The increasing demand for massive connectivity, high data rates, efficient spectrum utilization, and continuous power supply to sensors in next-generation wireless networks presents significant challenges, including bandwidth scarcity and energy constraints. To address these issues, we consider a backscatter-enabled cognitive full-duplex (FD) wireless powered communication network enhanced with intelligent reflecting surfaces (IRS) and cluster non-orthogonal multiple access (NOMA), serving as a secondary network (SN) coexisting with a primary network (PN). In this setup, SN sensors access the spectrum without interfering with PN operations and operate in FD mode to nearly double the spectrum efficiency compared to half-duplex transmission. Backscatter communication allows sensors with insufficient stored energy to transmit data without significant delays, supporting real-time applications. To enhance connectivity, mitigate deep fading effects, and provide massive connectivity, we deploy IRS and NOMA techniques. However, increasing the number of sensors raises the complexity of successive interference cancellation (SIC) in NOMA systems. To alleviate this, we employ cluster NOMA, grouping sensors into clusters to reduce SIC complexity via parallel processing. The joint optimization of hybrid access point beamforming, IRS phase shifts, which lie on a unit modulus nonconvex set, and sensor uplink power allocation, under practical constraints such as nonlinear energy harvesting, imperfect channel state information, self-interference due to FD transmission, and limited buffer capacity, results in a complex, high-dimensional, and NP-hard problem. In addition, simultaneous energy harvesting and data transmission in FD mode, together with the energy harvested in one frame that is used in subsequent frames, make the optimization problem frame dependent. To efficiently solve this ergodic problem, we propose a deep reinforcement learning framework based on the twin delayed deep deterministic policy gradient (TD3) algorithm. TD3 effectively handles high-dimensional decision variables and non-convex optimization problems inherent in this system design. Our results demonstrate that TD3 nearly doubles the system’s ergodic sum-rate compared to random allocation and achieves performance comparable to particle swarm optimization but with significantly lower computational complexity and overhead, making it more suitable for large-scale networks. In conclusion, our TD3-based framework provides a scalable and robust solution for complex wireless network architecture, maximizing performance while minimizing computational load. This work advances the state-of-the-art by offering a practical solution for dense networks in next generation networks, addressing key challenges in energy efficiency, fairness, and spectral efficiency.