Residential reinforced concrete building design relies on close collaboration between architectural and engineering offices to improve the distribution of living spaces while meeting structural regulatory requirements. Several studies have taken advantage of the vast data generated by both offices to create machine-learning models that streamline design processes and decision-making. Recent research proposed an artificial neural network (ANN) model for predicting the length and thickness of the rectangular segments that constitute the plan's walls based on the architectural data; however, it could not predict walls absent from the original design. This constraint was addressed by a convolutional neural network (CNN) model, demanding a larger dataset (by 137 times) and several rule-based filters for assembling the predicted plan, incurring high computational costs, and generating blurry predictions. Therefore, this study presents a new methodology to propose walls and columns not considered in the architectural design through an ANN model, which employs less data than CNN but with comparable results. First, this study creates a dataset of 165 Chilean buildings using a mapping function capable of generating neighborhoods within the floors and extracting their walls' geometric and topological features. Then, we trained an ANN model to predict unconsidered wall segments in architectural design, using a feature vector that addresses conditions such as thickness, wall connectivity, distance between elements, seismic zone, foundation soil type, and other engineering parameters, achieving outstanding results in terms of the coefficient of determination (R) of 0.95 for length, 0.93 for thickness, 0.94 for angle, and 0.97 for position (x, y). Finally, with an architectural plan, this model can propose different structural solutions, reducing the data used for training and validation to 8% concerning the CNN model, with comparable performance.

中文翻译：

---

通过具有基于结构邻居的特征的人工神经网络优化钢筋混凝土墙和柱

住宅钢筋混凝土建筑设计依赖于建筑和工程办公室之间的密切合作，以改善居住空间的分布，同时满足结构监管要求。多项研究利用两个办公室生成的大量数据来创建机器学习模型，以简化设计流程和决策。最近的研究提出了一种人工神经网络（ANN）模型，用于根据建筑数据预测构成平面墙壁的矩形段的长度和厚度；然而，它无法预测原始设计中不存在的墙壁。这一限制已通过卷积神经网络 (CNN) 模型得到解决，该模型需要更大的数据集（137 倍）和多个基于规则的过滤器来组装预测计划，从而产生较高的计算成本并生成模糊的预测。因此，本研究提出了一种新的方法，通过 ANN 模型提出建筑设计中未考虑的墙和柱，该模型使用的数据比 CNN 少，但结果可比。首先，本研究使用映射函数创建了 165 座智利建筑物的数据集，该映射函数能够生成楼层内的邻域并提取其墙壁的几何和拓扑特征。然后，我们训练了一个 ANN 模型来预测建筑设计中未考虑的墙段，使用特征向量来处理厚度、墙体连通性、元素之间的距离、地震带、地基土壤类型和其他工程参数等条件，在以下方面取得了出色的结果长度决定系数 (R) 为 0.95，厚度为 0.93，角度为 0.94，位置 (x, y) 为 0.97。最后，通过架构规划，该模型可以提出不同的结构解决方案，将用于训练和验证的数据减少到 CNN 模型的 8%，并且性能相当。