Knowledge-informed data-efficient characterisation of soil spatial variability

doi:10.1016/j.compgeo.2026.108403

Return

Knowledge-informed data-efficient characterisation of soil spatial variability

2026-07-07

PRE

Geng‐Fu He

Pin Zhang *

DOI：10.1016/j.compgeo.2026.108403

Original

Original request for help

Save

Abstract

En 中文

Despite expensive costs, site investigations remain indispensable in geotechnical design for revealing subsurface conditions and their inherent spatial variability. Although a variety of approaches have been proposed to characterise spatially varying soil properties, their practical performance is often constrained by sparse site investigation data and limited physical understanding. Built upon the recently proposed knowledge-informed data-efficient (KIDE) framework developed by Zhang and Yang (2026) for the development of soil property correlations, this study introduces an active learning (AL) guided knowledge-informed multi-fidelity learning framework. The framework integrates low-fidelity data collected from multiple sites with selectively acquired high-fidelity data from a target site through AL, enabling characterisation of quasi-site-specific spatial variability of soil properties. The practicability of the proposed method is demonstrated by predicting cone tip resistance (qt) and blow count profiles, with comparisons against approaches that neglect physical knowledge, AL and on-site low-fidelity data. The results indicate that using only half of the high-fidelity data, the proposed framework achieves the same accuracy and reliability as approaches without AL. The proposed framework can also identify the importance of low-fidelity qt profiles in predicting high-fidelity qt profiles, which generally decreases as the distance between the sites of these profiles increases. By incorporating physical knowledge, AL and on-site low-fidelity data, the proposed framework can characterise the spatial variability of high-cost geotechnical data in a practical and data-efficient way.