In many engineering applications, the use of Fiber-Reinforced Polymer (FRP) composite materials has increased due to their desirable properties. However, conventional theories are not sufficiently accurate in predicting the complex failure of these materials, as displayed by the World-Wide Failure Exercises (WWFEs) I and II. The development of data-driven models using Deep Neural Networks (DNNs) has shown promising results in predicting the failure strength of composite materials and has proven to be more accurate and time-efficient than conventional theories. A previous study by Fontes and Shadmehri (2023) used experimental failure data of laminates under biaxial loadings from WWFE-I to develop a data-driven failure model for FRP composite materials using a DNN framework. While this model effectively predicted failure envelopes consistent with the experimental data with an acceptable degree of accuracy, it had difficulty predicting the transition region from uniaxial to biaxial loading, and some sections of the envelope were non-convex. The present study aimed to improve the model by introducing novel inputs for loading stress ratio variables in the axial, circumferential, and shear directions. This new feature decreased the mean squared error, rectified the transition region issue, and enhanced the overall failure envelope predictions.