Data-driven design aids

Various data-driven methods and tools have been introduced to support engineering design ideation. For instance, FuncSion (Chakrabarti and Tang, Reference Chakrabarti and Tang1996) generates solutions for a specific set of functional requirements, using its database of functional elements and their relations. VISUALIZEIT (English et al., Reference English, Naim, Lewis, Schmidt, Viswanathan, Linsey, McAdams, Bishop, Campbell, Popna, Stone and Orsborn2010) uses a database of component flow graphs, which are created by applying graph grammar rules to functional graph models, to support concept generation. AskNature (Deldin and Schuknecht, Reference Deldin, Schuknecht, Goel, McAdams and Stone2014) is a web-based database of biological systems, which are organized by their biomimicry taxonomy, for the interest of biologically inspired designs. IDEA-INSPIRE (Chakrabarti et al., Reference Chakrabarti, Sarkar, Leelavathamma and Nataraju2006) represents both natural and artificial systems based on the SAPPhIRE (State-Action-Part-Phenomenon-Input-oRgan-Effect) ontology to support the search for natural or artificial solutions to a given design problem.

A growing stream of research has proposed methods to represent and retrieve patent documents as potential sources of design inspiration. In particular, function-related representations have attracted the greatest attention. For instance, Russo et al. (Reference Russo, Montecchi and Liu2012) adopted the Function–Behavior–Structure ontology (Qian and Gero, Reference Qian and Gero1996) to construct systematic queries to retrieve state-of-the-art patents for a specific design problem using Subject–Action–Object (SAO) structures. Fu et al. (Reference Fu, Cagan, Kotovsky and Wood2013a) used latent semantic analysis (LSA) of function verbs in patent texts to associate patents in a Bayesian network to guide the selection of patents as design stimuli. Murphy et al. (Reference Murphy, Fu, Otto, Yang, Jensen and Wood2014) represented functional aspects of designs in patents using the Bag-of-Words approach and mapped them with a vector space model to aid the search for functionally analogical patents to a specific query. Likewise, Liu et al. (Reference Liu, Li, Xiong and Cavallucci2020a) proposed a framework of representing patents with a bag-of-words approach using functional categories and classifying them with respect to the topmost hierarchy of the functional-basis (Stone and Wood, Reference Stone and Wood2000).

Some other methods and tools represent patents beyond functions. Mukherjea et al. (Reference Mukherjea, Bamba and Kankar2005) introduced the BioMedical Patent Semantic Web by extracting the biological terminology in patent abstracts in the biomedical field and then associating patents by utilizing the knowledge from biomedical ontologies. Berduygina and Cavallucci (Reference Berduygina and Cavallucci2020) mined the dependency structures between individual patent claims and linguistic features in claim jargon to support the use of the theory of inventive problem solving (TRIZ) (Altshuller and Altov, Reference Altshuller and Altov1996). Lee et al. (Reference Lee, Kang and Shin2015) constructed a morphological matrix for a specific technology based on patent metadata and keywords to identify novel patents for uses in technology opportunity analysis. Song et al. (Reference Song, Kim and Lee2017b) focused on the classification labels of patents in the Japanese Patent Classification System to discern patents for technology opportunity discovery. InnoGPS (Luo et al., Reference Luo, Sarica and Wood2019) uses a technology domain network map based on the international patent classification system from information science research (Alstott et al., Reference Alstott, Triulzi, Yan and Luo2017; Yan and Luo, Reference Yan and Luo2017a) to guide the search and retrieval of design documents and concepts across domains to inspire creative design concept generation via design analogy and synthesis.

A growing number of design ideation methods and tools have utilized large public semantic networks, such as WordNet and ConceptNet, rather than patent databases, as the backend knowledge base. WordNet (Miller et al., Reference Miller, Beckwith, Fellbaum, Gross and Miller1990) has been the most popularly utilized. For instance, WordTree (Linsey et al., Reference Linsey, Markman and Wood2012) uses brainstorming sessions and the WordNet's hierarchical structure to populate a tree structure, in which functional aspects of the design problem are represented with new additional verbs, to further guide the search for analogical solutions. Yoon et al. (Reference Yoon, Park, Seo, Lee, Coh and Kim2015) proposed a method to discover patents according to their function similarity assessed by leveraging WordNet's hierarchical structure. Georgiev and Georgiev (Reference Georgiev and Georgiev2018) developed WordNet-based metrics to measure divergence, polysemy, and creativity of the ideas from concept generation sessions. Nomaguchi et al. (Reference Nomaguchi, Kawahara, Shoda and Fujita2019) proposed to evaluate the novelty of function combinations in design ideas based on semantic similarities in WordNet, a word2vec model trained on Wikipedia, and these two metrics together, and reported a negative correlation between the human evaluations of novelty and the semantic similarity of the combined functions.

Other than WordNet that was collectively built via direct human efforts, a few other free online knowledge bases have also been employed in new design ideation methods or tools. For example, Chen and Krishnamurthy (Reference Chen and Krishnamurthy2020) proposed an interactive procedure to retrieve words and terms in ConceptNet to inspire designers. ConceptNet (Speer et al., Reference Speer, Chin and Havasi2017) is a large public knowledge graph automatically extracted from Wikipedia, built and maintained at MIT Media Lab. Han et al. (Reference Han, Forbes, Shi, Hao and Schaefer2020a) also proposed to evaluate new ideas by measuring the semantic similarity between design concepts using ConceptNet. Camburn et al. (Reference Camburn, He, Raviselvam, Luo and Wood2020) proposed a set of new metrics for automatic evaluation of the natural language descriptions of a large number of crowdsourced design ideas, and their evaluation was based on the Freebase (Bollacker et al., Reference Bollacker, Evans, Paritosh, Sturge and Taylor2008), another large public structured knowledge database managed by Google.

The growing uses of such public semantic network databases in the development of design ideation support tools have inspired the development of semantic networks based on engineering data. For instance, Shi et al. (Reference Shi, Chen, Han and Childs2017) and Liu et al. (Reference Liu, Wang, Li and Liu2020b) proposed the use of semantic networks mined from scientific papers as sources of inspiration for design concept generation. Chen et al. (Reference Chen, Wang, Dong, Shi, Han, Guo, Childs, Xiao and Wu2019) utilized the semantic concept network from Shi et al. (Reference Shi, Chen, Han and Childs2017) to retrieve implicit and explicit design stimulation at the semantic level and supported these stimuli with image synthesis to stimulate designers in ideation sessions.

On the other hand, a few recent studies employed the whole patent database instead of focusing on a specific domain and generated network maps to resemble technology space where all patent classes in the existing patent classification system are operationalized as nodes (Alstott et al., Reference Alstott, Triulzi, Yan and Luo2017; Yan and Luo, Reference Yan and Luo2017b). However, these studies are limited by the structure of patent classification systems and patent metadata; thus, they can only support high-level inspirations (Luo et al., Reference Luo, Song, Blessing and Wood2018). Furthermore, the studies on providing semantic-level design stimulation and evaluation generally rely on common-sense knowledge bases, such as WordNet and ConceptNet, or language models not trained specifically for engineering. In fact, the engineers’ perception of technical terms is biased and represented better by knowledge bases that are specifically trained on technological knowledge (Sarica et al., Reference Sarica, Luo and Wood2020).

Semantic networks

We intend to use a semantic network of technical terms, which covers generic engineering design concepts in all domains of technology, to support technical design inferences for idea generation. The past two decades, but especially the last decade, witnessed great progress in natural language processing (NLP) that allows researchers to introduce a few large-scale semantic networks based on generic human knowledge. For instance, WordNet (Miller et al., Reference Miller, Beckwith, Fellbaum, Gross and Miller1990), ConceptNet (Speer et al., Reference Speer, Chin and Havasi2017), never-ending language learning (NELL; Mitchell et al., Reference Mitchell, Cohen, Hruschka, Talukdar, Yang, Betteridge, Carlson, Dalvi, Gardner, Kisiel, Krishnamurthy, Lao, Mazaitis, Mohamed, Nakashole, Platanios, Ritter, Samadi, Settles, Wang, Wijaya, Gupta, Chen, Saparov, Greaves and Welling1998), and Yago (Rebele et al., Reference Rebele, Suchanek, Hoffart, Biega, Kuzey and Weikum2016) are available for public uses and have enabled various applications, such as information retrieval, semantic analysis, knowledge exploration and discovery, and artificial intelligence (AI), in many different fields. These knowledge databases are often generated by linguistically and statistically learning the information in knowledge repositories such as Wiktionary and Wikipedia, and merging the knowledge entities, and relations based on readily available or collaboratively created ontologies.

Despite these efforts on creating large-scale semantic networks for generic uses, only a few engineering researchers have carried out the same task in the context of engineering. For instance, Shi et al. (Reference Shi, Chen, Han and Childs2017) mined and analyzed 20 years of publications, nearly 1 million engineering papers, from ScienceDirect and 1000 design posts from blogs and design websites to extract technical terms and construct a large-scale semantic network. Despite the great amount of data collected, it is unclear if their data of different types can resemble comprehensive engineering knowledge in all domains. Liu et al. (Reference Liu, Wang, Li and Liu2020b) proposed a method to create a concept network by mining the concepts from the technical documents related to a specific design problem and associating them via the vector representations of these concepts by utilizing a word-embedding algorithm and synset relations of the WordNet. For the same purpose of creating a large-scale semantic network in the context of engineering and technology, we directed our attention to the patent data that provide more comprehensive coverage of the engineering knowledge base.

Patents are reliable and rich sources of engineering design knowledge (World Intellectual Property Organization, 2004; Asche, Reference Asche2017) since they have been examined rigorously to ensure their adequateness on defining the invention, their novelty, and usefulness. These characteristics of the patent examination process assure the quality of the patent data and avoid redundant designs to be documented in the patent database. Particularly, United States Patent and Trademark Office (USPTO) patent database is one of the largest patent databases, publicly available, organized in systematic catalogues of the inventions, and continuously evolving as inventors file patent applications and new technologies emerge.

In our recent work (Sarica et al., Reference Sarica, Luo and Wood2020), we have constructed a TechNet that consists of more than 4 million technology-related terms, which represent technical concepts in all domains of technology, and their semantic distance. The complete digitalized USPTO patent database from 1976 to October 2017 was utilized to construct TechNet. The utilization of the complete patent database was aimed to ensure the comprehensiveness of TechNet and the balanced coverage of knowledge in all domains of technology. In a benchmark comparison with other existing semantic network databases including WordNet, ConceptNet, and B-link (Shi et al., Reference Shi, Chen, Han and Childs2017), TechNet presented superior performances in terms of retrieval and inference tasks in the specific context of technology and engineering (Sarica et al., Reference Sarica, Luo and Wood2020). TechNet has been utilized to augment patent search (Sarica et al., Reference Sarica, Song, Low and Luo2019a), technology forecasting (Sarica et al., Reference Sarica, Song, Luo and Wood2019b), and idea evaluation (Han et al., Reference Han, Forbes, Shi, Hao and Schaefer2020a, Reference Han, Sarica, Shi and Luo2020b).

To construct TechNet, we first mined the raw patent texts to exact the terms (words and phrases) that represent meaningful engineering concepts (e.g., functions, components, configurations, and working principles), using NLP techniques for phrasing, denoising, lemmatization, and so on. On this basis, word-embedding models were trained (also in the total database) to project the terms to vectors and form a unified vector space that would represent the total engineering knowledge space. Then, a large TechNet can be forged by associating the technical terms by pairwise vector cosine similarity. The TechNet construction procedure is data-driven and unsupervised and summarized in Figure 1. Interested readers can refer to our prior publication (Sarica et al., Reference Sarica, Luo and Wood2020) for further details.

Fig. 1. A summary of the TechNet construction process (Sarica et al., Reference Sarica, Luo and Wood2020).

In the present paper, we introduce a methodology that utilizes the TechNet as a backend infrastructure to support idea generation. In contrast, prior studies have either used comprehensive common-sense knowledge databases or field-limited datasets to provide semantic-level aids for engineering design ideation generation. The novelty of the study is in the use of a comprehensive technology-focused semantic network trained from the technical patent database to aid in technical idea generation.

Step 1: concept retrieval in a focal design domain

The first step is to use TechNetFootnote ¹ to retrieve the terms in the technical documents representing the prior designs in a domain (i.e., identify the terms that are in both the technical documents and TechNet). These terms represent the technical concepts (e.g., functions, components, structures, and working principles) that have already been used in the designs of the focal domain to date. These prior documents need to be identified and curated first and have complete coverage of the prior designs of the domain. While we mine patent documents only in the case study of the present paper, other types of design documents (e.g., technical reports and engineering paper publications) may also be useful for the same purpose.

These concepts can be assessed according to their importance with regards to the focal design domain. For this purpose, we propose the following term-domain importance metric (tdi _t) by considering both a term's occurrence frequency in the focal domain and its specificity to the domain:

(1)$$tdi_t = tf_{t, D}\ast ts_{t, D}$$

(2)$$tf_{t, D} = \displaystyle{{\vert {\{ {d\in D\colon t\in d} \} } \vert } \over {\vert {\{ {d\colon d\in D} \} } \vert }}$$

(3)$$ts_{t, D} = \displaystyle{{\mathop \sum \nolimits_{d\in D} {\rm count}( {t, \;d} ) } \over {\mathop \sum \nolimits_{d\in P} {\rm count}( {t, \;d} ) }}$$

where t refers to a single term, d refers to a patent document, P stands for the set of patents in the whole patent database, and D refers to the set of patents in a design domain. This metric favors the terms that are more common in the domain but not as common in the entire patent database.

Term-domain frequency (tf _t,D) (0,1] is measured as the number of patents in the domain that contain the term relative to the total number of patents in the domain. It indicates how commonly a concept is used in the prior designs of the domain of interest. However, a commonly occurring concept in a domain might be a general one for all domains. Thus, we also incorporate the term specificity (ts _t,D) (0,1] metric, which is measured as the count of the term in the patents in the domain relative to the total count of the term in the total database. High term specificity indicates that a concept is specific to the design domain and may contain domain-specific characteristic design information.Footnote ²

Step 2: concept inferences beyond the domain

The second step is to infer, according to the term-to-term semantic similarity information in TechNet, from the terms in the prior design documents of the focal domain to additional technical terms that have not appeared in the prior design documents of the focal domain. The total of these additional terms forms the total white space to the focal design domain. Hereafter, we call those terms outside the focal domain the white space terms and the concepts they represent the white space concepts. The white space concepts have varying latent relevance to those already used in the original design domain.

To measure such a technical relevance (R _i), we assess the semantic relevance of the white space terms to those in the documents of the focal domain and calculate it as a potential white space term's weighted average semantic relevance to the terms in the focal domain:

(4)$$R_i = \displaystyle{1 \over n}\mathop \sum \limits_{t = 1}^n w_{i, t}tdi_t$$

(5)$$w_{i, t} = \displaystyle{{\mathop \sum \nolimits_{\,j = 1}^n v_{i, j}v_{t, j}} \over {\sqrt {\mathop \sum \nolimits_{\,j = 1}^n v_{i, j}^2 } \sqrt {\mathop \sum \nolimits_{\,j = 1}^n v_{t, j}^2 } }}$$

where n is the number of unique terms in the focal domain; w _i,t is the semantic similarity between the ith new term and the tth domain term; v _i is the vector representation of term i; and tdi _t is the weighting factor, term-domain importance, defined by Eq. (1) above.

On this basis, we further standardize the R metric by the z-score formulation as follows:

(6)$$ZR_i = \displaystyle{{R_i-\mu _R} \over {\sigma _R}}$$

where μ _R is the mean and σ _R is the standard deviation of the relevance scores (R) of white space concepts. A negative ZR score of a concept implies that it is more distant to the focal design domain than average white space concepts; vice versa.

Step 3: idea generation

Then, it comes to the third step, in which the designers make attempts to relate the previously unused white space concepts with the original design domain to generate new design ideas. Since the white space concepts, by definition, have not been previously used in the designs of the focal domain, new ideas that relate and synthesize them with the focal domain naturally derive novelty. Therefore, our proposed procedure in principle ensures the novelty of new ideas when they are generated. In particular, the relevance score (R) or its normalized form (ZR) may further guide the search and retrieval of the white space concepts to relate to the focal design domain for potential syntheses into new design ideas.

According to the extensive design creativity literature (Gentner and Markman, Reference Gentner and Markman1997; Ward, Reference Ward, Holyoak, Gentner and Kokinov1998; Christensen and Schunn, Reference Christensen and Schunn2007; Tseng et al., Reference Tseng, Moss, Cagan and Kotovsky2008; Chan et al., Reference Chan, Fu, Schunn, Cagan, Wood and Kotovsky2011, Reference Chan, Dow and Schunn2015; Fu et al., Reference Fu, Chan, Cagan, Kotovsky, Schunn and Wood2013b; Chan and Schunn, Reference Chan and Schunn2015; Song et al., Reference Song, Srinivasan and Luo2017a; Srinivasan et al., Reference Srinivasan, Song, Luo, Subburaj, Elara, Blessing and Wood2018), near stimuli to the target design domain can stimulate more ideas and more feasible ideas (Gick and Holyoak, Reference Gick and Holyoak1980; Weisberg, Reference Weisberg2006; Fu et al., Reference Fu, Chan, Cagan, Kotovsky, Schunn and Wood2013b; Chan et al., Reference Chan, Dow and Schunn2015; Keshwani and Chakrabarti, Reference Keshwani and Chakrabarti2017; Srinivasan et al., Reference Srinivasan, Song, Luo, Subburaj, Elara, Blessing and Wood2018), whereas far stimuli may stimulate fewer ideas and ideas with high infeasibility and abstractness, but give rise to the novelty of the generated ideas (Gentner and Markman, Reference Gentner and Markman1997; Ward, Reference Ward, Holyoak, Gentner and Kokinov1998; Tseng et al., Reference Tseng, Moss, Cagan and Kotovsky2008; Srinivasan et al., Reference Srinivasan, Song, Luo, Subburaj, Elara, Blessing and Wood2018). Therefore, when searching and choosing white space concepts as potential design stimuli, one may focus on the near-field concepts with high R values (e.g., ZR > 0) for ideation productivity, but anticipate common and non-surprising ideas. Alternatively, one may focus on the far-field concepts with low R values (e.g., ZR < 0) for the interest of generating novel ideas, but expect a lower chance of idea generation success and infeasible or vague ideas. Likewise, newly generated ideas can be instantly evaluated and compared in terms of novelty and feasibility according to how distant the adopted white space concepts are from the original domain, following the theories.

In particular, TechNet serves as the knowledge base and digital infrastructure for the methodology and workflow above. TechNet used in this study has 4,038,924 technology-related terms and roughly 8 × 10¹² undirected quantified semantic relevance values between each possible pair of terms and is larger than WordNet of 155,236 entities and 647,964 relations and ConceptNet of 516,782 entities and 1.3 × 10¹¹ relations. To the best of our knowledge, it is the largest technology-related semantic network to date. TechNet terms cover all domains of engineering defined in the Cooperative Patent Classification System. In particular, the distribution of terms is highly correlated with the distribution of patents across different technology domains, suggesting proportional and balanced coverage of relatively large or small domains (Sarica et al., Reference Sarica, Luo and Wood2020).

Thus, for a specialized design domain, TechNet enables the exploration in a sufficiently wide engineering knowledge space beyond the focal domain itself for the discovery of white space concepts for creative synthesis with those within the domain. Now, we apply the TechNet-based methodology to generating flying car design ideas.

Retrieving used concepts in prior flying car designs

We conducted an exhaustive search for flying car patents in the USPTO patent database that resulted in a set of 164 flying car patents between 1974 and 2018. The exhaustive search process combined patent text mining, citation analysis, inventor, and classification relationships to retrieve relevant patents. In particular, the set of relevant patents was expanded by employing an iterative process involving heuristic learning from patents retrieved in each iteration. Each iteration involves validity checks and expands query keywords, the citation network, and the set of inventors until convergence to a fixed set of patents. Hence, the whole process ensures the completeness and accuracy of the retrieved patent set (Song and Luo, Reference Song and Luo2017).

2396 unique technical terms in TechNet are retrieved from the titles and abstracts of the flying car patents. These retrieved terms cover various types of engineering concepts, such as working principles, main functions, sub-functions, structures, configurations, physical subsystems, modules, and components of prior flying car designs. Table 1 reports the top 30 terms with respect to tdi scores based on Eq. (1).

Table 1. Top 30 concepts already used in existing flying car designs

Some of these terms are synonymous notions of the flying car such as “aerocar” and “roadable aircraft”, which directly indicate inventors’ design tendencies, for example, a car that can fly or a plane that can be driven on roads. A subset of terms refers to specific concepts that define the working principles of flying cars in both air and road environments, such as “upward lift force”, “variable-environment”, and “ground propulsion”. Such terms as “center of gravity”, “telescopic wing”, “nose-height”, and “directional flight propulsion” point out the problem or solution foci of designers. The “center of gravity” of the vehicle becomes a challenging issue for dual-mode operations. “Nose-height” is an important parameter in airplane docking systems, which will be also an issue for flying cars. “Directional flight propulsion” is a concept used in both VTOL and hovering. Moreover, many terms in the list refer to physical components (e.g., “payload clamp”, “multiple ducted fan”, and “ground module”) and desired functional performances (e.g., “lowered energy consumption”, “larger lift”, “powerful suction”, “faster flight”, “advanced vertical agility”, and “easily convertible”).

In brief, these terms allow us to develop a quick understanding of frequently used technical concepts in the existing designs of flying cars.Footnote ⁶ In other words, these terms succinctly describe a domain-specific concept space of prior flying car designs. They serve as the starting and reference point for the inferences to and selection and evaluation of other white space concepts beyond prior flying car designs for new flying car design ideation.

Retrieve white space concepts

There are more than 4 million technical concepts in TechNet, whereas 2396 of them have already been used in prior flying car designs. Thus, there still exist enormous design opportunities via the synthesis of white space concepts with flying cars. In the meantime, the large size of the white space also demands structured guidance for the search and retrieval of concepts. Here, we focus on semantic distance as the guiding variable to retrieve white space concepts from the near-field to far-field to prior flying car designs as potential ideation stimuli and make attempts to relate and synthesize them with flying cars as the way to generate new design ideas. Based on design creativity theories, white space concepts with relatively high latent relevance to flying car designs are more likely to stimulate ideas and more feasible ideas, whereas the far-field concepts are anticipated to be less effective in stimulating new ideas but may contribute to greater novelty of the new ideas once they are generated.

We first retrieve the top 100 most relevant white space terms to each of the 2396 terms in the flying car patent set. The total set consists of 107,529 unique terms.Footnote ⁷ The ZR scores of these terms lie in the range of [−4.449, 7.565] where 89% of them have ZR > 0. The ZR score of each term is calculated, based on Eq. (6), against a random sample of 100,000 white space terms with a mean (μ _R = 0.2646) and standard deviation (σ _R = 0.0660) in the distribution of their R _i scores defined in Eq. (4). We consider this set of concepts approximate a relatively large and diverse but still confined “near-field” to the focal flying car design domain in the total TechNet. Note that, alternative white space term discovery methods can be adopted for this step as well.Footnote ⁸ Table 2 lists the top 30 white space concepts according to their ZR scores, which fall in the range of [6.460, 7.565]. We will use this set of concepts as the nearest stimuli for idea generation later.

Table 2. The ideas generated with the nearest 30 concept stimuli surrounding prior flying car designs

The table is sorted according to the ZR scores of the stimuli, whereas the sequence of the provision of the concepts to the engineers was randomized during the ideation process.

Next, we divide the total set of 107,529 terms into 10 equal-size quantiles from high to low ZR, and randomly sampled 10 terms in each quantile. This results in a set of 100 randomly selected near-field concepts in Table A1 in Appendix. This second set of white space concepts may provide a balanced coverage of a large and diverse near-field surrounding existing flying car designs. Their ZR scores lie in [−0.812, 5.964] and all are below the ZR scores of the nearest 30 concepts in the first stimuli set (Table 2).

On this basis, we further randomly sample 100 concepts from the total TechNet regardless of their semantic distance to the flying car domain (Table A2 in Appendix). Their ZR scores range in [−1.673, 4.311]. Figure 4 compares the distributions of these two randomly sampled sets of 100 white space concepts by ZR scores. The mean ZR score of the 100 near-field random concepts is 1.6 and greater than 0, whereas the totally random concept set exhibits a nearly symmetrical distribution with a mean ZR of −0.04.

Fig. 4. Distributions of the stimuli by semantic distance to existing flying car designs, including 100 concepts randomly selected from the nearest 107,529 concepts (red) and the 100 concepts randomly selected from the entire TechNet (textured).

In brief, the first two sets of white space concepts are retrieved by preferring technical concepts near those already used in existing flying cars, and many of them are from either car or aircraft domains (based on Tables 2 and Table A1 in Appendix). In contrast, the third set of white space concepts are retrieved without any constraint with regard to their semantic distance to flying cars and may come from any domain in the total technology space. Although the second set is also randomly sampled, the random draws were within a confined near-field. Figure 5 summarizes the relations of the three sets of white space concepts resulting from different retrieval strategies.

Fig. 5. The three stimuli sets.

Idea generation with white space concepts as stimuli

Then, these three sets of 30, 100, and 100 white space concepts are used as inspirational stimuli for generating new flying car design ideas. Three of the four coauthors of the paper, with basic engineering knowledge of flying cars, went through each concept in the three stimuli sets one by one and made attempts to apply the ideation heuristic of relating and synthesizing each white space concept with flying cars to generate new design ideas. We spent 30–45 min on each set of stimuli to generate ideas individually.

Summary of the case study

The three ideation runs demonstrate three different strategies to retrieve white space concepts as stimuli to inspire idea generation for a focal design domain. The first strategy favors the extremely nearest stimuli. The second strategy allows sampling a wider spectrum of stimuli in terms of their semantic distance to the focal domain, but still favors the near-field. The third strategy is a total random retrieval regardless of the semantic distance of stimuli to the focal design domain. The fact that these different strategies all retrieved inspirational stimuli shows the general effectiveness of the TechNet for idea generation.

At the same time, the idea generation efficacy, and the novelty and feasibility of the ideas generated vary with respect to the stimuli from different strategies. With the nearest stimuli, the first exercise had the highest stimulation efficacy, indicated by the portion of provided stimuli that invoked ideas. Both the first and second exercises favor near-field stimuli, despite the different extents present higher stimulation efficacies than the third exercise when the stimuli were randomly drawn from the large TechNet of more than 4 million concepts.

While almost all the ideas generated in the first exercise are highly feasible given the nearest stimuli that appear mostly from the car and aircraft domains, these ideas are also not surprising or novel. In contrast, by retrieving and providing stimuli with a wider range of semantic distance to the focal domain, the second and third exercises allowed us to generate more novel ideas with varied feasibility. In general, the ideas generated using highly relevant (i.e., positive and high ZR values) stimuli also appear to be highly feasible. While the ideas generated with semantically distant stimuli (i.e., negative and low ZR values) are naturally more novel, they also appear to be less feasible or more vague.

The stimuli set of the third exercise included farther stimuli and more stimuli with ZR < 0 than the stimuli set of the second exercise (Fig. 4). While the concepts from far fields may unleash the potential for novel idea generation, the engineers find it difficult to conceive their relevance to the original domain. As reported in Table A2 in Appendix, we attempted but failed to make sense of most of the far-field stimuli with ZR < 0. We were unable to obtain inspiration from any (and a large number) of far stimuli with ZR < −0.945 (Fig. 6b). By contrast, in the second exercise, we only retrieved a small number of far stimuli with ZR < 0 (with all of them having ZR > −0.815) but were able to find inspiration from most of them (Fig. 6a). These results suggest, despite the need to explore far stimuli for potential novel ideas, overly far concepts in the white space may be ineffective for idea stimulation. A balanced stimuli retrieval strategy with regards to the semantic distance from the stimuli to the original domain is needed.

Fig. 6. Distribution of ZR _i of white space stimuli in the (a) second and (b) third exercises. Red histograms denote the distribution of all stimuli retrieved and provided to engineers. Textured histograms represent the distribution of the effective stimuli that invoked ideas.

These findings from our case study resonate with the prior studies on design stimulation that have suggested near-field stimuli are more effective to inspire designers and produce more feasible ideas (Fu et al., Reference Fu, Chan, Cagan, Kotovsky, Schunn and Wood2013b; Chan et al., Reference Chan, Dow and Schunn2015; Srinivasan et al., Reference Srinivasan, Song, Luo, Subburaj, Elara, Blessing and Wood2018; Goucher-Lambert and Cagan, Reference Goucher-Lambert and Cagan2019), whereas far-field stimuli are less effective but may contribute to novelty of the ideas once they are generated (Gentner and Markman, Reference Gentner and Markman1997; Ward, Reference Ward, Holyoak, Gentner and Kokinov1998; Chan et al., Reference Chan, Dow and Schunn2015; Luo et al., Reference Luo, Song, Blessing and Wood2018; Srinivasan et al., Reference Srinivasan, Song, Luo, Subburaj, Elara, Blessing and Wood2018). Observing such tradeoffs, Fu et al. (Reference Fu, Cagan, Kotovsky and Wood2013a, Reference Fu, Chan, Cagan, Kotovsky, Schunn and Wood2013b) hypothesized that the stimuli from the “middle ground” may be desirable. He and Luo (Reference He and Luo2017) suggested that the most valuable inventions are based on mainly conventional combinations of prior work and a minor insertion of highly novel combinations.

In practice, different designers may follow these theoretical understandings to explore, retrieve, and mix design stimuli of varied semantic distance to a focal design domain according to their differed preferences in ideation effectiveness and idea feasibility versus idea novelty and radical innovation potential. In particular, the quantified semantic distance (R or ZR scores) between white space concepts and the focal design domain in the large TechNet provides a new basis for guiding the search, retrieval, and selection of near to far stimuli to support design idea generation and evaluation.