We present a novel spectral decomposition approach that combines Physics-Informed Neural Networks (PINNs) with the Lattice Boltzmann Method (LBM) to address fundamental spectral bias limitations in turbulent flow simulation. By partitioning the solution into low-frequency large-scale structures handled by LBM and high-frequency residuals learned by PINNs from DNS training data, our method achieves accurate reconstruction of the full turbulent energy spectrum while maintaining computational tractability. This approach represents a hybrid reconstruction/super-resolution technique that leverages high-fidelity DNS data for PINN training. The method’s capability to capture both large-scale flow patterns and high-frequency fluid dynamics makes it particularly valuable for critical applications in medical and pharmaceutical sectors, including drug delivery optimization, protein stability analysis during injection, and tablet dissolution modeling. Experimental validation on canonical turbulence problems demonstrates 5 \(\times \) improvement in Kolmogorov \(k^{-5/3}\) scaling accuracy, 83% better energy conservation compared to LBM-only methods, and 11.6 \(\times \) improvement in velocity profile accuracy. This breakthrough demonstrates the viability of hybrid numerical-ML simulations for high Reynolds number turbulent flows and establishes a new paradigm for multi-scale fluid modeling in biomedical applications.