{"title":"Multifidelity-based Gaussian process for quasi-site-specific probabilistic prediction of soil properties","authors":"Geng-Fu He, Pin Zhang, Zhen-Yu Yin, Siang Huat Goh","doi":"10.1139/cgj-2023-0641","DOIUrl":null,"url":null,"abstract":"Conventional empirical equations for soil properties prediction tend to be site-specific, exhibiting poor reliability and accuracy. Meanwhile, alternative data-driven methods require large datasets for training. To address these issues, this study proposed a novel multifidelity residual neural-network-based Gaussian process (MR-NNGP) modelling framework. A soil property low-fidelity (LF) prediction model is first trained using abundant LF data collected from worldwide sites for generating preliminary estimation. A high-fidelity (HF) model is subsequently trained on sparse HF data from the specific site of interest for calibrating the LF model to make quasi-site-specific predictions. An infinitely wide NN-inspired NNGP is adopted as the baseline algorithm for training LF and HF models. The compression index of clays is selected as an example to examine the capability of the proposed MR-NNGP. The results indicate that the compression index of clays can be well captured by MR-NNGP, exhibiting superior accuracy and reliability compared with one-shot training without using MR modelling and other baseline algorithms such as GP. The MR-NNGP framework alleviates data dependency and improves model performance through hierarchical modelling on relatively simple correlations using a superior algorithm. Unified LF data and efficient hyper-parameter optimization indicate the flexibility for broader applications in various sites worldwide.","PeriodicalId":9382,"journal":{"name":"Canadian Geotechnical Journal","volume":null,"pages":null},"PeriodicalIF":3.0000,"publicationDate":"2024-05-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Canadian Geotechnical Journal","FirstCategoryId":"89","ListUrlMain":"https://doi.org/10.1139/cgj-2023-0641","RegionNum":3,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"ENGINEERING, GEOLOGICAL","Score":null,"Total":0}
引用次数: 0
Abstract
Conventional empirical equations for soil properties prediction tend to be site-specific, exhibiting poor reliability and accuracy. Meanwhile, alternative data-driven methods require large datasets for training. To address these issues, this study proposed a novel multifidelity residual neural-network-based Gaussian process (MR-NNGP) modelling framework. A soil property low-fidelity (LF) prediction model is first trained using abundant LF data collected from worldwide sites for generating preliminary estimation. A high-fidelity (HF) model is subsequently trained on sparse HF data from the specific site of interest for calibrating the LF model to make quasi-site-specific predictions. An infinitely wide NN-inspired NNGP is adopted as the baseline algorithm for training LF and HF models. The compression index of clays is selected as an example to examine the capability of the proposed MR-NNGP. The results indicate that the compression index of clays can be well captured by MR-NNGP, exhibiting superior accuracy and reliability compared with one-shot training without using MR modelling and other baseline algorithms such as GP. The MR-NNGP framework alleviates data dependency and improves model performance through hierarchical modelling on relatively simple correlations using a superior algorithm. Unified LF data and efficient hyper-parameter optimization indicate the flexibility for broader applications in various sites worldwide.
期刊介绍:
The Canadian Geotechnical Journal features articles, notes, reviews, and discussions related to new developments in geotechnical and geoenvironmental engineering, and applied sciences. The topics of papers written by researchers and engineers/scientists active in industry include soil and rock mechanics, material properties and fundamental behaviour, site characterization, foundations, excavations, tunnels, dams and embankments, slopes, landslides, geological and rock engineering, ground improvement, hydrogeology and contaminant hydrogeology, geochemistry, waste management, geosynthetics, offshore engineering, ice, frozen ground and northern engineering, risk and reliability applications, and physical and numerical modelling.
Contributions that have practical relevance are preferred, including case records. Purely theoretical contributions are not generally published unless they are on a topic of special interest (like unsaturated soil mechanics or cold regions geotechnics) or they have direct practical value.