ÌìÃÀ´«Ã½

The construction of shafts in London Clay generates ground movements that are critical to the safe management of adjacent infrastructure. Empirical design curves — while convenient — are limited in scope and unable to account for the full range of shaft geometries, construction sequences, and soil stiffness profiles encountered in practice.

Surrogate models based on artificial neural networks (ANNs), trained from databases of detailed PLAXIS finite element analyses, are shown to reproduce vertical and horizontal displacements around newly excavated shafts with high accuracy across the full parameter space. A universal model was developed to cover the diversity of ground conditions encountered in London Clay, enabling rapid pre-design assessments at a fraction of the computational cost of the underlying finite element analyses. The trained models and associated datasets are freely available.

Publications

Ruiz López, A., Taborda, D.M.G., Tsiampousi, A., Pedro, A.M.G. & Hardy, S. (2024). Applying machine learning to the development of surrogate models for shafts in clay. Proc. ECSMGE 2024, Lisbon.
 
Surrogate model predicting vertical displacements around newly excavated shafts in clay. Zenodo · Dataset: Zenodo
 
Universal surrogate model for predicting ground movements resulting from shaft construction in London Clay. Zenodo · Dataset: Zenodo

Predicting the response of deep excavations in complex urban environments requires high-fidelity finite element analyses with realistic soil constitutive models. However, the computational expense of such analyses makes their direct use in design optimisation or back-analysis workflows impractical. A persistent additional challenge is determining how many training samples are needed: prescribing the database size a priori risks both under-sampling complex regions and over-sampling smooth ones.

CV-BASHES (Cross-Validation Batch Adaptive Sampling for High-Efficiency Surrogates) addresses this directly. Starting from a small initial set of analyses, the algorithm iteratively expands the database by placing new samples in regions where the surrogate error is largest, continuing until a cross-validation quality criterion is satisfied. The result is an automatically determined database rather than one prescribed a priori, reducing wasted computation without sacrificing accuracy.

Applied to a deep excavation in London supported by a multi-propped retaining wall — with London Clay modelled using the IC MAGE M01 constitutive model in PLAXIS 2D — CV-BASHES produced a training database of just 90 samples. The resulting ANN surrogate reproduces wall displacement profiles across multiple construction stages with high accuracy, and is freely available alongside its training dataset.

Publications

Yang, Y., Taborda, D.M.G., Tsiampousi, A. & Ruiz Lopez, A. (2025). A novel adaptive sampling approach with batch selection for the automatic generation of surrogate models in geotechnical engineering. Data-Centric Engineering, Cambridge University Press.
 
Yang, Y., Taborda, D.M.G., Tsiampousi, A. & Ruiz Lopez, A. (2025). Surrogate model for predicting the response of a deep excavation in London supported by a multi-propped retaining wall. Zenodo 

Yang, Y. (2025). Response of a deep excavation in London supported by a multi-propped retaining wall with parametrised geometry (1.0) [Data set]. .

The Observational Method — in which design decisions are updated progressively as monitoring data become available during construction — has long been advocated as a means of achieving more efficient and sustainable geotechnical designs. Its widespread adoption in practice has nonetheless been limited, in part because the back-analysis of soil parameters from field observations is computationally prohibitive when performed directly using finite element models.

Surrogate models remove this barrier. A framework coupling ANN surrogate models with a Genetic Algorithm (GA), on any other global optimisation algorithm, enables the back-analysis of soil parameters to be performed automatically and efficiently, exploiting the ability of the surrogate to evaluate hundreds of candidate parameter sets per minute. The surrogate is trained on synthetic data from a high-fidelity finite element model and shown to reproduce wall displacement profiles across all construction stages with high accuracy. The GA then identifies the set of soil parameters that best matches the accumulated monitoring data from preceding stages, and the calibrated model is used to make forward predictions for subsequent stages — a workflow directly aligned with the sequential logic of the Observational Method.

The framework was demonstrated through application to a well-documented 11.4 m deep multi-propped diaphragm wall excavation in London Clay, for which inclinometer data across multiple construction stages are available in the literature. Forward predictions obtained using back-analysed parameters were shown to outperform those based on the initially assumed soil properties, demonstrating the practical value of incorporating monitoring data systematically into the prediction workflow. A more extensive treatment of the methodology and its application — including sensitivity analysis identifying which soil parameters govern wall behaviour at each stage — is provided in a paper currently under review.

Publications

Thermal Integrity Profiling (TIP) is a non-destructive testing method for assessing cast-in-place concrete piles during early curing, based on monitoring temperature variations arising from the exothermic hydration of concrete. Deviations from expected temperature profiles indicate the presence of anomalies such as soil inclusions, air voids, or rebar cage misalignment. Conventional TIP interpretation relies on manual visual inspection, which is inherently subjective and unable to reliably quantify defect location or severity.

A data-driven framework has been developed to automate and standardise TIP interpretation using deep neural networks trained on numerically generated databases of hydration temperature profiles. The initial framework, based on two-dimensional finite element simulations, uses an ANN classifier to detect the presence of defects and a regressor to estimate their properties, achieving high predictive accuracy across a wide range of pile geometries and ground conditions.

The framework has been substantially extended to a fully three-dimensional formulation, overcoming the limitations of the 2D approach. The 3D FE model — implemented in COMSOL Multiphysics and automated via Livelink for MATLAB — generates a database of 19,445 simulations covering randomised pile lengths, layered soil stratigraphies, cage misalignments, and multi-defect configurations of up to three spherical voids per pile. Two deep convolutional neural networks are trained on this database: a multi-label classifier identifies the presence or absence of defects in each soil layer, and a regressor estimates defect volume, three-dimensional location, reinforcement cage displacement, and concrete hydration parameters. The classifier achieves a per-defect accuracy of 93.6% (96.8% per sample), and the regression model achieves near-perfect accuracy for cage displacement (R² = 0.992) and defect volume (mean absolute error of 0.03 m³), with defect location reaching an adjusted R² of 0.92 once intrinsically undetectable micro-defects are excluded. The classification threshold can be tuned to suit the risk tolerance of the application: lowering it increases sensitivity at the cost of more false positives, a useful capability in safety-critical foundation assessment.

Publications

Sánchez Fernández, J., Ruiz López, A. & Taborda, D.M.G. (2025). A novel machine learning-based approach to thermal integrity profiling of concrete pile foundations. Data-Centric Engineering, 6, e33.

Sánchez Fernández, J., Ruiz López, A. & Taborda, D.M.G. Thermal integrity profiling of concrete piles incorporating three-dimensional defects using deep convolutional neural networks. Geomechanics and Geoengineering: An International Journal. Under review

Datasets

Sanchez Fernandez, J. (2025). Thermal Integrity Profiling modelled data for ML training (1.0). .

Sanchez Fernandez, J. (2025). Three-dimensional Thermal Integrity Profiling numerical data for ML model training (1.0). .

Thermo-active foundations — piles and diaphragm walls equipped with heat-exchanger pipes — offer a low-carbon route to heating and cooling buildings by exchanging thermal energy with the ground. Optimising large systems involving many such elements requires rapid assessment of thermal performance across a wide range of configurations, pile geometries, ground conditions and operational parameters, making direct finite element simulation impractical at the design stage.

For thermo-active piles, a surrogate model was developed using a parametrised 3D thermo-hydraulic finite element model in COMSOL Multiphysics. An ANN trained on 800 LHS samples — spanning pile geometry, layered soil stratigraphy, pipe arrangement, fluid velocity, and inlet and initial temperatures — predicts the power output per unit length at 16 transient time steps over the first year of operation as well as at steady state. The model achieves R² consistently above 0.97, with steady-state mean absolute error of 2.4 W/m. SHAP analysis confirms that the inlet–ground temperature difference, number of pipe loops, and concrete conductivity dominate early performance, while fluid velocity becomes less significant at steady state.

A companion study extended the approach to thermo-active diaphragm walls, with a 500-sample database covering wall depth and thickness, concrete and soil conductivities, convective heat transfer coefficient at the exposed boundary, fluid velocity, and temperature differential. The trained ANN achieves an average R² of 0.987 and a steady-state MAE of just 0.8 W/m². SHAP analysis reveals a physically meaningful shift in governing factors: temperature differential and concrete conductivity control early-time behaviour, while the convective heat transfer coefficient at the exposed face dominates long-term output — consistent with the transition from transient conduction-dominated to boundary-condition-dominated thermal response.

Publications

Sanchez Fernandez, J., Provost, A., Ruiz Lopez, A. & Taborda, D.M.G. (2025). A data-driven approach to predicting the long-term thermal performance of thermo-active piles. Proc. 3rd International Conference on Energy Geotechnics (ICEGT 2025), Paris.
 
Sánchez Fernández, J., Taborda, D.M.G. & Ruiz López, A. (2026). Application of machine learning algorithms for power output prediction in thermo-active diaphragm walls. Proc. 21st International Conference on Soil Mechanics and Geotechnical Engineering (ICSMGE 2026), Vienna

The geotechnical design of offshore suction bucket foundations — widely used as anchors and supports for renewable energy infrastructure — requires evaluation of load-displacement behaviour under complex loading conditions across a range of ground profiles. The computational cost of detailed three-dimensional finite element analyses makes direct parametric studies and design optimisation challenging.

A data-driven macroelement model was developed to predict the load-displacement behaviour of suction buckets in sand, enabling design optimisation for varying ground conditions. The approach demonstrates that surrogate models can be deployed directly within design workflows, capturing the essential hydro-mechanical behaviour at negligible computational cost relative to the underlying numerical analyses.

Publications

Ruiz López, A., Taborda, D.M.G. et al. (2025). A data-driven macroelement model for suction buckets in sand. 5th International Symposium on Frontiers in Offshore Geotechnics (ISFOG2025), Nantes.