APPLICATION DES RESEAUX DE NEURONES POUR LA PREDICTION DES SOLS GONFLANTS

Abstract

Swelling soils, particularly those rich in smectite-type clays, represent a critical geotechnical issue in many regions around the world. Their ability to undergo significant volume changes due to variations in water content poses a serious threat to the stability of infrastructures, including building cracks, pavement heaving, and disruptions to underground networks. Traditionally, the determination of key parameters—namely, swelling pressure and swelling potential relies on experimental tests (oedometer, free swell, double oedometer, etc.) that are time consuming, costly, and technically demanding. Moreover, these methods require specialized equipment and strict testing conditions to ensure reliable results. This study proposes an innovative approach to predicting soil swelling parameters using artificial neural networks (ANN). Based on a rich and diverse experimental database (438 samples for pressure, 291 for potential), two predictive models were developed. The results demonstrate that the use of artificial neural networks offers a reliable, fast, and cost- effective alternative to laboratory testing for forecasting the swelling behavior of soils. The developed models successfully captured the nonlinear and multi-parameter nature of the swelling phenomenon with satisfactory accuracy, particularly for swelling pressure. This approach thus opens new perspectives for broader integration of artificial intelligence in geotechnical studies.