<p>A deep unfolding neural network, termed IRLS-NET, is proposed for single-snapshot direction-of-arrival (DOA) estimation within sparse array systems. The framework is designed to address the challenges of manual parameter tuning and high computational overhead inherent in traditional sparse recovery algorithms. Under the compressive sensing framework, the DOA estimation problem is reformulated as a sparse signal recovery task, and the iterative reweighted least squares (IRLS) algorithm is subsequently unfolded into a layer-wise network architecture. Within this architecture, the algorithm’s hyperparameters, such as regularization coefficients, are recast as learnable parameters and are optimized end-to-end via backpropagation. This methodology preserves the inherent structure and interpretability of the original IRLS algorithm while automating the parameter tuning process. Through comprehensive simulations, the estimation accuracy of IRLS-NET is shown to approach the Cramér-Rao Lower Bound and is comparable to that of CVX-based methods, particularly in the context of resolving closely spaced sources. Furthermore, a significant super-resolution capability is demonstrated, wherein the network successfully resolves a number of sources exceeding the number of physical sensors in underdetermined scenarios. The effectiveness and robustness of the algorithm for practical single-snapshot DOA estimation are further confirmed through experimental validation using measured data from a millimeter-wave radar. A key finding is that the required number of iterations and the computational runtime are substantially reduced compared to conventional iterative methods, positioning IRLS-NET as a robust and efficient solution for real-time applications.</p>