Brittle fracture modeling in linear elastic fracture mechanics (LEFM) remains challenging because crack-tip stress singularities and steep gradients require high resolution, while conventional finite element method (FEM) often depend on costly local mesh refinement or tip-enrichment techniques such as the extended finite element method (XFEM). Physics-informed neural networks (PINNs) [1] have emerged as an alternative for solving partial differential equations (PDEs) by incorporating PDE residuals into the loss function of a neural network. Their meshfree formulation offers flexibility in discretization. However, PINNs still struggle with localized features such as cracks, since fully connected networks have limited ability to capture singular components, including stress concentrations or steep gradients near crack tips. To address this issue, this research proposes a selective mode-based tip-enrichment strategy for PINNs to improve robustness and accuracy in LEFM simulations, building on earlier work by Gu et al. [2]. In this approach, the feed-forward displacement field is expressed as a superposition of a smooth neural-network component capturing the global response and an analytically enriched crack-tip component capturing the dominant localized (singular) behavior. Furthermore, to ensure the enrichment adapts naturally to different loading conditions, including mixed-mode cases, a Gumbel-Softmax reparameterization [3] is introduced to enable automatic mode identification without manual switching by the user. The proposed framework is evaluated on benchmark LEFM problems, and the predicted results show strong agreement with reference solutions. The approach also outperforms the earlier formulations, including conventional PINN, demonstrating improved robustness in crack-tip characterization and providing an efficient, scalable machine-learning-in-mechanics approach for fracture analysis.