A holistic assessment of the subsurface is achieved through an integration of subsurface data in terms of seismic and wells, using a geological foundation. A robust reservoir characterisation and a predictive subsurface model are of paramount importance in various phases of hydrocarbon exploration and production. Often, these processes are time-consuming as they involve seismic inverse modelling followed by reservoir or property modelling to derive the elastic and reservoir properties of the subsurface. Compared to the development phase, the exploration stage has its limitations due to sparce data, making it riskier and rife with uncertainties. Moreover, characterising a Carbonate reservoir has been a challenge due to its enigmatic behaviour owing to the multiple diagenetic changes it has undergone over millions of years. Recent advances in Deep Learning Methods such as Convolutional Neural Networks (CNNs) have encountered success in various industries and are revolutionising many fields. To train CNN, a huge amount of labelled data is required, which is a limitation in the exploration stage due to a limited number of wells. In this study, to address this limitation, rock physics-guided deep learning is implemented to predict a subsurface model for Eocene Carbonates of Western Offshore Basin, India. A clay-rich-carbonate rock physics model is established for the study area using a drilled well. This model is subsequently incorporated during the simulation of various geological possibilities such as variations in reservoir properties like volume of clay and effective porosity and reservoir thickness. Each of these various scenarios served as a synthetic well. To create synthetic seismic gathers, a real seismic wavelet is convolved with the angle-dependent reflectivity of the elastic curves generated from the synthetic wells. These seismic gathers, representing each scenario, serve as the labelled data which were required to train the CNN. Transfer Learning is also tested and thereafter implemented to make more efficient predictions before the validation and production with the real seismic dataset. The comparison of Modelled properties and the original log properties depicts that Rock Physics Guided Deep Learning has a better correlation coefficient of the properties than the seismic inverse modelling methods, thus saving time, enhancing efficiency and minimising the risk of hydrocarbon exploration.