A knowledge-informed data-efficient framework for soil correlation modelling

Pin Zhang

doi:10.1016/j.compgeo.2026.108190

A knowledge-informed data-efficient framework for soil correlation modelling

2026-05-11

0

PRE

AI

DOI：10.1016/j.compgeo.2026.108190

Original

Original request for help

Save

Abstract

En 中文

Accurate prediction of soil properties is essential in geotechnical engineering, yet high-quality data are often scarce, variable, and noisy. This study presents a knowledge-informed data-efficient framework for soil correlation modelling, demonstrated for the compression index (Cc). Active learning (AL) was integrated with neural networks (NN) to identify the most informative data from limited datasets, maximising model efficiency. A Positive Output Neural Network (PONN) was proposed to ensure strictly positive predictions, while Monte Carlo (MC) dropout quantified predictive uncertainty. The results indicate that the AL-selected active set was substantially smaller than the raw dataset, yet the NN trained on it achieved comparable accuracy. Compared with traditional curve-fitting correlations, the NN-based model provided higher accuracy and reliable uncertainty estimates. An online tool incorporating both the NN-based and traditional correlations allows Cc to be predicted directly from initial void ratio, liquid limit, and plastic limit. This framework demonstrates the potential of combining AL and knowledge-informed NNs with uncertainty quantification to maximise the value of sparse geotechnical data and standardise soil correlation modelling.

Keywords:

Active learning

Neural networks

Soil correlation modelling

Uncertainty quantification

Data-efficient framework

A knowledge-informed data-efficient framework for soil correlation modelling

Abstract

Journal

Organization

Cited Papers