Load-deformation prediction of bored piles using sequential soil profile encoding with transformer architecture: A study of Bangkok subsoil
Journal article
Authors/Editors
Strategic Research Themes
Publication Details
Author list: Youwai S.; Thongnoo C.
Publisher: Elsevier
Publication year: 2025
Volume number: 275
ISSN: 0957-4174
eISSN: 1873-6793
Languages: English-Great Britain (EN-GB)
Abstract
This paper introduces a transformer-based framework for predicting the load-deformation behavior of bored piles in Bangkok subsoil by treating soil profiles as sequential data. The proposed model innovatively encodes each soil layer and its properties as tokens, analogous to words in natural language processing, creating a continuous representation of the subsurface conditions. The model architecture consists of an encoder that processes soil profile sequences and a decoder that generates load predictions, incorporating pile features and previous deformation history. Soil properties, including Standard Penetration Test (SPT) values and soil types, are tokenized and processed through self-attention mechanisms to capture layer-wise relationships. The model is trained on static pile load tests from Bangkok, achieving a mean absolute percentage error of 4.08% on test data. Validation results demonstrate the model's capability to accurately predict load-deformation curves for different pile dimensions and soil conditions. The framework effectively handles the complex stratigraphy of Bangkok subsoil, characterized by alternating clay and sand layers. The trained model is publicly available and can be fine-tuned for different geotechnical conditions. This research represents the first application of sequential encoding and transformer architecture to pile foundation analysis, offering a new approach to geotechnical predictions that better captures the natural sequencing of soil layers and load-deformation behavior. © 2025 Elsevier Ltd
Keywords
No matching items found.
Contact Information