal-to-Interference-plus-Noise Ratio (SINR), and thus spectral efficiency. Yet, using the SIC method, the system successively detects, decodes, and cancels intense signals, reducing multi-user interference. The DQN-based algorithm improves the SIC process by dynamically choosing the order of detection and cancellation policy, improving spectral efficiency and supporting increased data rates in high-density network environments. Another important advantage is enhanced energy efficiency, realized through smart power allocation. Conventional power control policies result in energy wastage through inefficient resource utilization. On the other hand, the DQN-based approach learns optimal power allocation policies through real-time adjustment of transmission power levels based on current energy demand and interference limitations. Through reduction of energy wastage and optimal signal quality, the system optimizes energy consumption and network performance, leading to sustainable and energy-efficient 6G communications. In addition, the SIC optimization framework based on DQN ensures robust communication in highly dynamic and dense network environments. As 6G networks are designed to support ultra-low latency, massive connectivity, and high data rate applications, channel adaptation needs to be considered under highly dynamic network environments. The reinforcement learning model learns to adapt its policy by finding patterns in past data and approximating future network trends, hence facilitating real-time adaptive interference management and resource allocation. This makes the system highly resistant to channel fading, mobility, and network congestion, providing robust and smooth communication even under harsh deployment conditions. Deep Q- Network is chosen because it has less computational complexity and can be used in discrete action space than PPO and DDPG. In the suggested system, the decisions of resources allocation can be represented as discrete actions, which makes DQN more effective and consistent. PPO and DDPG are also developed to handle problems in continuous control, and they demand more computational resources, which might not be explicit in real-time Edge AI application. Second, DQN offers quicker convergence and reduced implementation which is why it is a viable option in dynamic wireless settings with tight latency and computing limits. Algorithm The deep Q-network (DQN)-based SIC optimization The deep Q-network (DQN)-based SIC optimization The intelligent edge AI-driven resource allocation framework supported by the DQN-based SIC will broadly increase spectral and energy efficiency, reduce interference, and guarantee high capacity in global communication; hence becomes a fundamental technical enabler of autonomous next-generation 6G wireless systems. The proposed DRL-based resource allocation framework thus provides intelligent and adaptive resource allocation to meet energy-efficient high-capacity communication in 6G MIMO MC-CDMA systems. The always learning and optimizing DRL resource distribution will greatly improve spectral efficiency, reduce interference, and minimize energy consumption. With this intelligent Edge AI-based resource allocation, a DQN-based SIC enhances spectral and energy efficiencies, cuts interference, and guarantees a high-capacity world communication system; hence it now becomes a core enablement technology of next generation autonomous 6G wireless systems. Computational complexity analysis MIMO processing, Successive Interference Cancellation (SIC), beamforming operations, and Deep Reinforcement Learning (DRL) based optimization have an impact on the computational complexity of the proposed Edge AI enabled MIMO MC-CDMA system. The number of users and subcarriers depend on the processing complexity of O(K × N) for MIMO MC-CDMA. Multiple users’ iterative interference canceller can be obtained with computational complexity of O(K2). The complexity of the update in the beamforming is mostly dependent on the complexity of the matrix inversion and optimisation of the beam weights and is of the order O(N3) with N being the number of antennas. The DRL component is expensive in terms of computation, due to the forward inference and backpropagation procedure performed in the neural network layers. The complexity of the DRL inference mainly depends on the number of hidden layers, number of neurons, number of training episodes and replay memory size. The proposed framework involves a more computational complex than the conventional MMSE only or non-learning approaches, because of the adaptive learning process that involves iterative SIC and DRL. But the proposed system provides better interference suppression, adaptive resource allocation and better spectral and energy efficiency. The number of users, number of antennas, replay memory size are factors influencing the scalability of the proposed framework. The more users that are connected to the network, the more capacity will be provided, but the more complex the SIC will be due to the sequential interference canceller. In the same way, the greater the number of antennas the greater the spatial diversity and the more reliable the communication, but the more the beamforming/computing/channel estimation calculations. There is a trade-off between more memory used and complexity of training, and better the stability and accuracy of learning by DRL, the larger the size of the replay memory. While there is more overhead to computation, Edge AI has the benefit of processing data at the point of use with minimal latency compared to cloud-based optimization. Edge-level processing is used to shorten the inference delay per scheduling interval, allowing for quicker real-time resource allocation, interference management, and network adaptability for dense and delay-sensitive 6G wireless communication environments. Clarification and strengthening of novelty The innovative aspect of this work consists of the application of Edge AI to MIMO MC-CDMA (multi-carrier code division multiple access) combined with Successive Interference Cancellation (SIC) and Deep Reinforcement Learning (DRL) providing real-time adaptable optimization solutions to 6G networks. The application of this technology is significantly different than current implementations of MIMO-OFDM or hybrid precoding systems as it allows for decentralized decision-making to occur at the edge of the network, thus reducing latency and the processing burden on the network. As the DRL agent optimizes and re-optimizes power allocation, user scheduling, and interference mitigation based upon real-time channel state information (CSI) the use of Edge AI allows for light-weight processing and fast convergence to an optimal solution. The joint optimization of these technologies greatly enhances spectral efficiency, energy efficiency, and SINR (signal-to-interference-plus-noise ratio) performance, especially in a high-density multi-user environment. Results and discussions In this section, simulative results of the developed Edge AI-enabled MIMO MC-CDMA (6G) system is discussed. Major areas of interest like DRL for resource management, spectral efficiency, energy efficiency as well as interference management techniques are discussed. The number of iterations in the DRL model is evaluated using Q-values and reward functions and its learning characteristic is analyzed as exhibiting stability. Differences between DRL based and the conventional scheme provide insights in to the spectral efficiency and energy efficiency rise under different user’s load. In addition, various interference cancellation techniques are also analyzed and the SIC is shown to be the best one among them. Comparisons with other existing models show that overall, the proposed system outperforms the existing system in terms of SE, EE, SINR, and BER. The findings of this study justify the improvement that arises when Edge AI and DRL are to be implemented in 6G MIMO MC-CDMA networks, and it unveils the findings that make it evident that the integration of such models is an effective solution for future wireless communication systems. The Signal-to-Noise Ratio (SNR) value of each user in multi-user MIMO MC-CDMA system is calculated independently taking into account the received signal power, multi-user interference and additive white Gaussian noise (AWGN). SNR of k-th user is the ratio of the actual power of the desired signal to the total power of all the interference and noise signals.15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${SNR}_{K}=\frac{{P}_{signal,k}}{{P}_{interface,k}+{P}_{noise,k}}$$\end{document} In this case, the signal power is determined based on channel state information (CSI) and beamforming weights, and the interference comes from other users. The final interference level is further reduced after Successive Interference Cancellation (SIC) processing. After demodulation and SIC decoding, the bit sequences transmitted and received by each user are compared to compute the Bit Error Rate (BER). It is the ratio of the incorrect bits received to the number of bits sent.16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{BER}} = \frac{{N_{error bits} }}{{N_{total bits} }}$$\end{document} Monte Carlo simulations are run several times to guarantee statistical accuracy and an average BER is calculated. Moreover, the system level parameters like channel model (Rayleigh fading), number of users, antenna configuration and transmit power levels are defined explicitly. The MATLAB simulation environment features AWGN channels and realistic fading conditions. The DRL training process is also set up with certain hyperparameters like the size of the replay memory, the number of episodes that will be trained, and the learning rate. Lastly, the performance is repeatedly tested using metrics such as spectral efficiency, energy efficiency, SINR, BER etc. to achieve consistency and reproducibility in dynamic network conditions shown in Eqs. (15 and 16). The research utilizes the simulation parameters shown in Table 2 to evaluate the proposed 6G MIMO MC-CDMA system. MATLAB served as the platform to conduct the network simulations which operated using frequencies at 28 GHz as well as 100 GHz. Realistic wireless conditions were simulated through the implementation of Rayleigh channel model parameters. The system operates at bandwidths from 400 MHz to 1 GHz and handles user numbers from 10 to 100 with base station antennas ranging from 64 to 256. Each user operating in the 6G MIMO MC-CDMA system possesses a setup with 64 subcarriers whose SNR values can range from 0 to 30 dB. Users’ transmission power operates between 0.1 and 2 W while the specified noise power stands at – 100 dBm. The signal detection process benefits from the implementation of Successive Interference Cancellation (SIC) technology. The Deep Reinforcement Learning (DRL) model demands a replay memory buffer spanning 104 to 10⁶ while using a gamma value 0.90 to 0.99 with a learning rate range of 0.001 to 0.01. The system goes through 5000 to 10000 training sessions to achieve its best operational outcomes. Table 2Simulation parameters.ParameterValue/RangeSimulation toolMatlabNetwork type6G MIMO MC-CDMACarrier frequency28 GHz/100 GHzChannel modelRayleighBandwidth400 MHz–1 GHzNumber of users (K)10–100Base station antennas64–256Subcarriers per user64SNR range0–30 dBTransmission power per user0.1W–2WNoise power (σ2)− 100 dBmInterference cancellationSuccessive Interference Cancellation (SIC)Replay memory size104–10⁶Discount factor (γ)0.90–0.99Learning rate (η)0.001–0.01Training episodes5000–10,000 Table 2Simulation parameters.ParameterValue/RangeSimulation toolMatlabNetwork type6G MIMO MC-CDMACarrier frequency28 GHz/100 GHzChannel modelRayleighBandwidth400 MHz–1 GHzNumber of users (K)10–100Base station antennas64–256Subcarriers per user64SNR range0–30 dBTransmission power per user0.1W–2WNoise power (σ2)− 100 dBmInterference cancellationSuccessive Interference Cancellation (SIC)Replay memory size104–10⁶Discount factor (γ)0.90–0.99Learning rate (η)0.001–0.01Training episodes5000–10,000 Simulation parameters. Convergence of DRL-based resource allocation The Fig. 7 displays Deep Q-Network (DQN) learning stability through Q-value and reward function R versus training episode numbers. The figure x-axis shows the Training Episodes while the y-axis shows the Q-value and Reward Value associated with each episode of training. The blue line demonstrates Q-values but reward values appear in the red line. During the earliest training period Q-values together with rewards exhibit substantial changes thus reflecting learning instability. Learning capabilities alongside better decision-making performance emerges as training continues which leads to increasing values. The variations in Q-values and rewards achieve stability at episode 604. The convergence level assessment demonstrates that Q-values experience 13.67% coefficient variation compared to 3.30% reward variation which reveals reward stabilization performs better than Q-values. The stability of performance throughout time indicates how well the DQN model learns to establish an optimum policy. Fig. 7Stability of learning using Q-values and reward functions. Stability of learning using Q-values and reward functions. Different deep reinforcement learning (DRL) models with Deep Q-Network (DQN), Deep Deterministic Policy Gradient (DDPG) and Proximal Policy Optimization (PPO) exhibit their reward function learning curve during various episodes in Fig. 8. The figure demonstrates the learning patterns as well as convergence behaviors of different models throughout time. The DDPG model presents irregular performance because of its design for continuous action spaces yet DQN maintains consistent and stable learning. The PPO algorithm shows faster convergence compared to DQN although it brings increased variance during optimization. The established results show that DQN allows stable learning behaviour which fits perfectly for optimizing power control and interference management within MIMO MC-CDMA systems for future 6G networks. This figure compares how DQN, PPO and DDPG methods learn over time. The x-axis is divided into number of episodes and the y-axis shows Rewards for each of those respective episodes this Figure compares how DQN, PPO and DDPG methods learn over time. The x-axis is divided into number of episodes and the y-axis shows Rewards for each of those respective episodes. Fig. 8Comparison of learning curve. Comparison of learning curve. Spectral efficiency (SE) vs. Number of users Figure 9 also depicts spectral efficiency against number of users for Traditional SIC, while the second one shows the number of users against Spectral efficiency for DRL-Based SIC. This Figure also depicts spectral efficiency against number of users for Traditional SIC, while the second one shows the number of users against Spectral efficiency for DRL-Based SIC. Thus, it can be noted that both the approaches are reaching out in their utilization of the system and with more and more users, the system is becoming spectrally efficient in both systems and is having a linear scalability also. The SIC Using DRL is slightly different from the Traditional SIC and has a slightly better performance at the higher user loads. This reveals that the deep reinforcement learning (DRL) achieves the interference cancellation in a better way thereby improving spectral efficiency of dense networks. This along with other results gives a glimpse that DRL has a strong impact in enhancing the other key performance indicators in 6G wireless communication systems. Fig. 9Spectral efficiency (SE) vs. Number of users. Spectral efficiency (SE) vs. Number of users. Energy efficiency (EE) vs. User load The Fig. 10 shows how energy efficiency changes as the user count increases when using both this proposed DRL-Based Successive Interference Cancellation (SIC) method and regular SIC method. Also, This Figure shows Energy Efficiency performance results with User Load on the x-axis and presents Energy Efficiency measured in bits/Joule. Performance measurements indicate that DRL-Based SIC demonstrates superior performance than traditional SIC methods at all observed user load conditions. Higher energy efficiency is achieved through the DRL-based approach because its intelligent resource allocation and adaptive interference cancellation strategies work together efficiently for performance maintenance. Energy efficiency stands lower in traditional SIC methods because they demonstrate limited capability to combine interference management with power consumption optimization particularly at high user scales. The user density increases burdens conventional methodologies because it worsens interference and misallocates resources thus reducing overall performance. The proposed DRL-Based SIC method achieves dynamic network adaptation because it optimizes power spectrum usage throughout changing network conditions. Fig. 10Energy efficiency (EE) vs. User load. Energy efficiency (EE) vs. User load. The achieved energy efficiency of 14.8 bits/Joule is relatively better than the traditional wireless systems, but this enhancement is explained by the incorporation of the Edge AI, DRL-based power optimization, and SIC-supported interference reduction. The energy efficiency in an average 5G base station system typically varies between 2 to 10 bits/Joule and varies depending on network load, transmission power, and deployment conditions. The proposed system goes beyond this limit because of the dynamic power distribution, fewer retransmissions, and effective interference control. Also edge-based processing reduces the energy usage tied to the centralized computation, with additional efficiency. Thus, the value obtained is realistic in the settings of optimized 6G-driven intelligent communication systems. Performance comparison Different interference cancellation strategies for MIMO MC-CDMA systems are evaluated through Signal-to-Interference-plus-Noise Ratio measurements across multiple users which are shown in Fig. 11. The user index appears on the x-axis whereas the y-axis shows SINR values in dB. Users experience severe signal interference when “No Cancellation” method is used because it generates extremely negative SINR values across most users. The “Zero-Forcing” method provides superior SINR but has limitations due to its increased noise level effects within faulty channel conditions. MMSE exceeds the performance of Zero-Forcing through its ability to combine optimized interference suppression with reduced noise enhancement. "Successive Interference Cancellation (SIC)" gives the best performance since it provides maximum SINR values to most users which demonstrates outstanding signal detection capabilities. The results demonstrate the superiority of SIC as the best interference cancellation method suitable for MIMO MC-CDMA applications in 6G networks because of its high-performance level. Fig. 11Comparison for interference cancellation. Comparison for interference cancellation. Table 3 shows a comparison of the proposed Edge AI-enabled MIMO MC-CDMA system with the existing models on four different parameters, namely Spectral Efficiency (SE), Energy Efficiency (EE), Signal-to-Interference-plus-Noise Ratio (SINR), and Bit Error Rate (BER). The lowest SE is depicted as 12.5 bits/s/Hz with the lowest EE of 5.2 bits/Joule with an 18 dB of SINR and higher BER according to the conventional MC-CDMA and CMFB methods \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${10}^{-2}$$\end{document}, which points to the low performance of the mother under interference conditions. MIMO-OFDM is a baseline for 5G and offers high SE of 17.8 bits/s/Hz and EE of 7.4 bits/Joule with higher BER. Hybrid Precoding mmWave Massive MIMO has obtained a better performance where the system attained SE of 22.6 bits/s/Hz with EE of 9.1 bits/Joule and the BER also reduced to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${10}^{-3}$$\end{document}. Based on the Deep Learning-Based MIMO Detection the performance is further boosted with SE equal to 25.3 bits/s/Hz, EE equal to 11.2 bits/Joule and significantly reduced BE 6 × \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${10}^{-5}$$\end{document}. The systems discussed above and the designed Edge AI-enabled MIMO MC-CDMA system for 6G work for a similar number of users but the later provides the highest SE (32.7 bits/s/Hz), EE (14.8 bits/Joule), SINR (34 dB) and the least BER \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${10}^{-5}$$\end{document} ensuring that it has better efficiency, ability to handle interferences and minimizing errors. Table 3Comparison with existing models.MethodSE (bits/s/Hz)EE (bits/Joule)SINR (dB)BER (at SNR = 10 dB)Conventional MC-CDMA and CMFB2612.55.218 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${10}^{-2}$$\end{document} MIMO-OFDM (Baseline for 5G)2717.87.4223 × \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${10}^{-2}$$\end{document}Hybrid Precoding mmWave Massive MIMO2822.69.126 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${10}^{-3}$$\end{document} Deep Learning-Based MIMO Detection2925.311.2286 × \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${10}^{-5}$$\end{document}Proposed Edge AI-Enabled MIMO MC-CDMA (6G)32.714.834 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${10}^{-5}$$\end{document} Table 3Comparison with existing models.MethodSE (bits/s/Hz)EE (bits/Joule)SINR (dB)BER (at SNR = 10 dB)Conventional MC-CDMA and CMFB2612.55.218 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${10}^{-2}$$\end{document} MIMO-OFDM (Baseline for 5G)2717.87.4223 × \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${10}^{-2}$$\end{document}Hybrid Precoding mmWave Massive MIMO2822.69.126 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${10}^{-3}$$\end{document} Deep Learning-Based MIMO Detection2925.311.2286 × \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${10}^{-5}$$\end{document}Proposed Edge AI-Enabled MIMO MC-CDMA (6G)32.714.834 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${10}^{-5}$$\end{document} Comparison with existing models. The theoretical equations of SINR in equation are in line with the reported results. Spectral efficiency in Eqs. (10 and 9) spectral efficiency. The spectral efficiency improves along with the increasing SINR given by the Shannon capacity relationship as SINR increases as a result of successful SIC and DRL-based optimization. The resulting effect of this increment in the SINR is reduced Bit Error Rate (BER) as witnessed in Table 3. Thus, positive SE and negative BER improvement are directly related to the enhancement in SINR proving the correctness and validity of the proposed system model and simulation outcomes. Simulation outcomes remain consistent between the SNR range in Table 2, SINR in Fig. 11 and BER findings in Table 3. The channel conditions are the SNR range of 0–30 dB. Provided proper optimization of the SIC and the DRL, the SINR is enhanced over the SNR of the input up to a maximum of 34 dB. This improvement results in a decrease in the values of BER, since the increase in the value of SINR increases the accuracy of signal detection. Thus, the measured BER performance is in line with the enhanced SINR and the specified SNR range. Figure 12 illustrates how different communication methods perform regarding Spectral Efficiency (SE), Energy Efficiency (EE), Signal-to-Interference-plus-Noise Ratio (SINR) as well as Bit Error Rate (BER). The implementation of Edge AI functionalities in MIMO MC-CDMA (6G) results in superior metrics across Spectral Efficiency and Energy Efficiency and Signal-to-Interference-plus-Noise Ratio and Bit Error Rate. The communication method demonstrates peak SE (32.7 bits/s/Hz) along with peak EE (14.8 bits/Joule) which leads to optimal data transmission and energy optimization. A strong signal quality appears with an SINR rating of 34 dB according to this measurement. A BER measurement of 10⁻5 enables superior transmission reliability because it indicates fewer mistakes will occur. Experimental results demonstrate the superiority of proposed method by demonstrating improved operational efficiency alongside enhanced reliability and performance as opposed to current system standards. Fig. 12Performance comparison. Performance comparison. Discussions The advanced AI-based optimization techniques help the proposed Edge AI-enhanced MIMO MC-CDMA or the future generation 6G wireless communication system to outperform multiple current existing model systems. To this end, the system is designed to incorporate deep reinforcement learning for resource allocation and interference management to realize improved SE, EE, SINR, and BER. According to the simulation result, the proposed model is efficient and perform better compared to other techniques such as MIMO-OFDM, Hybrid Precoding mmWave Massive MIMO, and Deep Learning-Based MIMO Detection. The findings also demonstrate the following aspects of increased efficiency, namely, SE, EE, SINR, and BER. The proposed system gains an SE of 32.7 bits/s/Hz, this is much higher than conventional MC-CDMA and CMFB which is 12.5 bits/s/Hz and even higher than the state-of-art deep learning technique for MIMO detection which is 25.3 bits/s/Hz. Similarly, EE is enhanced to 14.8 bits/Joule, which can be attributed to better power efficiency than that of the earlier models like Hybrid Precoding mmWave Massive MIMO (9.1 bits/Joule). This clearly proves that signal quality can be enhanced through interferences through having a SINR of 34 dB; better than the conventional methods like MIMO-OFDM of 22 dB and Hybrid Precoding MIMO of 26 dB. With this the BER is equal to 10⁻5 implying better transmission reliability mostly at higher SNR. Here the DRL – based Successive Interference Cancellation (SIC) approach improves the system performance by providing dynamic allocating resources, low interference and stable system operation even in the large number of users. Another incentive of the system is the capability to provide an adaptive power control and interference management with DRL based successively interference cancellation. This intelligent approach increases SINR by a considerable extent so that the quality of signal transmission in dense network is improved. Furthermore, the design of the proposed model introduces higher spectral efficiency, it can serve more users and ensure a reliable communication at the same time. The use of deep learning in the methodology also contributes to energy efficiency since the resources are well distributed, power is not wasted, as well as enhancing the stable network. Also, the system minimizes the bit error rate (BER) and thus lowers the incidences of data transmission errors and increases the reliability of the system. As for the performance of the proposed model, it is of surprisingly low BER, which is achievable even in extreme network environment. Both features namely the Edge AI and the MIMO MC-CDMA make sure that the system to be anticipated for network requirements so that the quality is guaranteed to be stable. The results clearly show the effectiveness and applicability of AI for improving the characteristics of dispensing wireless communication systems, which makes the proposed strategy a promising solution for modern networks of high-speed communication. Overall, the Edge AI-enabled MIMO MC-CDMA (6G) system effectively addresses key challenges in next-generation wireless networks by improving spectral efficiency, optimizing energy consumption, enhancing signal quality, and ensuring reliable data transmission. The integration of AI-driven optimization and interference cancellation strategies makes this approach highly suitable for dense and high-speed communication environments, demonstrating a clear advantage over conventional and baseline 5G technologies. The effectiveness of the proposed system in terms of high-performance is demonstrated with the help of the comparative analysis of the proposed system with the current techniques like MIMO-OFDM and traditional MC-CDMA. The findings are always better in spectral efficiency, energy efficiency, SINR and BER. Such gains are attained as a result of the combination of the DRL-based resource allocation and the SIC-based interference cancellation. Moreover, the simulations are done in realistic channel conditions and averaged with numerous runs, which assures reliability. Thus, the achieved advancements are reasonable and testify to the efficiency of the suggested solution to 6G wireless networks. The spectral efficiency of 32.7 bits per second per Hertz is justified by the SE expression in Eq. (10) which is fired on the Shannon capacity formula. Based on this, SE is growing logarithmically with SINR. SIC reduces interference in the proposed system, increasing the values of SINR (approximately 34 dB). The replacement of this enhanced SINR with the SE formula yields increased data rates per unit bandwidth. Also, DRL-based resource allocation optimizes the power consumption among users, and it also provides greater spectral efficiency in a multi-user setting. Ablation analysis Ablution analysis is done to determine the value of each element in the suggested system. Tests on the performance are done by deleting major modules DRL optimization and SIC separately. The system loses its energy efficiency and resource allocation without the presence of DRL. In the absence of SIC, the interference is higher, which causes reduced SINR and high BER. The combination of Edge AI, DRL, and SIC leads to the highest performance of the full model. This discussion supports that all elements are important towards enhancing efficiency of the system. The performance is compared considering four configurations to check the effectiveness of every component. The basic communication is offered by the baseline MIMO MC-CDMA system (Eqs. (1–4)) but it has high interference. The inclusion of SIC (Eqs. (6−10) enhances SINR and minimizes BER through the cancellation of interference. Resource allocation and energy efficiency (Eqs. (11−14)) benefit with the use of DRA alone though interference is not completely eliminated. The entire proposed structure, which is a combination of MIMO MC-CDMA, SIC and DRL, offers the optimal performance with respect to spectral efficiency, energy efficiency, SINR, and BER. The comparison can be used to measure the contribution of each component in the proposed system. The initial model demonstrates poor results because of interference and lack of optimization. The former SIC adds value to signal quality (reduced interfering) and the latter DRL adds value to allocation of resources and energy-efficiency. All these are functional when combined to a large extent to enhance spectral efficiency, SINR, and BER. Therefore, the analysis of the ablation shows that every one of the modules leads to the overall enhancement of the proposed framework performance. Conclusion and future work This paper provides the detail investigation on the Edge AI-based MIMO MC-CDMA (6G) system and verifies that it performs better than other systems in the contemporary wireless networks. When applying DRL for the resource allocation and interference management proposed system enhances various main performance indicators: SE, EE, SINR and BER. It was found out that the solutions exceptions of the proposed approach of arriving at faster data rates, lower error levels of transmission, and better power control than other methods like the MIMO-OFDM and the hybrid precoding. In applying SIC with the use of DRL, the adaptability of the system intensifies and improves the functionality under high user loads. The research discusses possibilities of the role of next generation communication architectures governed by artificial intelligence in higher generation networks including 6G to address issues relating to spectrum use and management, interference and energy consumption. In so doing the research confirms that with the use of adaptive resource management system with assist of artificial intelligence it is possible to improve networks’ performance with least computational cost. This kind of demand of ultra-reliable and high-capacity communication have made this research work as a strong background for the advancement of intelligent, scalable and sustainable wireless communication. For the future work, the following solutions can be proposed to improve the system and its productivity in terms of robustness and expansibility. First, the use of adjustment approaches like federated learning and transfer learning cohesively can enhance logicality to networks while enabling real-time decision-making. Further, if more research was conducted on channel models, for example the Rician and mmWave fading models the behaviours of the system under different propagation environments would be ascertained. One of the other areas to consider is the collaboration of edge devices where the optimal allocation of resources for AI can be defined with regard to various base stations in case of ultra-dense networks. Finally, this model can be expanded to multi-access edge computing (MEC) and intelligent reflecting surfaces (IRS) to find more energy-saving solutions and spectral efficiency in future 6G smart spaces.