Deterministic Lateral Displacement (DLD) devices are vital tools in microfluidics, enabling size-based, label-free separation of cells and particles. These devices play an essential role in cancer diagnostics by effectively isolating circulating tumor cells (CTCs) from blood samples. However, traditional methods used to evaluate and optimize DLD devices, such as computational fluid dynamics (CFD) simulations, are often costly, complex, and very time-consuming. While machine learning (ML) methods, particularly deep learning, offer potential improvements, current models typically require extensive modifications to physical datasets and domain restructuring, limiting their accuracy and ability to generalize to new scenarios. This research introduces an advanced Physics-Informed Deep Neural Network (PIDNN) that significantly enhances the prediction of velocity fields within DLD devices. Unlike conventional ML methods, PIDNN uniquely integrates essential physics principles directly into its architecture, enforcing critical boundary conditions and initial condition. This integration ensures physically accurate predictions, substantially improving the reliability of the model. The PIDNN is trained using detailed velocity field data generated by COMSOL Multiphysics simulations. Model inputs include critical parameters such as non-dimensional diameter (F), period number (N), Reynolds number (Re), and spatial coordinates. Furthermore, an innovative data sampling technique is introduced to enhance data density near crucial device boundaries, effectively capturing essential flow features. The application of PIDNN drastically reduces the time needed for evaluating device performance, enabling rapid selection of optimal design parameters that greatly enhance the effectiveness of particle and cell separation. When combined with a particle trajectory solver, the PIDNN can accurately predict particle trajectories and critical non-dimensional diameters (Dc) with average error less than 5% , essential for efficient cell separation. This innovative approach streamlines the development of versatile, high-performance DLD devices, significantly advancing their practical applications in cancer diagnostics.