Modeling and Evaluation of Vertical Strain at the Top of the Subgrade in Airfield Pavements Using Neural Networks

The application of three-dimensional finite-element models in airfield pavement design provides high accuracy in predicting structural responses, allowing the simulation of loading conditions and interactions between pavement layers. However, this process requires considerable computational time compared to the layered elastic design (LED) method, which limits its use in routine design analyses. The objective of this study was to develop and evaluate a predictive model based on artificial neural networks (ANNs) capable of estimating vertical strain at the top of the subgrade in airfield pavements. To achieve this goal, the network was trained using a database generated through three-dimensional simulations performed in ABAQUS software, considering five aircraft with single-axle dual-wheel configurations, totaling 5,000 simulations. The ANN architecture was defined with two hidden layers, containing 10 and 8 neurons, using the hyperbolic tangent activation function. The model’s performance was evaluated using the mean absolute error and R2 metrics, and the results were compared with values calculated by ABAQUS and multilayer elastic analysis, which applies the LED method. Additionally, a sensitivity analysis was carried out using elasticity coefficients, enabling the identification of the most influential structural variables affecting pavement response. The results showed that the ANN achieved high predictive accuracy and presented a shorter inference time compared to the finite-element method, demonstrating its potential for large-scale applications. It is concluded that the use of ANNs trained with robust data can represent an efficient alternative for the structural analysis of airfield pavements, supporting the optimization of the design process and aiding decision-making in pavement projects and maintenance planning.