An open-source platform for shear capacity prediction of stirrup-free FRP-reinforced concrete beams integrating design codes and ensemble machine learning

FRP-ShearPred is a cross-platform desktop application for shear strength assessment of fibre-reinforced polymer (FRP) reinforced concrete beams without shear reinforcement. It brings together five international design codes and eight ensemble machine learning algorithms in a single graphical environment, so that code predictions, ML predictions, and experimental data can all be compared without switching tools.

The software is aimed at structural engineers and researchers who work with FRP-RC experimental databases. A typical workflow is to load a dataset, train one or more ML models with Bayesian or evolutionary hyperparameter optimisation, inspect model behaviour through SHAP and partial dependence plots, and export results for reliability or statistical post-processing.

Prediction Tab — enter beam parameters, load a trained model bundle, and compare all design-code predictions against the ML model in a single view.

Training Summary — ranked performance table for all eight algorithms after Bayesian optimisation, with live convergence curves.

Algorithm Configuration — per-algorithm hyperparameter editor with lock/free toggles and user-defined search ranges for each optimisation strategy.

Feature Importance — Gini impurity-based ranking for the best model (Extra Trees shown); all diagnostic plots are exportable as CSV or image files.

• Five design codes side-by-side: GB 50608-2020, ACI 440.1R-15, CSA S806-12, BISE (1999), JSCE (1997)
• Eight ensemble ML algorithms: GBDT, XGBoost, LightGBM, CatBoost, Random Forest, Extra Trees, AdaBoost, KNN
• Three hyperparameter optimisation strategies: Bayesian search (Optuna TPE), TLBO, and NSGA-II multi-objective
• 95 % split conformal prediction intervals (Vovk et al., 2005) attached to every ML prediction, with a distribution-free finite-sample marginal coverage guarantee
• Applicability-domain check: per-input training-range tooltips warn when a query falls outside the data envelope tree-based models can support
• Non-negative output constraint applied to both design-code formulas and ML predictions (shear capacity is physically ≥ 0)
• Model interpretability via SHAP beeswarm plots, Gini feature importance, and response surface analysis
• Batch prediction over entire CSV/Excel databases with one-click export
• Portable .frpmdl model bundles that carry the fitted model, scaler, encoder, and metadata in one file
• Interactive beam cross-section schematic with annotated parameter labels

• Python 3.9 or later
• Operating system: Windows 10/11, macOS 12+, or Linux (Ubuntu 20.04+)

git clone https://github.com/hunter137/FRP-RC-Shear.git
cd FRP-RC-Shear
conda env create -f environment.yml
conda activate frpshear
python main.py

git clone https://github.com/hunter137/FRP-RC-Shear.git
cd FRP-RC-Shear
pip install -r requirements.txt
python main.py

CatBoost is excluded from the default install due to its large package size (~400 MB). Install separately if needed:

pip install catboost

1. Open the Prediction tab.
2. Enter beam parameters (a/d, d, b, f′c, ρf, Ef, FRP type).
3. Load a pre-trained model bundle (.frpmdl) or train your own in the Training tab.
4. Click Predict to obtain results from all design codes and the loaded ML model simultaneously. A 95 % conformal prediction interval is shown below the point estimate.
5. Click the small ? button beside any input to view that parameter's training-data range; values outside the range are flagged because tree-based models do not extrapolate.
6. Click Export CSV to save results.

1. Open the Model Retraining tab.
2. Load an experimental database (.xls, .xlsx, or .csv).
3. Map columns to required features via the interactive column-mapping dialog.
4. Select algorithms and an optimisation strategy (Bayesian / TLBO / NSGA-II).
5. Click Train; live metrics and a progress bar are displayed throughout.

# Train all models with Bayesian optimisation using the included example dataset
python train_frp_models.py --data data/testdata.xls

# Train specific models only
python train_frp_models.py --data data/testdata.xls --only LightGBM KNN

# Reduce trials for a faster exploratory run
python train_frp_models.py --data data/testdata.xls --trials 200

# Set a wall-clock time limit (minutes)
python train_frp_models.py --data data/testdata.xls --time-limit 60

1. Click Batch Prediction in the Prediction tab.
2. Select an Excel/CSV file containing beam parameters in the required column format.
3. Results are computed for all design codes and the loaded ML model.
4. Export to CSV for downstream reliability or statistical analysis.

Trained models are saved as .frpmdl files (compressed joblib archives). Each bundle contains:

• Fitted model object(s)
• MinMaxScaler fitted on training data
• Feature column names and OneHotEncoder for categorical inputs
• Training/test metrics and cross-validation scores
• SHAP calibration subsample (up to 400 rows)
• Metadata: algorithm name, hyperparameters, training timestamp, software version

FRP-RC-Shear/
├── main.py                  # Application entry point
├── app.py                   # MainWindow: assembles all tabs
├── train_frp_models.py      # Command-line training script
├── formulas.py              # Design code formula implementations
├── config.py                # Global constants and colour palette
├── column_mapping.py        # Database column auto-detection and mapping
├── metrics.py               # Regression evaluation metrics (R², RMSE, MAE, …)
├── model_io.py              # Model bundle save/load (.frpmdl)
├── optimization.py          # Hyperparameter search: TLBO, Bayesian, NSGA-II
├── qt_compat.py             # PyQt5/PySide6 compatibility shim
├── widgets.py               # Shared UI helper widgets
├── requirements.txt         # pip dependencies
├── environment.yml          # conda environment specification
├── LICENSE                  # MIT License
├── README.md                # This file
├── CITATION.cff             # Machine-readable citation metadata
├── CHANGELOG.md             # Version history
├── data/
│   └── testdata.xls         # Example experimental database (728 specimens)
├── models/
│   └── README.md            # How to obtain pre-trained model bundles
├── docs/
│   └── screenshots/         # Application screenshots
├── tests/
│   ├── test_formulas.py     # Unit tests — design code formulas
│   └── test_metrics.py      # Unit tests — evaluation metrics
├── tools/
│   ├── diagnose_crash.py    # Crash diagnostics tool (run instead of main.py)
│   └── README.md            # Usage instructions for tools
└── tabs/                    # GUI tab modules and supporting dialogs/threads
    ├── data_tab.py
    ├── train_tab.py
    ├── eval_tab.py
    ├── code_tab.py
    ├── predict_tab.py
    ├── interp_tab.py
    └── …                    # dialogs, helpers, worker threads, hyperparameter editor

If the application exits silently or crashes without a visible error message, run the diagnostics tool instead of main.py:

python tools/diagnose_crash.py

This generates crash_diag.log in the project root. Attach that file when opening a bug report. See tools/README.md for the full list of what the tool captures.

python -m pytest tests/ -v

FRP-ShearPred targets a specific structural configuration; users should be aware of the following before applying it in practice:

• Intended configuration: rectangular cross-section concrete beams, longitudinal FRP reinforcement only, no transverse reinforcement (stirrups), monotonic loading.

• Training-data ranges of the bundled 728-specimen dataset (from the fitted MinMaxScaler):

  Parameter	Min	Max	Unit
  a/d	0.55	16.22	—
  d	73	1111	mm
  b	89	1000	mm
  f′c	20	93	MPa
  ρf	0.09	3.98	%
  Ef	29	192	GPa
  FRP type	CFRP, GFRP, BFRP, AFRP

  The Prediction tab flags inputs that fall outside these ranges. Tree-based ensemble models cannot extrapolate and will repeat boundary predictions for out-of-range queries.

• Uncertainty estimates: the 95 % conformal interval is a statistical predictive interval with marginal coverage; it is not an engineering safety factor and is not a substitute for code-prescribed material/resistance factors.

• Recommended use: the ML predictions are intended as a complement to design-code formulas — useful for parameter screening, sensitivity studies, and educational purposes. Safety-critical design decisions should follow the relevant code provisions.

This work was supported by the National Key R&D Program of China (Grant Nos. 2024YFC38098 and 2024YFC3809803), the Liaoning Xingliao Talents Program for Science and Technology Innovation Team (No. XLYC2404005), and the Technology Research and Development Program of Shenyang Science and Technology Bureau (Grant No. 24-213-3-33).

If you use FRP-ShearPred in your research, please cite the following SoftwareX article:

@article{liang2026frpshearpred,
  author  = {Liang, Deyu and Cao, Jingwen and Liu, Jinlong and
             Cui, Yujun and Zhang, Yuzhuo and Xue, Xingwei and Xu, Lei},
  title   = {{FRP-ShearPred}: An open-source platform for shear capacity
             prediction of stirrup-free {FRP}-reinforced concrete beams
             integrating design codes and ensemble machine learning},
  journal = {SoftwareX},
  volume  = {35},
  year    = {2026},
  pages   = {102811},
  doi     = {10.1016/j.softx.2026.102811},
}

If you use the software itself (the released code/binaries), you may additionally cite the Zenodo archive:

@software{liang2026frpshearpred_software,
  author  = {Liang, Deyu and Cao, Jingwen and Liu, Jinlong and
             Cui, Yujun and Zhang, Yuzhuo and Xue, Xingwei and Xu, Lei},
  title   = {{FRP-ShearPred}},
  year    = {2026},
  doi     = {10.5281/zenodo.19503522},
}

This project is licensed under the MIT License — see the LICENSE file for details.