TY  - SER
AU  - Napattarapong Kaenpuek
AU  - Chao,Kuo Chieh
AU  - Avirut Puttiwongrak,
ED  - Royal Thai Government Fellowship,
TI  - Evaluation of factors influencing soil-water infiltration behavior using machine learning approach
T2  - Thesis
PY  - 2025///
CY  - Pathum Thani, Thailand
PB  - Asian Institute of Technology
KW  - Landslide hazard analysis
KW  - Data processing
KW  - Soil science
KW  - Machine learning
N1  - A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Geotechnical and Earth Resources Engineering; Thesis (M. Eng.) - Asian Institute of Technology, 2025
N2  - Landslides pose significant threats to lives, infrastructure, and the environment, often triggered by rainfall-induced soil infiltration that reduces slope stability. Accurate  prediction of soil-water infiltration under varying rainfall and slopeconditions is crucial  for landslide risk mitigation. Traditional deterministic models, while physically  rigorous, require detailed input parameters that are difficult to obtain in practice,  whereas purely data-driven approaches may lack physical consistency. This study  develops and evaluates a Physics-Informed Neural Network (PINN) framework that  integrates the governing Richards{u2019} equation with observational data to predict soil  moisture profiles in unsaturated slopes subjected to rainfall. Laboratory experiments  were conducted using a physical slope model filled with red clayey sand and white  sandy clay under controlled rainfall intensities (10, 50, and 90{u202F}mm/h) and slope angles  (0°, 15°, and 30°). Soil properties, including soil-water characteristic curves, were  determined experimentally, and Finite Element Method (FEM) simulations were  calibrated against observed data. The PINN was first trained using observational data  only and then extended to a hybrid strategy combining FEM and observations. Results  show that the observation-only PINN captured the general infiltration trend but  produced smoother, less accurate wetting fronts, particularly at greater depths. In  contrast, the hybrid-trained PINN improved accuracy, capturing infiltration dynamics  and wetting front propagation more effectively, with lower error metrics and better  alignment with observations and FEM results. These findings indicate that, with denser  sensor networks in the future, PINNs trained on richer observational data could further enhance prediction performance, supporting real-time landslide hazard assessment
UR  - http://203.159.5.9/ait-thesis/detail.php?q=B23569
ER  -