Abstarct: Quantitative structure–property relationship (QSPR) is an efficient approach for predicting the properties of various compounds baxsed on their molecular structures. In this study, QSPR models were developed to predict the retention time of a set of volatile organic compounds (VOCs) present in tea leaves using molecular descxriptors. To this end, a comprehensive set of descxriptors was calculated and subsequently reduced by applying stepwise regression (SW), random forest (RF), and particle swarm optimization (PSO) as variable selection methods. Artificial neural networks (ANNs) were employed to model potential nonlinear relationships between the molecular descxriptors and retention time. In order to construct an optimal neural network model with adequate performance, all influential network parameters were optimized and trained. Model performance was evaluated using various approaches, including statistical parameters, leave-one-out validation, applicability domain analysis, residual distribution, Y-randomization, and external validation. The results demonstrated that the RF–ANN models exhibited relatively superior performance compared to the other models, indicating strong predictive capability, although the other models also produced acceptable results. Nevertheless, the SW–ANN model achieved a comparable level of performance using a smaller number of descxriptors, leading to a simpler and computationally more efficient model. The findings of this study confirm the effectiveness of ANN-baxsed QSPR models and demonstrate that stepwise regression can serve as an effective strategy for descxriptor selection without compromising the predictive ability of the model.