Wear at the wheel--rail interface remains a persistent challenge in railway engineering, affecting operational safety, maintenance planning, and system reliability. Although finite element simulations are widely used to analyze stress distribution and material deformation, accurately predicting wear and rolling contact fatigue (RCF) remains difficult due to the nonlinear nature of the phenomenon and the interaction between geometry, loading conditions, and material response. Consequently, recent studies have explored data-driven and hybrid strategies to complement traditional FE analyses and improve predictive capability [1,2]. According to [3], RCF tends to occur in non-conformal contacts, whereas conformal or closely conformal contacts are typically associated with wear-dominated behavior. These authors also introduced the maximum separation parameter (s), defined as the distance between the rail and wheel surfaces along the rail centroid after contact, as a measure of contact conformity. Furthermore, [4] proposed the severity index I_{wr}, which incorporates conformity and contact length to characterize the overall intensity of the wheel--rail interaction. In this work, we propose an Artificial Intelligence methodology using Convolutional Neural Networks (CNNs) to classify both the conformity (closely conformal, conformal, non-conformal) and severity (severe, non-severe) of the wheel--rail contact. The model receives measured wheel and rail profiles as input and directly predicts the conformity level (s < 0.1, 0.1 < s < 0.4, s > 0.4) and severity (I_{wr} < or > 0.35). The database was generated using a validated two-dimensional quasi-static finite element framework implemented in Abaqus, supported by a Python routine capable of running large numbers of simulations over diverse geometric configurations. After training and validation, the AI achieved an accuracy of 88.2%, an AUC-PR of 91.9%, and an F1-score of 86.2%. These results indicate that the proposed workflow allows direct assessment of wheel--rail contact conditions without extensive finite element simulations, providing a reliable alternative for large-scale or real-time evaluation.