Flexible Pavement Analysis Using Physics-Informed Machine Learning Methods
Advisor: Professor Imad Al-Qadi
Zoom Link: https://illinois.zoom.us/j/81060176584?pwd=UjhmZnpMMk5CbkdIem8vSFVTRjQxZz09 (Meeting ID: 810 6017 6584, Password: 578613)
According to US Transportation Statistics in 2023, over 3.2 trillion vehicle miles were traveled in 2021, while a concerning 43% of public roadways exhibited poor to mediocre conditions. This decline in roadway quality translates to a staggering financial toll on motorists, approximating an annual cost of $130 billion in vehicle repairs and associated operational expenses. To enhance time and cost efficiencies, the engineering domain has witnessed an escalating emphasis on advanced computational techniques in pavement design and management. A primary challenge is the time and computer resources required to simulate pavement structure responses using three-dimensional (3D) finite element (FE) analyses. This study provides an accurate and efficient approach to predict pavement response using graph neural network (GNN) and optimal model to predict IRI over the pavement service life, using ANN, incorporating maintenance activities.
This dissertation introduced the use of GNN to predict pavement responses utilizing an extensive FE analysis dataset. A database was created as a central resource for this study, containing 240 FE cases of the 819 3D FE simulations of flexible highway pavements conducted over two decades. The GNN was used to build pavement FE nodes into graph nodes, embedding physics-based interaction between nodes within its machine learning (ML) architecture. The dynamic behaviors of pavement FEs were computed via learned message-passing between two graphs within two continuous timesteps. This model achieved a one-step mean square error (MSE) as minuscule as 1.21×10−8 and a rollout MSE of approximately 4.55×10−5. This prediction framework, astoundingly efficient, requires a mere 5 minutes, presenting a stark contrast to traditional 3D FE analyses that could span hours to weeks. A systematic exploration of hyperparameters accentuated the optimal configuration of 10 message-passing steps (M) and a single historical time step (C), balanced adeptly between model performance and computational expediency. Additionally, preliminary data normalization surfaced as a pivotal step, considerably attenuating simulation noise.
The pre-trained GNN model, which was originally developed for highway pavement analysis, was adapted for airfield pavement structural responses. Using 11 3D FE simulation cases,
226 individual steps were utilized in analysis. Utilizing the strategies model scaling and graph
pooling, satisfactory mean squared errors (MSE) of 1.24 × 10−7 and 2.26 × 10−7 were resulted
in one-step predictions, respectively. It was found that the model scaling might be susceptible to
overfitting in the base layer, while graph pooling might introduce higher errors in the subbase and
This dissertation also delved into the domain of pavement management, particularly the
International Roughness Index (IRI) prediction (IRI progression over the pavement’s service life
without maintenance/rehabilitation and the drop in IRI after maintenance models). The first model
utilizes the recurrent neural network (RNN) algorithm using data extracted from the Long-Term
Pavement Performance (LTPP) database. A RNN based long short-term memory network (LSTM)
was used to solve the vanishing gradient problem. The second model utilized an artificial neural
network (ANN) algorithm to correlate the impacting factors to the IRI value after maintenance.
Combining the two models allowed for the prediction of IRI values over AC pavement’s service