Banca de DEFESA: Anderson Gustavo Villamil Gonzalez

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : Anderson Gustavo Villamil Gonzalez
DATE: 03/08/2023
TIME: 14:00
LOCAL: on line (teams)
TITLE:

EVALUATION OF ARTIFICIAL NEURAL NETWORK TO PREDICT SAND-GEOMEMBRANE INTERFACE SHEAR STRENGTH

KEY WORDS:

Artificial Neural Network, Geomembrane, Interface shear strength, Multilayer Perceptron

PAGES: 97
BIG AREA: Engenharias
AREA: Engenharia Civil
SUBÁREA: Geotécnica
SUMMARY:

The implementation of waterproofing systems during the construction of dams, landfills, and artificial channels is crucial to avoid fluid infiltration into the foundation soil, which can lead to structural damage and other potential hazards. Geomembranes are widely used in civil engineering as they offer excellent waterproofing capabilities, low permeability, and favourable mechanical properties. Like other construction materials, evaluating the strength at the interface between geosynthetics and in-contact material is necessary to ensure enough resistance to potential failures. Various laboratory tests can be conducted to ensure adequate interface resistance, such as direct shear, ring shear, and inclined plane. These tests determine the friction angle, a critical factor in determining the interface shear strength between granular soil and geosynthetics. However, these tests can be time-consuming and expensive and may only sometimes be feasible in project planning. Therefore, finding alternative methods to obtain the necessary information is essential. One possible solution is to use reference results from other research. In this way, a database of previous results can be compiled, and a predictive model can be created to estimate the required interface strength values. This study aims to assess the effectiveness of using an Artificial Neural Network (ANN) methodology to predict the shear strength at the interface of sand and geomembrane. A Multi-Layer Perceptron (MLP) architecture was chosen to configure the ANN models, and the training process is a supervised one that involves a Back-Propagation (BP) training algorithm coupled with the Differential Evolution (DE) optimization algorithm. The input data for the models were defined from 428 laboratory tests reported in previous investigations, including 14 input parameters and sand-geomembrane interface strength results. Four ANN models were analysed and compared, differentiated in terms of their number of inputs (9 or 14) and the number of hidden layers (1 or 2). The ANN model with the architecture 14-71-342-1 displayed the most satisfactory results for the training and testing phase in terms of the predicted values' distribution compared to the trend line (R²: 0.919 training, R²: 0.852 testing), a lower number of residual values outside the acceptable range (4% training, 11.6% testing), and excellent prediction performance according to statistical metrics for both phases (RMSE: 1.92, MAE: 1.32, MAPE: 5.03% training, RMSE: 0.852, MAE: 3.36, MAPE: 7.13% training). Based on the results, the ANN technique can be defined as an effective approach for predicting sand-geomembrane interface strength values (friction angle) for the collected data.

COMMITTEE MEMBERS:
Externo à Instituição - PAULO CÉSAR LODI - UNESP
Interno - 404375 - ENNIO MARQUES PALMEIRA
Presidente - 2517970 - GREGORIO LUIS SILVA ARAUJO
Interno - 1549489 - MANOEL PORFIRIO CORDAO NETO