Intravoxel Incoherent Motion (IVIM) is a diffusion-weighted Magnetic Resonance (DW-MR) technique used to quantify tissue perfusion and diffusion properties. Traditional methods for IVIM fitting often suffer from sensitivity to noise and computational inefficiency, thus limiting the utility of IVIM in clinical settings. Recently, Deep Neural Networks (DNNs) have shown promise for estimating IVIM maps with improved accuracy and robustness. This study investigates a physics-informed convolutional neural network (CNN) ensemble for IVIM parameter estimation, integrated with the uncertainty quantification of the generated maps. A physics-informed U-Net was trained five times on 4200 simulated Shepp-Logan phantoms under varying Signal-to-Noise Ratio (SNR) conditions (25, 50, and 100). Accuracy was assessed on a simulated test set using the Median Absolute Error (MedAE), while variability in homogeneous regions was quantified using the Robust Coefficient of Variation (RCV). Additionally, an in-vivo test set of DW-MR mouse brain images was used to compare the CNN’s parameter estimation with the Bayesian approach. Predictions from the five trained models were aggregated using a deep-ensemble approach, with the mean prediction across the ensemble providing improved accuracy and robustness on the simulated test set. This also enabled to generate variance maps highlighting regions with higher uncertainty, providing a qualitative assessment of parameters reliability. The in-vivo mean parameters estimates were in broad-agreement with the Bayesian estimates, especially for D and D*, despite the CNN being trained only on simulations. In conclusion, the proposed physics-informed CNN ensemble shows potential to enhance IVIM parameter estimation, enabling uncertainty quantification and improving robustness.