Developing New Empirical Formulae for the Resilient Modulus of Fine-Grained Subgrade Soils Using a Large Long-Term Pavement Performance Dataset and Artificial Neural Network Approach

Abstract

Artificial neural network (ANN) has been successfully used for developing prediction models for resilient modulus (Mr). However, no reliable Mr formula derived from these models has been proposed in previous studies, although engineers/researchers need empirical formulae for hand calculation of Mr. Therefore, this study aimed to propose reliable empirical formulae for the Mr of fine-grained soils using ANN. For this purpose, thousands of ANN models were developed using the long-term pavement performance (LTPP) and external datasets. The input parameters were the percentage of soil particles passing through #200 sieve (P200), silt percentage (SP), clay percentage (CP), liquid limit (LL), plasticity index (PI), maximum dry density ([ρdry]max), optimum moisture content (wopt), confining pressure (σc), and nominal maximum axial stress (σz). The ANN models were compared with several constitutive models. The results indicate that the constitutive models failed to predict the Mr, and the best Mr predictions were obtained by the ANN-C9 (P200, SP, CP, LL, PI, σc, and σz), ANN-C10 (P200, SP, CP, [ρdry]max, wopt, σc, and σz), and ANN-C11 (P200, SP, CP, LL, PI, [ρdry]max, wopt, σc, and σz) models. Thus, the structures of these ANN models were formulated and proposed as the new empirical formulae for the Mr of fine-grained soils. Sensitivity analysis was also performed on these ANN models. It was determined that (ρdry)max is the most influential parameter in the ANN-C10 model, and LL is the most influential parameter in the ANN-C9 and ANN-C11 models. On the other hand, σc and σz are the least influential parameters.