Dataset for a framework for assessing the impact of geometric imperfections in concrete shell structures using deep learning
This dataset contains scripts and data supporting the following following thesis: Pollet, M. (2025). Rapid structural analysis of prefabricated thin concrete shells using deep learning (Thesis). University of Bath.
Concrete thin-shells are materially efficient structures, which can be used to reduce the environmental impact of concrete structures. However, geometric imperfections, which may occur during production can negatively impact their structural behaviour. While this impact can be assessed through Finite Element Analysis (FEA), a faster analysis method, such as surrogate modelling, could benefit concrete shell manufacturers.
This dataset contains deep learning models – Multilayer Perceptrons, and Convolutional Neural Networks – that have been trained to predict the buckling factor and stress fields of geometrically imperfect concrete thin-shells of various shapes under design loads. It also contains the Python scripts that were used to train these models and assess their performance. Running these scripts necessitates the associated ConcreteShellFEA dataset to be downloaded. Further details about this data can be found in the related thesis.
Cite this dataset as:
Pollet, M.,
Shepherd, P.,
Hawkins, W.,
2026.
Dataset for a framework for assessing the impact of geometric imperfections in concrete shell structures using deep learning.
Bath: University of Bath Research Data Archive.
Available from: https://doi.org/10.15125/BATH-01532.
Export
Data
models.zip
application/zip (1GB)
Creative Commons: Attribution 4.0
This folder contains the trained deep learning models used to achieve the predictive performances highlighted in the thesis. Files formats: PyTorch (.pth).
results.zip
application/zip (85MB)
Creative Commons: Attribution 4.0
This folder contains the predictions made by the deep learning models, which were used to calculate used to the predictive performances highlighted in the thesis. Files formats: PyTorch (.pth).
Code
datasets.py
text/x-script.python (11kB)
Software: MIT License
A Python script with classes that can be used to load the datasets.
metrics.py
text/x-script.python (8kB)
Software: MIT License
A Python script with error metrics classes.
models.py
text/x-script.python (6kB)
Software: MIT License
A Python script with the deep learning model classes.
tools.py
text/x-script.python (11kB)
Software: MIT License
A Python script that contains some tool functions.
train_CNN_buckling.py
text/x-script.python (16kB)
Software: MIT License
A Python script that can be used to train and evaluate a CNN for buckling prediction.
train_CNN_stress.py
text/x-script.python (19kB)
Software: MIT License
A Python script that can be used to train and evaluate a CNN for stress prediction.
train_MLP_buckling.py
text/x-script.python (15kB)
Software: MIT License
A Python script that can be used to train and evaluate a MLP for buckling prediction.
train_MLP_stress.py
text/x-script.python (18kB)
Software: MIT License
A Python script that can be used to train and evaluate a MLP for stress prediction.
Creators
Maxime Pollet
University of Bath
Paul Shepherd
University of Bath
Will Hawkins
University of Bath
Contributors
University of Bath
Rights Holder
Documentation
Data collection method:
The methods used to generate this data can be found in the related thesis.
Technical details and requirements:
The data in the models and results folders was generated using the Python code in scripts folder. These scripts rely on the dependencies listed in requirements.txt.
Additional information:
The original folder structure is given in README.md. To reproduce it, create a new folder and extract the "models.zip" and "results.zip" folders inside. Additionally, create a "scripts" folder and store all Python scripts inside. The path to the ConcreteShellFEA dataset needs to be specified in each script, under the DATASET_ROOT variable.
Methodology link:
Pollet, M., 2025. Rapid structural analysis of prefabricated thin concrete shells using deep learning: (Alternative Format Thesis). Thesis (PhD). University of Bath. Available from: https://researchportal.bath.ac.uk/en/studentTheses/rapid-structural-analysis-of-prefabricated-thin-concrete-shells-u.
Documentation Files
requirements.txt
text/plain (589B)
Creative Commons: Attribution 4.0
README.md
text/plain (3kB)
Creative Commons: Attribution 4.0
Funders
University of Bath
https://doi.org/10.13039/501100000835
Publication details
Publication date: 9 February 2026
by: University of Bath
Version: 1
DOI: https://doi.org/10.15125/BATH-01532
URL for this record: https://researchdata.bath.ac.uk/1532
Related theses
Pollet, M., 2025. Rapid structural analysis of prefabricated thin concrete shells using deep learning: (Alternative Format Thesis). Thesis (PhD). University of Bath. Available from: https://researchportal.bath.ac.uk/en/studentTheses/rapid-structural-analysis-of-prefabricated-thin-concrete-shells-u.
Related datasets and code
Pollet, M, Shepherd, P, Hawkins, W, and Costa, E, n.d. ConcreteShellFEA: A surrogate modelling dataset for the buckling and stress behaviour of concrete thin-shells. University of Bath Research Data Archive. Available from: https://doi.org/10.15125/BATH-01519.
Contact information
Please contact the Research Data Service in the first instance for all matters concerning this item.
Contact person: Maxime Pollet
Faculty of Engineering & Design
Architecture & Civil Engineering
Research Centres & Institutes
Centre for Digital, Manufacturing & Design (The Foundry)
