Dataset and code for: Machine learning prediction of direction-dependent nonlinear pushover capacity for torsionally irregular RC buildings

This dataset supports the paper "Machine learning prediction of direction-dependent nonlinear pushover capacity for torsionally irregular RC buildings" (Journal of Building Engineering). It contains the OpenSeesPy models, the generated building population and its multi-directional pushover capacity dataset, and the machine-learning training and evaluation code needed to reproduce all reported results and figures. A population of 253 low-rise reinforced-concrete moment frames is generated by Monte Carlo sampling of design-stage parameters. Torsional (plan) irregularity is introduced through a continuous stiffness eccentricity, produced by grading the square column depths across the plan while the floor mass is kept uniform. Each frame is modelled in three dimensions in OpenSeesPy (force-based elements with concentrated Gauss-Radau fibre hinges, rigid diaphragms, P-Delta) and analysed by nonlinear static (pushover) analysis at 36 loading directions (0-360 degrees in 10-degree steps). Each base-shear- roof-drift curve is reduced to a peak-bounded equal-energy bilinear backbone (V_y, d_y, V_max, d_max). A supervised multi-output regression maps design-stage features (geometry, materials, reinforcement, member sizes, eccentricities e_x/e_y, loading angle, and physics-informed angle-eccentricity interaction terms) to the four backbone parameters. Nine regression algorithms are benchmarked under leave-building-out cross-validation, with CatBoost the most accurate. Contents: - generate_dataset.py : OpenSeesPy model generation and multi-directional pushover; produces the building population and the per-direction capacity dataset. - dataset_buildings.csv / dataset_curves.csv : the generated building parameters and the per-direction capacity/backbone data used for training and evaluation. - param_backbone.py, benchmark.py : feature construction, the nine ML models, the two- stage model, leave-building-out cross-validation, and all reported metrics. - make_figures.py : reproduces the manuscript figures from the dataset and models. - README.txt : file descriptions, software dependencies and versions, and step-by-step instructions to reproduce the results. Software: Python 3.x with OpenSeesPy, NumPy, pandas, scikit-learn, XGBoost, LightGBM, CatBoost, and Matplotlib. Exact versions are listed in README.txt. These materials allow full reproduction of the dataset, the trained surrogates, the accuracy metrics (overall R2 = 0.95, within-building directional R2 = 0.91), and the figures reported in the article.本数据集支撑论文《面向扭转不规则钢筋混凝土（Reinforced Concrete, RC）建筑的方向相关非线性推覆承载力机器学习预测》（发表于《Journal of Building Engineering》）。数据集包含复现论文所有报道结果与图表所需的OpenSeesPy模型、生成的建筑种群及其多向推覆承载力数据集，以及机器学习训练与评估代码。 研究通过对设计阶段参数开展蒙特卡洛抽样，生成了包含253栋低层钢筋混凝土弯矩框架的建筑种群。通过连续刚度偏心引入扭转（平面）不规则性：在保持楼层质量均匀的前提下，对平面内方形柱截面尺寸进行梯度调整，以此生成连续变化的刚度偏心。所有框架均在OpenSeesPy中构建三维模型（采用带集中高斯-拉多纤维铰的基于力的单元、刚性楼板与P-Delta效应），并开展36个加载方向（0°~360°，步长10°）的非线性静力推覆分析。将每一条基底剪力-屋顶位移曲线简化为受峰值约束的等能双线性骨架曲线（V_y, d_y, V_max, d_max）。采用监督式多输出回归模型，将设计阶段特征（几何参数、材料属性、配筋信息、构件尺寸、偏心距e_x/e_y、加载角度，以及物理信息驱动的角度-偏心交互项）映射至上述四项骨架曲线参数。针对9种回归算法开展留建筑交叉验证（leave-building-out cross-validation）基准测试，其中CatBoost模型精度最优。 数据集内容如下： - generate_dataset.py：用于OpenSeesPy模型生成与多向推覆分析，可生成建筑种群与各方向承载力数据集。 - dataset_buildings.csv / dataset_curves.csv：用于模型训练与评估的建筑参数集与各方向承载力/骨架曲线数据集。 - param_backbone.py、benchmark.py：用于特征构建、9种机器学习模型实现、两阶段模型构建、留建筑交叉验证，以及所有报道指标的计算。 - make_figures.py：可基于数据集与训练模型复现论文图表。 - README.txt：包含文件说明、软件依赖与版本信息，以及复现研究结果的分步操作指南。 所需软件：Python 3.x环境下的OpenSeesPy、NumPy、pandas、scikit-learn、XGBoost、LightGBM、CatBoost与Matplotlib库，精确版本信息详见README.txt。 本数据集可完整复现论文报道的数据集、训练得到的替代模型、精度指标（整体决定系数R²=0.95，建筑内部方向相关决定系数R²=0.91）以及所有研究图表。