Deep clustering has emerged as an effective paradigm that unifies representation learning and cluster assignment within a single optimization framework. Despite significant progress, existing deep clustering models often rely on highly non-convex formulations, making them sensitive to initialization and prone to suboptimal convergence. These limitations motivate the need for principled approaches that combine the expressive power of deep learning with the stability and interpretability of convex optimization. This paper presents a Hybrid Optimization Framework for Deep Clustering that in tegrates convex relaxation with graph-based regularization in a unified deep architecture. The proposed model formulates clustering as a bi-level optimization problem, where a convex surrogate for the affinity matrix is jointly optimized with a Laplacian regularizer to enforce manifold smoothness and geometric consistency in the latent space. This hybrid design stabilizes training dynamics and enhances robustness against non-convexity. Theoretical analysis establishes convergence to a stationary point and clarifies the role of the convex subproblem in reducing sensitivity to poor initialization. Extensive experiments on MNIST, Fashion-MNIST, COIL-20, and USPS demonstrate that the proposed method consis tently outperforms state-of-the-art baselines in accuracy, normalized mutual information, and adjusted Rand index. Additionally, the framework exhibits efficient and monotonic convergence behavior while maintaining competitive computational complexity.