Uncertainty quantification of compressive stress response in expanded polystyrene foams using evidential neural networks

Abstract

This study investigates the application of Deep Evidential Regression in shallow feed-forward neural networks to model and quantify the compressive stress response of expanded polystyrene foam. This foam material, widely utilized for impact protection and packaging, exhibits distinct mechanical behavior characterized by elasticity, plateau, and densification stages during compressive loading. This research adopts a data-driven approach, leveraging artificial neural networks enhanced with evidential learning to predict the distribution of stress responses, thereby addressing both aleatoric and epistemic uncertainties. The methodology involves organizing stress-strain data into training, validation, and test sets, adding noise to simulate real-world conditions, and training models with evidential layers. Results demonstrate that the proposed models maintain high predictive accuracy, with coefficients of determination exceeding 0.90 for noisy test data and above 0.99 for noise-free data. The evidential regression models also provide robust uncertainty quantification, essential for applications where data quality varies. This study’s findings highlight the efficiency and effectiveness of Deep Evidential Regression in enhancing the reliability of stress-strain predictions for EPS foam, offering significant potential for broader application to similar foam materials.