<p>Traditional seismic inversion is expensive and relies heavily on unstable stochastic methods. Deep learning (DL) is a cost-effective method that excels in extracting intricate features from data and making precise predictions. However, while challenges such as the lack of interpretability, scarcity of borehole label data, and insufficiency of synthetic data to generate scenarios consistent with actual geology persist, this study demonstrates how we can improve the robustness and applicability of inversion solutions. We propose a hybrid method that leverages physical constraints provided by a rock-physics equation while harnessing the efficiency of DL to effectively approximate inversion solutions. We used a porosity model along an interpreted seismic horizon based on well logs to simulate the reservoir’s bulk and shear moduli with different saturating fluids. This model provided a priori knowledge of the modelling process to strengthen stability and consistency with local geologic conditions. We used the obtained moduli in empirical equations to compute the saturated reservoir seismic wave velocities. This provided the necessary labels for training four baseline networks with different topologies and hyperparameters to invert seismic amplitude data into Vp/Vs, Poisson’s ratio, and acoustic impedance. Subsequently, we weighted individual baseline model predictions based on their accuracies and aggregated them to achieve robust models. The proposed method demonstrated superior performance compared with conventional inversion and popular DL approaches in a real-world application. The study results are critical for fluid discrimination and understanding the reservoir’s potential for oil and gas production.</p>