Form Finding via Machine Learning 基于机器学习的形态生成
This is a collaboration between computer science andarchitecture to complete an exploratory revolution in structureand space.
If one day, the computer can replace the architect to completethe intellectual level of architectural design, then where shouldthe architecture go? Can architecture open up new designstrategies and design areas in the highly computational future?
This project focuses on how designers use computationalthinking to redefine designs and recreate designs. By usingmachine learning, we try to let computers learn the generationof architectural forms in three-dimensional space and explorethe prospects of machine learning in the field of architecturalform generation. Through data collection and data training inneural network, new architectural forms are created dependingon the training output. These machine-designing forms arecollected to form the future city scenarios.
这是计算机科学与建筑学的合作，共同完成结构与空间的探索性革命。如果有一天，计算机能够取代建筑师完成建筑设计的智力层面，那么建筑将走向何方？在高度计算化的未来，建筑能否开辟新的设计策略和设计领域？
本项目聚焦于设计师如何运用计算思维重新定义和再创造设计。通过机器学习，我们尝试让计算机学习三维空间中建筑形式的生成，探索机器学习在建筑形式生成领域的前景。通过数据收集和神经网络的数据训练，依赖训练输出创造新的建筑形式。这些机器设计的形式被收集起来，构建未来城市的场景。
Data Preperation 数据准备
Through machine learning of big data, the design logichidden behind data will be revealed, thus creating anartificial intelligence model to predict the architecturalforms that match the form-finding goals. This neural network model concludes design strategiesfrom learning procedure, storeing design rules.and generating architectural forms by recalling the parameters. To match with the feasible data structure for machine learing. Details that have itle effects on the overall forms will beremoved. Simplified models can be built by extracting the main section curves and lofting them together. The basic idea is to translate a surace into a series of real number by extractina the coordinates of the controlling pointsto match the data structure of the machine learning network. The collection of the real numbers together represent theform of the tower, showing a more efficient data structure than 3D volumes with a large number of voxels. 通过大数据的机器学习，揭示隐藏在数据背后的设计逻辑，从而创建一个人工智能模型，用于预测与形态生成目标相匹配的建筑形式。该神经网络模型通过学习过程总结设计策略，存储设计规则，并通过回调参数生成建筑形式。 为了匹配适合机器学习的数据结构，将去除对整体形式影响较小的细节。通过提取主要截面曲线并将其联合成型，可以构建简化的模型。基本思想是通过提取控制点的坐标，将表面转换为一系列实数，以匹配机器学习网络的数据结构。这些实数的集合代表塔楼的形态，展示出比具有大量体素的3D体积更高效的数据结构
Data Training 数据训练
In order to test the performance of the idea of the network with real-world fomms, 300 models of the towers buit by differentdesigners were collected and used as a dataset for training in the next step. Similarly to the designed methodology, thereal-data of the coordinate system were extracted from the collected models. 为了测试该网络在现实世界形式中的表现，收集了由不同设计师建造的300个塔楼模型，并作为下一步训练的数据集。与设计方法类似，从收集到的模型中提取了坐标系统的真实数据。 The input neurons contain the reguired information to start a design. in this case, the input data is the boundary curvatureand the height of the building, and the feature parameters showing the design stvle. The output data is the generated formunder the input condition, showing the predicted design outcome from a given architect or a combination of styles fromdiferent architects. The hidden layer works to map the input and output data, mathematically expressing the design rules. 输入神经元包含启动设计所需的信息。在此情况下，输入数据为建筑的边界曲率和高度，以及显示设计风格的特征参数。输出数据为在输入条件下生成的形式，展示了给定建筑师或不同建筑师风格组合的预测设计结果。隐藏层的作用是将输入和输出数据进行映射，数学地表达设计规则。
Testing Process 测试过程
To test the performance of the network, the testing dataset was generated and inputted to the network. in the trainincprocess, the network will only update itself by learning the training dataset without loading the testing dataset, socomparing the expected torms and the predicted torms ot the tesung dataset is the best wav to evaluate the bertommanceof the network. 为了测试网络的性能，生成了测试数据集并将其输入网络。在训练过程中，网络仅通过学习训练数据集进行自我更新，而不会加载测试数据集。因此，将预期的形式与测试数据集的预测形式进行比较，是评估网络性能的最佳方式。。
