See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/348087327

Training deep convolution network with synthetic data for architectural

morphological prototype classification

Article in Frontiers of Architectural Research · December 2020

DOI: 10.1016/j.foar.2020.12.002

CITATIONS READS

10 7,967

2 authors, including:

Chenyi Cai
Future Cities Lab - ETH Zurich
12 PUBLICATIONS 42 CITATIONS

SEE PROFILE

All content following this page was uploaded by Chenyi Cai on 13 June 2021.

The user has requested enhancement of the downloaded file.

Frontiers of Architectural Research (2021) 10, 304e316

Available online at www.sciencedirect.com

ScienceDirect

journal homepage: www.keaipublishing.com/foar

Research Article

Training deep convolution network with

synthetic data for architectural
morphological prototype classification
Chenyi Cai, Biao Li*

School of Architecture, Southeast University, Nanjing 210096, China

Received 13 July 2020; received in revised form 4 December 2020; accepted 5 December 2020

KEYWORDS Abstract The use of architectural morphological analysis and generative design is an impor-
Deep convolution tant strategy to interpret current designs and to propose novel ones. Conventional morpholog-
network; ical features are defined based on qualitative descriptions or manually selected indicators,
Architectural which include subjective bias, thus limiting generalizability. The lack of public architectural
morphology; morphological datasets also leads to setbacks in data-driven morphological analysis. This study
Prototype proposed a new method for generating topology-based synthetic data via a rule-based system
classification; and for encoding morphological information to promote morphological classification via deep
Feature extraction; learning. A deep convolution network, LeNet, which was modified in the output layer, was
Generative design trained with synthetic data, including five spatial prototypes (central, linear, radial, cluster,
and grid). The performance of the proposed method was validated on 40 practical architec-
tural layouts. Compared to the ground truth, the proposed method provided an encouraging
accuracy of 97.5% (39/40). Interestingly, the most possible mistakes of the LeNet were also un-
derstandable according to the architect’s intuitive perception. The proposed method consid-
ered the statistical and overall characteristics of the training samples. This work
demonstrated the feasibility and effectiveness of the deep learning network trained with syn-
thetic architectural patterns for morphological classification in practical architectural layouts.
The findings of this work could serve as a basis for further morpho-topology studies and other
social, building energy, and building structure studies related to spatial morphology.
ª 2020 Higher Education Press Limited Company. Publishing services by Elsevier B.V. on behalf
of KeAi Communications Co. Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).

* Corresponding author.
E-mail address: [email protected] (B. Li).
Peer review under responsibility of Southeast University.

https://doi.org/10.1016/j.foar.2020.12.002
2095-2635/ª 2020 Higher Education Press Limited Company. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co. Ltd.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Deep CNN for morphological classification 305

1. Introduction phase of the proposed process was time-consuming,

because there were few shared datasets related to archi-
According to Li and Han (2011), architectural design re- tectural morphology.
quires an integrated balance of complex adaptive systems To develop a morphology-to-data transmission
(CAS). One solution is generative design, which focuses on approach for morphological classification, two sets of
the translation and simulation of design concepts using obstacles were observed: (1) the availability of architec-
computational models, including decision making, linkage tural morphology dataset for inputting deep learning al-
construction, and design optimization. The computational gorithms and (2) the quantification of architectural
model is established based on the large amounts of data morphological features. In the computer science field,
and the extraction of rational rules to achieve novel design feeding deep learning algorithms with synthetic data for
proposals (Li, 2012; Soddu, 1998). the recognition of actual conditions could achieve
From the rationalist view, the morphological approach considerable results (Srisuchinnawong et al., 2018), with
proposes the idea that morphology has the potential to be the synthetic data generated based on morphological
the driving force behind the urban design process (Rossi similarity and diversity. However, the application of syn-
et al., 1982). Given that architectural morphology de- thetic training data for testing practical data in the
termines the spatial footprint pattern and influences the architectural morphological field has yet to be fully
urban fabric (Levy, 1999), many studies have linked explored.
morphological types to the aspects of the building energy, According to the previous discussion, generating syn-
building interior, social effects, urban evolution, etc. For thetic data based on topo-morphological similarity may
example, a study on urban neighborhoods (Ariga, 2005) offer an opportunity to provide effective training samples
used the morphological classification and clustering of for deep learning algorithms. A fully automated feature
footprint patterns over time in order to extract evolu- extraction method may overcome the drawbacks of manu-
tionary patterns that result in sustainable urban neighbor- ally selecting indicators. The pictorial processing of the
hoods. In addition to conceptual and qualitative studies on morphology could be taken as the data source for images,
morphological classification, recent morphological similar- so a possible method is feature mapping of the images. An
ity studies have been carried out quantitatively by consid- image is a 3D matrix (RGB) that contains spatial informa-
ering shape, size, spatial proportion, and other geometrical tion. For example, spatially adjacent pixels have similar
measures. values, whereas pixels that are farther apart have little
In quantitative evaluations, researchers utilized various correlation. Therefore, the spatial information of images
morphology-to-data transmission methods by selecting and hides essential features that are worth extracting. Feature
adjusting indicators to represent morphology. For compu- mapping is a transmission from the original data to feature
tational analysis and design, quantifying architectural vectors that contain the information of overall
morphology and function is important in building effective characteristics.
computational models. In the research on urban renewal in In a convolution neural network (CNN), the feature
Roma (Tang et al., 2019), scalar indicators (e.g., block ID, mapping of the images extracts features by convolution
plot area) and geometric indicators (e.g., edge length, kernels. Since its introduction in 1988, LeNet (Lecun
shape) were used to represent block morphology for et al., 1998), the progenitor of CNN, has undergone
searching morphologically similar blocks. However, some continuous improvements and is considered one of the
missing factors remained. This is especially true for geo- classic models of CNN. Neural networks are implemented
metric information that is difficult to represent numeri- for solving problems of pattern recognition of two-
cally, such as the composition of buildings, thus leading to dimensional (2D) images (Krizhevsky et al., 2012) and
subjective bias and limited generalizability. multiple manifolds, such as graphs and two-dimensional
Deep learning1 algorithms promote new methodologies (3D) models (Tom et al., 2018). Researchers have
in morphological analysis and generative design. It statis- applied the manifolds to various morphological domains,
tically provides automatic feature extraction and learning such as style recognition (Yoshimura et al., 2019) and
strategies for morphological analysis and design (Li et al., architectural element classification (Yetis and Yetkin,
2019). Morphology-to-data transmission methods, such as 2018).
image data-based (RGB), numerical labeling, and semantic The current paper aims to develop a novel method for
segmentation (Chaillou, 2019), are dedicated to feeding architectural morphology-related studies. The proposed
samples into neural networks with continuous and infor- method is a combination of the rule-based systems and
mative features. In a study of typo-morphology in Lisbon data-driven approaches, which generate and encode syn-
(Gil et al., 2012), the k-means clustering algorithm was thetic morphological training samples for neural networks
used to classify block and street types based on prepared performing in actual conditions. In this research, a deep
plans. However, the clustering algorithm was limited in CNN was trained with concise feature vectors of synthetic
terms of data size, because the learning algorithms morphological patterns. The synthetic data were generated
required a large number of training samples to ensure by extracting the volume organization rules of architectural
learning accuracy. Furthermore, the data preparation spatial prototypes, taking five types as examples. Thus, the
typological similarity was integrated with the actual

1
Deep learning uses neural networks as its main model and is widely applied in data mining, computer vision, text processing, auton-
omous driving, speech recognition, and other fields.
306 C. Cai, B. Li

conditions. This work further completed the classification tasks. Then, computational models are constructed to
and prototype identification of practical architectural simulate the self-organization of architectural elements
spatial organization. This paper makes the following con- and to achieve design generation in dynamic processes
tributions to the literature: (Tang et al., 2019).

This is an interdisciplinary study that integrates archi- 2.2. Data-driven methods in generative design
tectural design with computer science and applies the
latest deep learning techniques to the study of archi- Data-driven methods use a statistical approach to describe
tectural morphological analysis. and solve problems consisting of a flexible number of ele-

The proposed method effectively performs the auto- ments, connections, and variables (Hovestadt, 2010). Digi-
matic generation of the architectural morphological tal techniques have been widely used in architecture,
patterns as training samples for neural networks, generating large amounts of precise data. For example, 3D
thereby reducing the difficulties in collecting data, scanning, depth detection, and synchronous location and
cleaning data, and labeling data in real applications. mapping (SLAM) technologies have been increasingly used

This study provides a feasible method of comprehen- along with online maps. These techniques provide archi-
sively and automatically encoding the overall morpho- tects with large amounts of data quickly and efficiently.
logical characteristics with feature mapping based on They also help establish a database of architectural cases
pixels, excluding artificial work on morphological in- with feature information and retrieve cases with similar
dicators selection, thus avoiding subjective bias. features (Tang et al., 2019). With a data-driven approach,

The previous methods working on the same tasks are designers need to extract features from abundant of
either the qualitative description or informative inde- training data by designing a network that can support
pendent indicators. In comparison, the proposed design decision-making through data analysis methods.
approach considers both aspects simultaneously.

The signified neural network output can be changed 2.3. Neural network in generative design
arbitrarily to meet a certain demand. Therefore, the
new output can be conveniently adapted for the further Similar to conventional feature engineering, rule-based
application of building energy efficiency evaluation, systems are based on pre-defined rules and translate
building structure, and generative architectural design design rules into programming principles. It means that the
projects related to spatial morphology. information synthesis is limited by the architect’s stock
knowledge and experiences, which may result in subjective
bias and inadequate linkage construction of design vari-
2. Concept model of generative design ables. In comparison, neural network algorithms can effi-
ciently parse out general features from large amounts of
Two generative design approaches are taken in the data and apply those to new data in testing. Neural net-
computational model construction: rule-based system and works are used for feature extraction, prediction, and
data-driven approaches. With the rule-based system, the clustering (Nielsen, 2015). However, limited by the avail-
architect builds a physical or mathematical simulation able data, the application of the neural network in the
model and solves specific architectural problems under architecture field is restricted. Nevertheless, it is an
well-defined rules (Li, 2012). It is an analytical problem- effective way to communicate and simulate the architec-
solving method that can help the architect to efficiently tural design concept by integrating the two approaches. It
complete the linkage construction and optimization steps. supports database construction and feature extraction and
In contrast, the data-driven approach creates statistical then guides the computational model to accurately corre-
models and mines through vast amounts of data. It solves late design elements and variables, ultimately completing
the problems of prediction, recommendation, and feature design evaluation and in-depth optimization.
extraction. Furthermore, it helps to summarize and identify Two key points in the research of the application of deep
non-logical rules, such as personal preference, subjective learning in architecture are (1) the accessibility of training
tendency, style definition, and hand-drawn images, which samples and (2) the information encoding of sample fea-
support architects during the design decision step. tures. Previously published studies construct a training
dataset by crawling online data or manual semantic seg-
2.1. Rule-based system in generative design mentation, which are time-consuming and have a limited
sample size (Chaillou, 2019; Huang and Zheng, 2018; Liu
Digital technology allows architectural elements to be et al., 2019). This process can also generate subjective
freely described and interconnected without losing the noise. A more efficient and accurate approach is the use of
systematic mapping (Hovestadt, 2010). With the rule-based an automatic data preprocess pipeline, which can feed the
system approach, designers need to manually define and architectural morphology to the algorithms.
quantify features, such as the architectural concept, the Fig. 1 represents the general workflow of our work. First,
building elements, the physical environment of the build- five architectural spatial organization prototypes were
ing, and the rules of the elemental association. Further- extracted. The architectural morphological patterns were
more, such algorithms such as evolutionary algorithms, generated based on the prototype abstraction. Second, the
multi-agent systems (Caiet al., 2019), and integer pro- features of the generated samples were mapped into
gramming (Hua et al., 2019), are required to match design feature vectors for the neural network input dataset. Third,
Deep CNN for morphological classification 307

Fig. 1 The workflow of architectural morphological prototype classification based on neural network.

Fig. 2 Five spatial prototypes of architectural morphological organizations.

308 C. Cai, B. Li

Fig. 3 Generation rules and process of the five types of morphological pattern samples.

the output layer modified LeNet was trained compared with conventional data augmentation operations, such as
a simple fully-connected neural network. Finally, 40 prac- rotation, mirroring, etc., are used for making minor
tical architectural layouts were taken to test the perfor- changes to existing datasets in order to acquire more
mance of the trained neural networks for morphological samples to reduce overfitting. To augment our training
prototype classification. samples, we used the generative method to obtain various
morphologies based on the topological similarity. We
3. Architectural morphological pattern added random function to parameters, such as direction,
radius, edge type, angle, etc. Hence, the samples are
generation
topologically similar but different in detail in terms of
morphology, thereby ensuring the training quality. Fig. 32
In the work entitled “Architecture: Form, Space, and shows the generation logic of the five types of morpho-
Order,” Cheng (2005) elaborated on the basic principles logical patterns.
and syntax of architectural design as a classical theory in
which the architectural space is interconnected and
1. The centralized pattern consisted of multiple basic
combined into a coherent prototype, including central-
polygons arranged around the center point. The basic
ized, linear, radial, clustered, and grid (Fig. 2). It reflects
polygon was constructed from three elements: center,
the architects’ general judgment on prototype classifica-
radius, and the number of sides. The centralized
tion of architectural layouts. We used the above-
morphological patterns were obtained after the triple
mentioned five spatial prototypes in this work.
recursion of the basic polygons.
Training neural networks require adequate input data
2. The linear patterns were generated by designing the
in continuous features. We used algorithms, such as the
linear axis. The growth of the axis through the Linden-
multi-agent system (Cai et al., 2019) and L-system (Chan
mayer System (L-system) was simulated, starting from
and Chiu, 2000), which are bottom-up methods to obtain
the growth point, with three possible directions of
a community by simulating the action and the relation of
growth: forward, left, and right with certain angles. The
components. A technique called data augmentation is
growth stopped if the growth point was out of the panel
used for enlarging the dataset by artificially adding the
or too close to the original growth point. Then, the
variation of training samples to obtain a robust network
rooms were generated along the axis.
(Goodfellow, Bengio and Courville,2016). The

2
The L-system is a type of formal grammar. It has been used to describe the behavior of plant cells and applied to model the morphology
of various organisms (Source: Wikipedia).
Deep CNN for morphological classification 309

Fig. 4 Examples of generated architectural morphological patterns.

Fig. 5 Fundamental filters in convolutional operations.

3. The logic for generating radial pattern axis was similar to Fig. 4 shows examples of the five types of architectural
the centralized pattern except that the vertices were morphological patterns with the following details.
randomly removed to obtain irregular radials.
4. In the clustered pattern, the rooms and courtyards were 4. Neural network model training and
simulated by rectangular multi-intelligent agents, with
prototype classification
both attraction and repulsion forces between each
other, thus obtaining a balanced situation.
5. The grid pattern was generated through grid agents The application of the neural network model can be divided
(constructed from the growth point and direction) with into two steps. The first step involved training by simulating
three types of edges: thick, bilinear, and dotted. This the signal propagation between neurons, thus extracting
was done to ensure that each agent had eight morpho- the general features of the training sample. The second
logical possibilities. Here, the growth of the grids was step was testing in which the trained network was applied
based on the growth of the agents. to the new samples. Due to the computational mechanism
of neural networks, we do not need to explicitly define the
310 C. Cai, B. Li

Fig. 6 Overview of the morphology-to-data process for generating samples.

Fig. 7 Partial visualization of the architectural morphological datasets.

function or process for target morphology. We just need to 4.1. Morphological pattern information encoding
define the input data and output targets. In other words,
the classification of neural networks can solve such a Based on the idea of feature mapping, the image was pro-
problem when the architectural morphology cannot be cessed with a “filter operation” for feature mapping in the
exhausted under a particular classification due to their di- convolution process. The process of feature mapping can be
versity and ambiguity of definition. described as taking a certain size kernel (3  3 as an
Deep CNN for morphological classification 311

example in Fig. 5), sliding it over the image pixels, and of size 4  4 and stride 4. As the samples were in black and
performing operations, such as multiplication and accu- white, we encoded them based on a single channel, taking
mulation, to obtain a feature matrix. the brightness intensity average and finally encoding the
In this experiment, the size of the generated morphology architectural morphology information in the sample as
sample was 448  448 pixels. To make the neural network 28  28 feature mapping data (Fig. 6). We took 1000
operation more efficient and to prevent the loss of image generated samples for each type of spatial organization
features, the image information was encoded in this pattern in the experiment. The spatial prototypes were
experiment by sweeping the image in two layers with filters labeled by 0e4 according to the morphological pattern. The

Fig. 8 Structure of the original LeNet and its modification.

312 C. Cai, B. Li

Fig. 9 Gradual reduction of learning errors and recognition failure during training.

ground truth number was the goal for the neural network test data. The output of the five neurons in the output layer
output to be trained as close as possible. Therefore, the represents the probability of the corresponding prototype.
dataset featuring the architectural morphology was con- We used a deep convolution neural network, LeNet,
structed from 5000 (28  28) encoded feature data with which was designed for training with MNIST3dataset. The
labels. Fig. 7 shows a visualization of the architectural original LeNet structure has a total of 11 layers, including
morphology encoded information. two convolutional layers. The input layer size was 28  28
with one channel, as the samples were black and white
4.2. Training of LeNet images. The output layer size was 10 with one channel. As
the output layer of this experiment required 5 neurons,
The basic function for weights and deviations between instead of 10, therefore, we took 1e8 layers of the LeNet.
neural layers is expressed as formula (1), where A(n) in- Then, we added three new layers: a linear layer, a softmax
dicates neurons in the nth neural layer, W(n) indicates the layer, and an output layer, with five neurons to constitute
weight propagated between each neuron and the previous the neural network adopted for this experiment (Fig. 8).
neural layer, b(n) indicates the bias propagated between This was implemented on Mathematica (Wolfram Research,
each neuron and the previous neural layer, and X indicates Champaign), and we used the ADAM optimizer. The batch
the data in the previous layer. The backpropagation of er- size was 100. In less than 20 s, the neural network was
rors in the neural network is expressed as formula (2), trained after 25 rounds. Fig. 9 illustrates the gradual
where E indicates the error between estimation and target, decrease in recognition failure as the number of training
and s indicates the learning rate. iterations increases.

AðnÞ Z X$W ðnÞ þ bðnÞ ð1Þ 4.3. Testing of the trained neural networks for
morphological prototype classification
vE
DW Z s ð2Þ
vW A total of 40 practical building layouts were selected to test
The neural network training continuously optimizes the performance of the neural network. These samples
weights (W) and bias (b) using gradient descent and error were chosen from the reference, and the author used these
inverse propagation method, so that the distance (E) be- examples to illustrate the spatial organization prototypes.
tween the output and the target becomes progressively Therefore, each sample was labeled with the ground truth
smaller. In this way, the output gradually gets closer to the according to the book. Due to the diversity of architectural
optimization goal. The neural network testing uses the W-b morphology, some samples could be considered to belong
matrix to obtain the training step and then operate with the to more than one prototype depending on the architect’s

3
It contains thousands of black and white images of handwritten 0e9 numbers. The trained neural networks can recognize new hand-
written numbers. The accuracy of recognition is often used to measure the performance of specific neural networks.
Deep CNN for morphological classification 313

different interpretation. As a comparison study, we tested The neural network predicted the respective probabili-
the 40 samples through a simple, trained, fully-connected ties of five output neurons and ultimately took the highest
neural network with three layers (an input layer, a hidden probability as the output result. The result’s accuracy re-
layer, and an output layer) and the trained LeNet. flects the quality of the synthetic training samples and the

Fig. 10 The outputs of the simple neural network and the LeNet on the architectural morphological prototype classification.
314 C. Cai, B. Li

performance of the neural networks. The size of the neural classified into more than one prototype, which is also re-
network input and output can be modified. The results can flected in the LeNet output. For example, some architects
be used as conditions for further programs, such as concept would take case 1 as a clustered prototype, although it is
strategy and case retrieval. labeled as a centralized prototype; cases 15, 16, and 20
could be taken as a linear prototype while being labeled as
5. Results and discussion a radial prototype; and cases 24 and 25 have the possibility
of being the centralized prototype while being labeled as a
clustered prototype.
The fully-connected neural network correctly generated
Fig. 11 shows the most possible mistake of the LeNet
outputs in 25 out of 40 samples (Fig. 10). The deep con-
compared with the architect’s definition variety, taking the
volutional neural network LeNet identified 39 out of 40
eight cases with high secondary probability output as ex-
samples correctly, compared to the ground truth labels.
amples. For example, in case 1, the probability of a
The failed case involved identifying a radial prototype, case
centralized prototype is 0.51, whereas that of a clustered
19, as a linear prototype (Fig. 10). The deep convolutional
prototype is 0.47; in case 20, the probability of a radial
neural network performed better than the simple fully-
prototype is 0.62, whereas that of the linear prototype is
connected neural network, as it had a more developed
0.37. This is a surprising outcome wherein the neural net-
structure.
work’s output in terms of architectural morphological pro-
Interestingly, we can intuitively see that case 19 can be
totype shows high accuracy and similarity mistakes with
classified as a linear prototype as it shows linear
architects. Architects take more than one possibility when
morphology. Furthermore, some other samples could be

Fig. 11 Most possible mistakes of the LeNet compared with the architect’s possible classification.
Deep CNN for morphological classification 315

classifying certain samples, and the LeNet reflects a similar performance fields. The quantitative descriptions of
decision process. architectural morphology provide a baseline for in-depth
As the training samples were generated based on a building interpretation and have received attention in
specific prototype, they integrated both topological simi- morphological studies in recent years. Conventional
larity and morphological diversity simultaneously. methods based on an intuitive perspective of forms often
Compared with the conventional data augmentation consist of statistical calculations, weight decisions, and
methods, the proposed method provides a more flexible indicator selections, which are used for the integrated
and diverse enhanced data source. We transmitted the description of a building. It leads to subjective bias and may
448  448-pixel image to 28  28 feature mapping data, greatly be influenced by the researcher’s knowledge.
thus boosting the efficiency of the neural network. In the Morphological analysis is promoted by the recent success
calculation process of the LeNet, the input data were of deep learning methods in which the intrinsic features
calculated through kernels of the n  n matrix containing can be extracted/learned automatically from a large
weight values. This is a robust technique for image pro- amount of data. The morphological features are quantified
cessing. In other words, the neural network is sensitive to by high-dimensional feature data through a deep CNN,
the distribution of the pixels in the image. For these rea- which contains the overall information of the morphological
sons, the neural networks performed well in classification characteristics, rather than one-to-one correlations of the
based on the generated samples. features.
Compared with other quantitative classification ap- We applied a rule-based system to generate morpho-
proaches, encoding morphology into feature data based on logical patterns, including five spatial prototypes, to
feature mapping is a fully automated method that includes construct a synthetic training dataset. This was done to
several aspects of the morphological criteria without the supplement the insufficient architectural morphological
need to balance the weights of indicators. Training with datasets and to save efforts on preprocessing the training
synthetic data saves effort in collecting, selecting, and la- datasets. To complete the morphology-to-data trans-
beling data. The training target could also flexibly perform mission, we used the convolutional approach based on
a certain task; for example, the training labels could be pixels to quantify the morphological features. A total of
modified to represent the building energy performance, 5000 synthetic training samples served as inputs for training
structural type, etc. the networks. The performance of the proposed method
The limitations of the proposed method are related to was validated on 40 practical architectural layouts.
the highly automated process, which increases the diffi- Compared to the ground truth from the reference book, the
culties of emphasizing the influence of a certain aspect by modified LeNet provided an encouraging accuracy of 97%
adjusting the weights. First, the generated samples could (39/40), whereas that of the simple, fully-connected neural
be more detailed, and more types are needed to achieve a network was 62.5% (25/40). Interestingly, the only mistake
trained network and architectural morphology with higher of the LeNet was similar to that of some experienced ar-
accuracy and better diversity, respectively. Second, given chitects. Furthermore, the most possible mistakes of the
that the samples are image-based and only contains 2D LeNet were similar to the experienced designers’ mistakes
information, 3D information like building height and when intuitively observing the testing samples. The LeNet
building form, may be lost. This drawback could be over- output shows similarity with the architects’ definition va-
come by adding one more dimension, such as the grayscale riety on some samples. The results indicate that, by
dimension in the picture to represent the building height or developing a proper training dataset, the neural network
the use of voxels instead of pixels. Third, for other tasks output can be highly accurate while still maintaining the
that need different training samples, the generative system diversity of the morphological definition. The proposed
should be redeveloped in order to obtain task-oriented method, therefore, can serve as a basis for further archi-
synthetic data. tectural typo-morphology-related studies.
This work demonstrated the feasibility and power of
using the deep learning network in architectural
6. Conclusion morphology. The finding of this work can help promote
morphological design in the future and potentially facilitate
Computers can rapidly perform 4- or even 10-digit opera- a greater understanding of architects’ designs. Our future
tions, because they are extremely powerful in terms of work will focus on technological improvements and the
following basic instructions. However, the human brain can following application scenarios:
easily distinguish information contained in pictures, such as
faces, cats, and furniture. In comparison, such a task is
More dimensions, such as grayscale value, can be added
quite difficult for a computer (Rashid, 2016). In relation to to the data source to indicate volume heights and
this, one of the goals of neural networks is to solve non- improve the model performance.
explicit instruction problems, such as morphological
More information could be added to the data source
analysis. according to specific scenes. For example, the gener-
Morphological similarity analysis and classification ated samples could include more details about green or
represent a useful analysis framework in many studies, such vertical spaces.
as those in the typo-morphological, historical evolution,
The training targets could be modified according to
predesign contextualization, and building energy specific application scenarios, such as morphology-
 316 C. Cai, B. Li

related building performance evaluations and building Hovestadt, L., 2010. Beyond the Grid: Architecture and Information
structure types, by changing the label interpretation. Technology: Applications of a Digital Architectonic. Birkhauser,

Further, the output of the neural network could be used Basel.
as one of the input parameters for generative designs Hua, H., Hovestadt, L., Tang, P., Li, B., 2019. Integer programming
for urban design. Eur. J. Oper. Res. 274 (3), 1125e1137.
based on the rule-based system. For example, the con-
Huang, W., Zheng, H., 2018. Architectural drawings recognition
ceptual hand-drawing scripts could serve as inputs for and generation through machine learning. Paper presented at
the trained neural network to obtain similarity compar- the. In: Proceedings of the 38th Annual Conference of the As-
isons for case-based studies. With the combination of sociation for Computer Aided Design in Architecture (Mexico
generative designs, the proposed method could facili- City, Mexico).
tate the decision-making and promote further building Krizhevsky, A., Sutskever, I., Hinton, G.E., 2012. ImageNet classi-
development. fication with deep convolutional neural networks. Adv. Neural
Inf. Process. Syst. 25 (2).
Lecun, Y., Bengio, Y., Haffner, P., 1998. Gradient-based learning
Declaration of competing interest applied to document recognition. Proc. IEEE 86 (11),
2278e2324.
We declare that we have no financial and personal re- Levy, A., 1999. Urban morphology and the problem of the modern
lationships with other people or organizations that can urban fabric: some questions for research. Urban Morphol. 3,
inappropriately influence our work, there is no professional 79e85.
or other personal interest of any nature or kind in any Li, B., 2012. GenerativeDesign: Exploration of Computer-Aided
product, service and/or company that could be construed Architectural Design Methods Based on Complex Adaptive Sys-
tems. Southeastern University Press.
as influencing the position presented in, or the review of,
Li, B., Han, D., 2011. Technical comprehension of architectural
the manuscript submitted. generative design and its prospects. Architectural Journal (06),
96e100.
Acknowledgments Li, B., Zhang, J., Ludger, H., Guo, Z., 2019. Analyzing the black box
of design with algorithm model. Architect 94e99, 01.
Liu, Y., Lin, W., Deng, Q., Liang, L., 2019. Exploring the building
This research was funded by National Key R&D Program of form generation by neural networks –Taking the free university
China (2017YFC0702302 in 2017YFC0702300), National of berlin as an example. Paper presented at the Proccedings of
Natural Science Foundation of China (Project No. 51978139) 2019. In: National Conference on Architecture’s Digital Tech-
and Fundamental Research Funds for the Central Univer- nologies in Education Research.
sities and Postgraduate Research Innovation Program of Nielsen, M.A., 2015. Neural Networks and Deep Learning, vol. 2018.
Jiangsu Province (Project No. KYCX20_0110). Determination press, San Francisco, CA, USA.
Rashid, T., 2016. Make Your Own Neural Network. CreateSpace
Independent Publishing Platform.
References Rossi, A., Eisenman, P., Ghirardo, D.Y., Ockman, J., 1982. The
Architecture of the City. The MIT Press, Cambridge.
Ariga, T., 2005. Morphology, sustainable evolution of inner-urban Soddu, C., 1998. Argenia, a natural generative design. Generative
neighborhoods in san francisco. J. Asian Architect. Build E ng. art: Proceedings of GA 98, 7e41.
4 (1), 143e150. Srisuchinnawong, A., Sakulkueakulsuk, B., Chaisangmongkon, W.,
Cai, C., Tang, P., Li, B., 2019. Intelligent generation of architec- 2018. Word recognition in captured images by CNN trained with
tural layout inheriting spatial features of Chinese garden based synthetic images. Paper presented at the. In: 2018 22nd Inter-
on prototype and multi-agent system - a case study on Lotus national Computer Science and Engineering Conference
teahouse in yixing. Paper presented at the. In: Proceedings of (ICSEC).
the 24th CAADRIA Conference. Victoria University of Tang, P., Li, H., Wang, X., 2019. Generative design on conservation
Wellington, Wellington, New Zealand. and inheritance of traditional architecture and settlement
Chaillou, S., 2019. AIþ Architecture: towards a New Approach. based on machine learning: a case study on the urban renewal
Harvard University. design of roma termini railway station. Architect (1), 100e105.
Chan, C.W., Chiu, M.L., 2000. A simulation study of urban growth Tom, K., Paul, G., Anthony, S., Wonka, P., Mitra, N., 2018. Frank-
patterns with fractal geometry. Proceedings of the Fifth Con- enGAN: Guided Detail Synthesis for Building Mass-Models Using
ference on Computer Aided Architectural Design Research in Style-Synchronized GANs. arXiv preprint arXiv:1806.07179.
Asia 55e64. Yetis‚, G., Yetkin, O., 2018. A novel approach for classification of
Cheng, D., 2005. Architecture: Form, Space and Order. Tianjin structural elements in a 3D model by supervised learning. Paper
University Press, Tianjin. presented at the. In: Proceedings of the 36th eCAADe Confer-
Gil, J., Beirão, J.N., Montenegro, N., Duarte, J.P., 2012. On the ence (Poland).
discovery of urban typologies: data mining the many dimensions Yoshimura, Y., Cai, B.Y., Wang, Z., Ratti, C., 2019. Deep Learning
of urban form. Urban Morphol. 16 (1), 27e40. Architect: Classification for Architectural Design through the
Goodfellow Bengio, Y., Courville, A., 2016. Deep Learning. MIT Eye of Artificial Intelligence (Paper presented at the annual
Press, Cambridge. conference on computers).

View publication stats