From type to network: a review of knowledge representation methods in architecture intelligence design

With the rise of the next generation of artificial intelligence driven by knowledge and data, the research on knowledge representation in architecture is also receiving widespread attention from the academia. This paper sorts out the evolution of architectural knowledge representation methods in the history of architecture, and summarizes three progressive representation frameworks of their development with type, pattern and network. By searching these three keywords in the Web of Science Core Collection among 4867 publications from 1990 to 2021, the number of publications in the past 5 years raised more than 50%, which show significant research interest in architecture industry in recent years. Among them, the first two are static declarative knowledge representation methods, while the network-based knowledge representation method also includes procedural knowledge representation methods and provides a way for knowledge association. This means the network representation has more advantage in terms of the logical completeness of knowledge representation, and accounts for 67% of the current research on knowledge representation in architecture. In the context of the rapid development of artificial intelligence, this method can realize the construction of architectural knowledge system and greatly improve the work efficiency of the building industry. On the other hand, in the face of carbon-neutral sustainable development scenarios, using knowledge representation, building performance knowledge and design knowledge could be expressed in a unified manner, and a personalized and efficient workflow for performance-oriented scheme design and optimization would be achieved.


Background
Under the background of the rapid development and application of big data and artificial intelligence (AI) technology, the first generation of knowledge-driven AI and the second generation of data-driven AI are moving towards unprecedented integration, forming the third generation of both knowledge and data driven AI. Academia and social practice have widely recognized that data-driven AI technologies such as the famous deep neural network in recent years are still incapable of performing some of the daily intelligent activities of human cognition, learning, understanding, reasoning, and adaptability. This hindered the real landing and industrialization of AI; therefore, the concept of knowledge has returned to the core discussion of AI. With the boom of big data driven artificial intelligence technologies, architecture, engineering and construction industries are rapidly transforming towards digitalization and intelligence, especially the division of labor becomes more refined and the projects become more complex, knowledge and its representation methods has received considerable attention from academia.
The word knowledge refers to "the fact or situation of something that is familiar through experience or association". Accordingly, architectural knowledge also covers rational and sensible, engineering and artistic aspects of things. According to the viewpoint of systematology, only by organically integrating isolated "knowledge points" can knowledge play its role. Based on this concept, knowledge representation (KR) is to extract architectural knowledge and its association from theories and cases, and forming structured data which could be understood and analyzed by computers. This is also the primary goal of research in the new field of next-generation intelligence.
By searching keywords (knowledge representation OR pattern language OR typology OR knowledge graph OR semantic network) for certain research directions (architecture OR Construction Building Technology OR Urban Studies in the Web of Science Core Collection) among 4378 publications from 1990 to 2021, with the widespread application of artificial intelligence technology in various fields, the number of publications in the past 5 years raised more than 50%, which shows significant research interest in architecture industry in recent years ( Fig. 1).
Up to now, there is still no clear process framework in the early stage of architecture design. During this stage, students and practitioners often spend a lot of time searching for information and making inefficient adjustments. If the knowledge related to design can be integrated and presented in an axiomatic and universal mode, it will not only realize knowledge reuse, reduce unnecessary duplication of work and work platform conversion in the design, evaluation or optimization links，but can also greatly speed up the learning process and improve the operability for non-professionals, so as to realize the public attributes of architecture design. Utilizing the powerful computing power of computers to organize construction-related knowledge is an effective way to promote industry development and multi-industry exchanges and cooperation. How to do a good job of KR research in architecture will not only brings challenges, but also brings new opportunities to the field.

Type-based knowledge representationfrom taxonomy to typology
The concept of typology originated from taxonomy in natural history, its original purpose was to systematically and accurately recognize biological categories in order to carry out correlation research. With the continuous expansion of knowledge, simple classification cannot meet the needs of KR, so the concept of type has gradually been developed. When it is taught as a kind of knowledge, people only need to get the type of things through In the second half of the eighteenth century, the concept of type was used to describe "Greek perfection" in painting and sculpture (Xue, 2016). Subsequently, it was used to express "a logical merger of a conception of origin and a conception of typical form" (Vidler, 1996) in architecture. In 1825, the French artist Quatremère de Quincy composed the "Architecture Dictionary", which made an exhaustive induction of classical architectural forms from the perspective of linguistic typology, which was regarded as the beginning of systematic architectural KR as types (Xue, 2016). In the following hundred years, many scholars have carried out research on typology on the basis of Quincy's theory. Aldo Rossi regarded type as a logical principle a priori to form (Rossi, 1982); Giulio Carlo Argan regarded type as a similar system of classification (Argan, 1996). They all affirm the close relationship between type and the generation of architectural form, and give a self-consistent theoretical explanation (Wang et al., 2019).
Based on type, a large number of scholars have done the representation research and related practice of architectural typology, and provided the knowledge framework of typology representation (Wang, 2003;Wei, 1990). Table 1 shows the related work done by some scholars. Since the twenty-first century, with the rise of new regionalism, due to its favorable representational effect on systematic sorting of traditional vernacular architecture, many scholars have carried out climate and cultural adaptability assessments of traditional architecture based on architectural typology, which is a good example for contemporary architecture. Architecture design provides design templates and inspiration (Anna-Maria, 2009;Peña-Huaman et al., 2022;Rachmayanti et al., 2022).
With the wide application of AI deep neural networks, some scholars have also carried out typologybased building generation exploration. After combining type features with rules, building or city can be translated into shape and description grammar, which can realize the generation of historical buildings (Stouffs & Tunçer, 2015) (Fig. 2) and prediction of building energy consumption (Caldas & Santos, 2012); in the application of engineering projects, types can be used as preassembled components in the BIM system to improve the modeling efficiency of functional partitions of the same type (Nagakura & Sung, 2017).

Table 1 Typology representation framework
Fundamental contents (Wang, 2003) Ratio and module Spatial mode and scale

Classification principle (Wei, 1990) Urban scale and planning Building scale and construction Detail scale and decoration
Application method (Wei, 1990) Type selection Type migration conversion

Pattern-based knowledge representationfrom prototype to pattern language
Type can represent a class of buildings in a more complete way in form. However, it has been separated from architecture design, because the generalization of building types in a narrow sense cannot help us face a series of new design problems, and for different projects, we cannot achieve type reuse owing to variety of types. Therefore, in the later stage of the theoretical development of architectural typology, scholars shifted their research focus to the "prototypes". Different from the generalization of architectural form characteristics by type, prototype is more concerned with the universality of architectural knowledge system. Quincy proposed three origins of architecture with reference to the way of life of primitive society: caves, tents and huts (Clausen & Lavin, 1992); German architect Gottfried Semper proposed the famous four elements of architecture with reference to anthropology (Semper, 1989) -hearth, platform, roof and its supports, and nonstructural enclosure… In early modernist architecture, architects such as Frank Lloyd Wright and Le Corbusier all paid tribute to prototypes.
With the development of the times, the prototype, which is close to naive reductionism, cannot be applied to the ever-changing and increasingly complex architectural activities. Many scholars have improved and optimized prototype theory based on humanistic thinking, the most famous of which is the "pattern language" (Alexander, 1978) proposed by American architect Christopher Wolfgang Alexander. The theory describes a total of 253 patterns in three scales of town, building and construction. These patterns blend rationally and perceptually and together constitute the necessary operational elements for architecture design. In the book of pattern language, other scale patterns related to this mode are noted before and after each mode (Fig. 3), providing us with the necessary knowledge associations.
Because of its better representation of complex models, pattern language was first used in the development of object-oriented programming in software engineering, where they were used to refer to reusable components in engineering (Ozel, 2007). In the field of architecture, patterns can summarize design activities very well. As a method of organizing knowledge into concepts, it encapsulates many tedious operations in architecture design into modules, allowing designers to better focus on the design itself.
"Pattern Language" is not only just a manual for design, but the inherent quantitative relationship contained in the pattern has also become a powerful tool for building evaluation and generative design, and is widely used in the field of architecture, engineering and construction (AEC). In terms of design generation and optimization, Zhao Qun summarized the ecological architectural pattern language based on the cave dwelling model by studying the settlement characteristics and performance of traditional dwelling buildings in Northwest China, which provided the basis for the ecological design strategy of new dwelling buildings (Zhao, 2005); Bukovszki et al. designed a coding engine based on the concept of pattern language by investigating the sustainable lifestyles of low-income people in Hungary, which is to realize the social housing participatory design of low-income people. The marking of daily life scenes, through cluster analysis, lead to decision-making goals for oriented design that participants usually can make 45% of the decisions in early design stage (Bukovszki et al., 2021) (Fig. 4); Na et al. proposed a new method for evaluating design performance based on pattern language. This method takes the design of kitchen space as an example, they first set the parameters of some patterns to define a standard pattern, and then extract the patterns in the BIM model. Afterwards, they employ evaluation equations and weights to calculate performance scores to automate evaluations and hopefully allow design practitioners to adjust designs based on the evaluations (Na et al., 2020) (Fig. 5).

Web-based knowledge representation-from semantic web to knowledge graph
Pattern language has made a great contribution in the construction of architectural knowledge system, but it also has its limitations. It is considered that Alexander's theory contains an ontology confuses objective and subjective phenomena, and rejected rigorous scientific arguments (Dawes & Ostwald, 2017). In actual projects, the design methods contained in the pattern language cannot lead to an excellent conceptual solution. However, pattern language is also considered to be the prototype of the architectural knowledge network structure. In 1968, after J. R. Quillian first proposed the Semantic Network (SN), this concept gradually migrated to the field of architecture. SN and pattern language have certain similarities. In SN, knowledge is represented by graph: nodes are used to represent concepts or entities, corresponding to Alexander's pattern; edges represent the logical relationship, corresponding to the organization between different patterns. SN realizes knowledge reasoning through inheritance and matching methods, similar to the KR method  (Na et al., 2020) based on type and pattern, it also belongs to the explicit declarative KR method.
Due to the better articulable of complex spatial relationships, SN has many applications in large-scale landscape and urban planning analysis. Some scholars combined the related theory of space syntax to convert the garden space sequence into a SN, and analyzed the layout of traditional Chinese royal gardens such as the Humble Administrator's Garden and Yu Garden (Dong et al., 2021;Yu et al., 2016) (Fig. 6). Some scholars also constructed a planning system for urban design (Dong, 2016).
Based on SN, in 2012, Google launched a knowledge graph (KG)-based search engine product (Amit, 2012). Based on sources such as DBpedia and Freebase, KG is a knowledge base that combines the interrelated descriptions of concepts, entities and associations. Nevertheless, KG integrates knowledge mining, fusion and processing in addition to KR. Under these circumstances, a simple triple-based KR cannot satisfy such complex data structure reasoning. Therefore, the technical basis of KG also includes natural language processing (NLP), Semantic Web (SW) (Berners-Lee, 2000) and ontology-based knowledge reasoning. Their common combination can realize the close cooperation of knowledge and data, extending the application scenes of AI. In the current AEC studies, researches related to KG is still in its infancy. The knowledge mining and representation work based on KG, and the realization of human-computer interaction is the focus of the academia.
In the face of complex architectural problems, pattern language and SN for architecture design are different from the declarative and static KR methods. KG can provide a framework of procedural KR to describe architectural activities to realize design knowledge spontaneous acquisition and design behavior predictions Gao, Zhang, He, et al., 2022;Gao, Zhang, Huang, & Shi, 2022) (Fig. 7). Gao Wen et al. extracted the 3D modeling event log into a data structure called command-object graph (COG) that could automatically represent the design generation process , and used the advanced model Transformer in NLP to realize the high-accuracy sequence-based data prediction of modeling commands (Gao, Zhang, He, et al., 2022). Based on process mining technology, creative automatic analysis and humancomputer interaction for architecture design can also  (Gao, Zhang, He, et al., 2022) b Examples of randomly picked command graphs in the experiment and c clustering result of all embedded command graphs from 99 designers' logs data (Gao, Zhang, He, et al., 2022) Li et al. Architectural Intelligence (2022 . 8 a Procedures of a graph-and feature-based building space recognition algorithm (Chen et al., 2017) b Correspondence between data structure and graph c Case model and its recognition result (Li et al., 2019) d Structure of ANN model and the relationship between shape parameters and energy consumption parameters be realized. This series of studies has changed the view that the design process is unrecognizable in academia, revealing the programmability of design behavior. Based on this finding, the knowledge mining part of knowledge representation has more powerful tools, and the realization path of design automation is also clearer.
In the early stage of the architecture design, the shape of the building has a significant impact on its performance (Becker, 2008), and it is of great significance to study the knowledge association between the architectural shape and performance indicators. Recently, researchers from school of architecture, Tsinghua University, has carry out some exploring with the knowledge relationship between the architectural shape and performance indicators such as energy consumption, daylighting quality and natural ventilation performance. Chen Hongzhong, Li Ziwei et al. proposed a representation method based on graph network for the 3D buildings model. This method can identify the geometric information such as nodes and edges of the building model, and generate the proxy model required by the energy consumption model. They further established artificial neural network (ANN) models for building performance prediction, specifically on building energy consumption and daylighting through (Chen et al., 2017;Li et al., 2018Li et al., , 2019 (Fig. 8). Finally, the knowledge association between building shape parameters (such as window-to-wall ratio, storey height) and energy consumption can be realized.
In addition, He Qiushi et al. took the building proxy model and its corresponding daylighting simulation results as input and output respectively, and used ResNet (CNN) and pix2pix(GAN) to realize the fast prediction of static and annual daylight metrics and illuminance distribution in space within 1 second. This method provides a deep neural network-based knowledge mining and prediction way for building performance. (He et al., 2021) (Fig. 9). In the future design for combined performance, the prediction of lighting can effectively improve the efficiency of environmental simulation, and realize an efficient workflow for co-frequency advancement of design and simulation. Apart from that, new knowledge mining and representation strategies are also being explored. For example, Pan et al. proposed a framework of "video 2entities" based on zero-sample learning technology and general KG (Pan et al., 2021). In this study, researchers extracted unseen entities, like small construction parts or worker construction process, in architectural decoration videos. The knowledge extraction framework can not only be used for KG update in AEC fields and realize the construction automation in the near future (Fig. 10). However, this research is still in the preliminary development stage and can only be used for entity extraction, the extraction of other contents in KR system needs to be further deepened.  (Pan et al., 2021) 3 Discussion

Completeness comparison of KR
The representation methods based on type, pattern and network have their own characteristics, but in terms of the completeness of KR, network-based knowledge construction is a necessary condition for the formation of knowledge systems. In Table 2, we compare the KR methods based on type, pattern and network, and can find that network-based KR methods are more logically complete. At present, the network-based KR research also occupies a major position in the research field, accounting for 67% (Fig. 11).

Application scenarios and realization path of KR oriented architecture design
KR research are conducted in various scenes (Table 3), while focusing on cross-scenario of design and construction, design and evaluation in the field of AEC. In recent years, with the rapid development of artificial intelligence, KG, as an advanced knowledge system that combines data and knowledge, has also experienced an emerging growth in related research. Based on the summary of the application scenarios in the above table, we can also try to summarize the implementation path of using the KR method in the design (Fig. 12): (a) Type-based KR method is simple and form-based.
When applying this method to design buildings, it is important to capture key geometric features like ratio, module and spatial relationships. During the designing stage, designers can use the data extracted from the sample to compose a building similar to its form. Therefore, in addition, this method is not fully applicable to the optimization of building. (b) Pattern-based KR method is not limited to the learning of buildings form. Designers acquire design knowledge by learning a macro design strategy. For instance, the atrium in buildings can be regarded as a pattern in design. Designers can insert the atrium into the building as a component, and establish the connection with other evaluation parameters, so as to realize the coordination of design and optimization. (c) Network-based KR method has better extrapolation capabilities in architecture design. Taking floor plan design as an example, designers can recognize Fig. 11 The proportion of researches on typology, patterns or networks in architectural KR Architectural geometric features based on shape grammar/ description grammar Pattern-based KR Design language mining (Zhao, 2005) Architectural form features/Settlement characteristics/Design pattern language Design evaluation and public participation (Bukovszki et al., 2021;Na et al., 2020) Design pattern language People's subjective evaluation of design Network-based KR Landscape and urban planning analysis evaluation (Dong, 2016;Dong et al., 2021;Yu et al., 2016) Semantic network Design flow prediction Gao, Zhang, He, et al., 2022;Gao, Zhang, Huang, & Shi, 2022) Knowledge graph/Modeling event log Performance prediction (Chen et al., 2017;He et al., 2021;Li et al., 2018Li et al., , 2019 Knowledge graph/Spatial network and performance parameters Knowledge graph update (Pan et al., 2021) Knowledge graph/Building workflow based on the video each space as a node and the spatial connection as an edge. The graph network established based on the spatial layout can be applied to a wider range of design scenarios.

Conclusion
In this paper, KR methods in the fields of architecture, built environment, urban research and other fields are sorted out from the type, to the pattern, and then to the network. However, in the work of KR, there is no general method that can fully represent the huge architecture design knowledge system. When facing different application scenes, it is necessary to choose and coordinate methods according to the knowledge attributes and application purposes. In the field of architecture, due to the high ubiquity of the network, the KR method based on the network has a more in-depth development. This is also a hot issue for future architectural KR work.
In the long course of scientific development, the acquisition and imparting of knowledge is the source of the continuation and progress of human civilization. As a practical discipline that studies human construction activities and improves the living environment, architecture has a huge and complex knowledge system, which is difficult to be unified. But for many decades, efforts to standardize, scientific and systematize architecture have never stopped. With the widespread application of AI in the field of AEC, architecture, which is an open and complicated macro-system, is also full of vigor. For the KR research in architecture, with the help of AI deep neural network and big data, it can not only help us extracting and summarizing massive architecture-related content in the network to construct knowledge associations. Furthermore, AI combining knowledge and data has great potential in design generation, design decision-making and even design evaluation.
At present, most of the related research on architectural KR is still at the theoretical level, and many related studies are not aware of the attributes of its role in KR; in addition, in the actual architecture design project, simple knowledge representation is not enough, and it is necessary to build an intelligence system for knowledge processing and invocation, so as to truly improve the efficiency of design activities; in the context of carbon neutrality, KR has great potential for performance assessment and optimization of buildings. Through the unified expression of building performance knowledge and design knowledge, a performance-oriented building design optimization vertical domain knowledge map can be constructed, which can truly realize the personalized optimization and interactive heuristic design of the scheme, and provide a scientific perspective and implement logic for design strategies of green and zero-carbon buildings.