Design exploration predicts designer creativity: a deep learning approach

Abstract

This study examined brain activation in graphic designers responding to pictorial stimulation during exploration tasks and determined the predictive effects of design exploration on designer creativity through a deep learning approach. The top and bottom 25% (10 each participants) were assigned high-creativity and low-creativity groups, respectively. The results provided the following indications. (i) Shallow architectures had higher prediction accuracy than deeper architectures. (ii) The prediction accuracy of shallow long short-term memory networks was higher than that of convolution neural networks. (iii) Bandpower exhibited increased prediction accuracy, and shallow LSTM networks with differing power spectra among independent components outperformed other deep learning methods. (iv) Direct acyclic graph networks did not improve prediction accuracy. (v) Design exploration could effectively predict designer creativity.

Introduction

Design is an iterative interplay between a particular problem and its solutions and alternatives. The characteristics and limitations of a current solution can evolve to become new criteria for redefining a problem and assist in generating new design strategies. This process is referred to as design exploration (Maher et al. 1996) and often occurs during conceptual design. Scholars have perceived design exploration as the coevolution of problems and solutions (Cross 1997; Dorst and Cross 2001; Studer et al. 2018) and have suggested that it has an essential role for inspiring creative performances and outcomes (Atman et al. 1999; Gero 2004). However, few studies have provided evidence of an association between design exploration and designer creativity (Goldschmidt 2011; Studer et al. 2018), and the current study aims to fill this gap in the literature.

Literature review

Design exploration and designer creativity

Design exploration is the process of generating and evaluating design alternatives that have not yet been considered (Navinchandra 1991, p. 67). This definition emphasises the focus on alternatives and reducing limitations that enables new dimensions to emerge for designers to consider. Taking into account the ill-structured nature of design problems, scholars have held that design exploration, which precedes research, involves the construction and incremental refinement of design problems and solutions (Gero 1993). In other words, design exploration precedes the initial phase in which designers become acquainted with a given problem by researching and evaluating relevant information. In the exploration phase, Goldschmidt (2005) indicated that designers test their hypotheses with questions. Fallman (2008) similarly noted that designers in this phase often assess their ideas by asking ‘What if?’ and while concurrently considering alternatives to transcend existing paradigms, are proactive and emphasise societal characteristics in their expressions.

Design knowledge, novel views, and project freedom may result in various solutions, leading to the variety and flexibility of designs. This variety and flexibility, generated through design exploration, is essential to developing radically innovative solutions (Clevenger et al. 2010; Navinchandra 1991). Capable designers conduct exploration activities within a deliberately limited space: The goal is a refined, rich, and full expression of a strong idea (Goldschmidt 2005), which is also the coevolution of a problem and solution (Cross 1997; Studer et al. 2018). Maher et al. (1996) stated that coevolution in design exploration supports the change of design solutions and requirements over time. Designers are typically explorative, future-oriented, and open to the unknown (Beverland and Farrelly 2011). Achievements in design exploration largely depend on the determination of design hypotheses and designers’ beliefs, knowledge, experience, styles, and willingness to take risks (Goldschmidt 2005).

Goldschmidt and Tatsa (2005) indicated that design creativity is associated with the number of highly interconnected moves. Goldschmidt (2005) continued that each move, as the fundamental design step, repositions designers in their exploration. Each move can be divided into smaller units that are not necessarily semantically complete entities. Each move is coupled with its preceding move to establish whether a link exists between them. These moves are encoded for particular research goals. Goldschmidt (2005) indicated that data regarding these encoded moves often serve as a basis for diverse statistical tests. A design solution consists of numerous interconnected partial solutions. Frequently shifting between these design moves is the strategy designers employ to ensure that their problem-solving processes remain focused.

Designers engage in creative processes by drawing inspiration from environmental attributes that they associate with their experiences until a cohesive whole is formed (Nam and Kim 2011). The capacity for exploring and interacting with unknown concepts is vital for developing creative designs (Banerjee et al. 2008). Studies have revealed that design exploration may inspire innovative product development (Molina-Castillo et al. 2011; Tabeau et al. 2017) and differentiation (Kim and Atuahene-Gima 2010; O’Cass et al. 2014). Studies have also demonstrated that design exploration affects product performance, such as the degree of innovativeness (Daly et al. 2018; Tabeau et al. 2017), cost efficiency (Kim and Atuahene-Gima 2010; O’Cass et al. 2014), and quality (Molina-Castillo et al. 2011).

Based on the aforementioned studies, we hypothesised that design exploration predicts designer creativity accurately. However, until now, no sufficient academic evidence has proven this relationship. This may be partly because of the lack of advanced research tools to obtain trustworthy validation (Gero 2010; Liang et al. 2017; Liu et al. 2018), thereby impeding the development of a practical method to screen and identify highly creative designers. The emergence of deep learning promises excellent pattern identification of the brainwave signs of mental state (Goshvarpour and Goshvarpour 2019; Zeng et al. 2018) and optimal choice (Zhang et al. 2017; Zommara et al. 2018), which are critical for designers during the creative process.

Deep learning approach

Deep learning methods have been successfully used in recognition studies (Karpathy et al. 2014; Krizhevsky et al. 2017; Silver et al. 2016). A convolutional neural network (ConvNet) is the most common method of deep learning for processing unstructured data. Long short-term memory (LSTM) networks are a method specializing in processing time domain data. Researchers in the deep-learning field often use these two methods to deform and concatenate them into various network structures (LeCun et al. 2015). ConvNet is a deep learning algorithm that tightly couples local feature extraction and global model construction. ConvNets can extract useful information from input data, assign significance (e.g. learnable weights or features) to diverse aspects of inputs, differentiate between them, and amplify crucial aspects of the input in deeper layers while also strengthening the less significant aspects (Bashivan et al. 2016; Schirrmeister et al. 2017). In other words, with sufficient training, ConvNets can learn critical features by focusing on the edges of parts of the input to increase and finally cover global features in deeper layers in the hierarchy of the input information.

Deep learning methods have also been applied to electroencephalography (EEG) signal decoding problems (Dose et al. 2018). Schirrmeister et al. (2017) combined temporal and spatial filters and treated EEG data as a 2-dimensional (2D) array, using time points as the width and electrodes as the height of inputs, to examine configurations of ConvNets, including shallow, deep, and hybrid ConvNets and ResNets. Bashivan et al. (2016) transformed EEG data into 2D images with trial duration and time windows, and ConvNets, LSTM networks, and their mixtures were examined. The results of applying visualisation methods revealed variations in brain activity, indicating that the mixed network configuration had the lowest test error and best performance.

The present study specifically evaluated directed acyclic graph (DAG) networks (Szegedy et al. 2015) and series networks (Krizhevsky et al. 2017) (Figs. 1, 2). ConvNets were conducted in configurations of both DAG and series networks, and LSTM and mixed networks were applied in series network configurations. Shallow networks, used for tuning model parameters, contained input layers, LSTM or convolutional blocks, fully connected layers, softmax layers, and classification output layers. Deep networks additionally connected cascaded or parallel configurations in convolutional or LSTM blocks. Adam (Kingma and Ba 2015), a variant gradient descent method, was selected as the parameter search method for both LSTM networks and ConvNets.

Fig. 1.

Configuration of DAG networks

Fig. 2.

Configurations of series networks

Research purpose

Although design exploration has been observed to play a critical role in inspiring creative performances and outcomes, few studies have provided evidence of an association between design exploration and designer creativity or determined their cause–effect relationship. Therefore, the present study applied EEG to investigate the brain activations of graphic designers when they responded to visual stimuli while engaging in design exploration tasks. The aim of this study was to determine the predictive effects of design exploration on designer creativity by applying a deep learning approach. The three research questions addressed are as follows: (i) How can design exploration activities predict designer creativity? (ii) What type of deep learning network architecture has the best prediction accuracy? (iii) Which features can enhance prediction accuracy? To answer these questions, we designed a series of visually stimulating design exploration tasks and employed a mixed network configuration to identify the distinct creativity levels of designers from their recorded EEG data.

Network evaluation

Parameter setting and the meanings of all network configurations are elucidated in this section. Parameter settings were determined through grid search with leave-one-pair-out cross-validation in this study (Bergstra and Bengio 2012; Hsu et al. 2008). For LSTM networks, the number of hidden units that contained information of passed time steps through state activation function and gate activation function were tuned from 10 to 200 units, increasing by 20 units. Network training iterations, maximum epoch, sampling size, and minimum batch were set to 20 and 150, respectively. A learning stabilisation parameter (i.e. gradient threshold), was tuned from 1 to 0.1, decreasing by 0.2 increments. For shallow LSTM networks, 140 hidden units and a gradient threshold of 0.9 achieved the best prediction accuracy.

For ConvNets, maximum epoch and minimum batch were set to 6 and 80, respectively, which were constrained by GPU memory and system resources. The initial learning rate of ConvNets was set to 0.1. Learning rate drop factor controlled the period of learning rate change. The learning rate drop factor was set to 0.1 after two learning epochs and minimum learning rates were set to 0.001. Because each independent component (IC) of EEG data was one-dimensional (1-D), a 1-D filter was chosen with a size of 1 × x, in which x is 4 in the first and 2 in the other convolution layer when a network configuration involved more than one convolution layer.

The size of a filter was constrained by shrinking data sizes in deeper convolution layers. The number of filters was set to 100 and moving stride horizontal and vertical filters were set to 2 and 1, respectively. The batch normalisation layer was added to maintain output of the convolution layer with zero mean and a unit variance (i.e. variance equal to one) distribution for a batch of training iterations. Furthermore, a drop layer was adapted in both the LSTM network and ConvNets to mitigate model overfitting (Srivastava et al. 2014). The number of class output of a fully connected layer was set to two, which was the same as the number of creative levels in the data.

Support vector machines (SVMs) and tree-based ensemble methods are commonly applied in machine learning to classify groups and were the baseline methods in the present study. Parameters, cost, and gamma in specific energy ranges of the SVM were tuned by grid search through leave-one-pair-out cross-validation. The parameters, numbers of weak learners, and specific learning rates of tree-based ensemble methods were also selected through grid search.

This study conducted leave-one-pair-out cross-validation for parameter determination and prediction evaluation. One pair of samples comprised two samples undergoing the same task, one from the high-creativity (HC) group and one from the low-creativity (LC) group. In this cross-validation, the number of folds equals the number of pairs in the data set. Therefore, the learning algorithm was applied once for each pair, using the selected pair as a single-item test set and using all other pairs as a training set. To improve accuracy and generalise the prediction model, we established random combinations of samples in the HC and LC groups to form several data sets, using one for tuning the parameters of all network configurations and baseline methods.

Methods

Participants

Forty graphic designers were invited to participate in an EEG experiment. All participants had worked in the design service industry for more than 10 years and were recognised for their achievements with awards in international design competitions. They had no history of drug or alcohol abuse and had normal or corrected-to-normal vision. The participants were guided to undergo a self-assessment with a creative personality scale (CPS; Gough 1979) after completion of the EEG measurements. CPS questionnaire contains 18 positive and 12 negative items. One point was added for selecting a positive item, and 1 point was subtracted for selecting a negative item. Because higher CPS total scores indicate higher creativity, the top and bottom 25% of participants were then divided into HC (CPS cutoff point = 6.05) and LC (CPS cutoff point = 3.96) groups for brainwave comparison analyses. Of the 10 HC designers, four were women and six were men, their ages ranged from 31 to 40 years, and seven were right-handed and three were left-handed. Of the 10 LC designers, six were women and four were men, their ages ranged from 28 to 38 years, and nine were right-handed and one was left-handed.

Equipment and materials

The 32-channel wearable EEG system was used in this experiment. It consists of two foam-based sensors which are applied only to the Fp1 and Fp2 sites. The scalp markers were located based on the international 10–20 system according to the human cerebral structure. This wireless system features dry electrodes and a soft cap, consolidating its precision and convenience. The dry sensors could be used repeatedly on hairy sites. This headset has a sampling rate of 250 Hz and a 16-bit quantisation and the electrode impedance was maintained as low as possible (≤ 5 kΩ). The brainwave data were received through portable devices by the Bluetooth protocol. Data collected from the experiment were exported in the ASCII (.txt) format.

Robust research concerns the effects of abstract images on design exploration and designer imagination (Cardoso and Badke-Schaub 2011; Cila et al. 2014; Liang et al. 2019). Therefore, the visual stimuli used for design exploration in this study were presentative paintings by Joan Miró, one of the most famous surrealist artists in the world. His geometric shapes, biomorphic forms, abstract objects, and inventive style had gloried himself and inspired later generations. Many of Miró’s works have become teaching material for design education in Taiwan, and most were published prior to 1960 and have entered into the public domain in most countries, mitigating copyright concerns. The usage of these artworks was restricted for the purpose of academic research in this study. Ten pieces Miró works were selected and randomly presented during the experiment.

Experimental procedure

This project was approved by the Research Ethics Office of National Taiwan University (NTU-REC No: 201706HM070). The participants signed consent forms and were briefed regarding the experiment. We then set up the EEG system with participants and recorded their 60-s resting-state brainwaves as the baseline for potential correction. During the experiment, the participants watched a PowerPoint presentation that led them to verbalise the design problems, purposes, and imagined outcomes of a personal design project.

Each participant was then asked to look at two randomly displayed Miró paintings and think of answers to the following questions: ‘Which parts of this painting do you want to explore further (e.g. the originality of the idea and the techniques used)?’, ‘How would you like to change your original design?’, and ‘What do you think are alternative outcomes?’ During this process, we recorded EEG data for 1 min. The designers then verbalised their answers for 2 min. Each round for two particular artworks lasted 3 min, for a total of 15 min for the 10 Miró paintings. The 2-min narration subsession helped us to comprehend designers’ explorations while methodically comparing brainwave activation and narrative information. These subsessions also served as intervals to avoid recording overlapping brain responses. The experiment duration was approximately 30 min, including the experiment explanation, EEG system setup and testing, design project description, and baseline signal acquisition. A conceptual diagram of the research procedure is illustrated in Fig. 3.

Fig. 3.

Conceptual diagram of the research procedure

Data acquisition and pre-processes

Malfunctioning channels were first removed using five standard deviations from the mean as the threshold after measuring kurtosis in EEGLAB (Delorme and Makeig 2004). A low-pass filter with a cutoff frequency of 50 Hz and a high-pass filter with a cutoff frequency of 1 Hz were used to eliminate line noise, muscle movement, and oculomotor activity on the basis of finite impulse response filters. The 60-s EEG signals were split into 1.6-s epochs (Liu et al. 2018) that would not overlap and were then pared to avoid edge artefacts. The differences between the 60-s time slots and the baseline reached statistical significance.

Subsequently, we used EEGLAB to deconstruct the filtered EEG data through independent component (IC) analysis with the Infomax algorithm (Liu et al. 2018), which reverses the superposition and allows the EEG data to be separated into mutually independent scalp maps. The scalp topography of each IC could be further analysed using DIPFIT2, an EEGLAB plug-in, to detect the 3-dimensional location of equivalent dipoles on the basis of a boundary element model of the human head. The threshold of residual variance was set at 15%. This process was applied to the concatenated epochs from all conditions, and each IC was normalised to a standard normal distribution.

The power spectra of components were derived by the Fast Fourier Transform function, through which all components were filtered into delta, theta, alpha, beta, and low gamma bands. In addition, given the large number of model parameters, data were often cropped during data augmentation (Schirrmeister et al. 2017). All ICs were divided into overlapping 400-ms windows that were shifted by 40-ms increments. As a result, thirty 400-ms samples were separated from 1.6-s IC epochs. The time range of each IC epoch did not have specific physiological meaning, but could increase their size by a factor of 30. The features of the IC epochs included normalised ICs (one dimension), signal bands (five dimensions), the average power of signal bands (five dimensions), and the power spectra of ICs (fifty dimensions). There were a total of 61 dimensions of features in each 400-ms segment (one sample).

Results

The deep learning methods were performed with Matlab R2018a using a standard PC with 3.7 GHz Intel(R) i7-8700K CPU, 64 GB RAM and a NVIDIA GTX 1050Ti Pascal GPU with 4 GB RAM. EEG data of 20 participants were grouped into 10 pairs, each with one LC and one HC designer. Both HC and LC participants were randomly shuffled to generate 10 combinations, of which one was used for parameters tuning and the remaining nine were used for repeated prediction evaluations. Table 1 presents the data profile of the parameter tuning combination. We also tested the prediction accuracy of this combination with features of signal bands and power spectra of ICs and found that pairs 4, 7, and 8 had relatively lower prediction accuracies than the others, revealing a need for further testing of distinct combinations that were nontrivial (Table 1). The calculation of prediction accuracy was defined as follows

$$\text{Accuracy}=\frac{\text{True}\text{positive}+\text{True}\text{negative}}{\text{True}\text{positive}+\text{True}\text{negative}+\text{False}\text{positive}+\text{False}\text{negative}}$$

Table 2 presents the prediction accuracy of shallow LSTM networks with various features and baseline methods (i.e. SVM and tree-based ensemble methods). Using normalised ICs directly as features with LSTM networks resulted in low accuracy, but using power spectra of ICs resulted in significantly improved accuracy.

Table 1.

Profile and accuracy tests of combinations of tuned parameters

Paired number	1	2	3	4	5	6	7	8	9	10	Average
Number of samples
LC	1740	60	180	1860	1680	1920	1860	1620	1620	1620	–
HC	1740	1620	1800	1560	1800	1800	1740	1680	1380	180	–
Total	3480	1680	1980	3420	3480	3720	3600	3300	3000	1800	–
Accuracy	1	1	1	0.46	1	1	0.48	0.54	1	0.9	0.84

1. Profiles display the sample numbers of each pair

2. Prediction accuracy tests use features of signal bands and power spectra of ICs

Table 2.

Average prediction accuracies of LSTM networks with different features and baseline methods

Method (feature)	Averaged prediction accuracy
Shallow LSTM (normalised ICs)	0.39 (10 hidden neurons)
Shallow LSTM (five signal bands)	0.50 (60 hidden neurons)
Shallow LSTM (five signal bands + power spectral of ICs)	0.84 (140 hidden neurons)
SVM (five signal bands + power spectral of ICs)	0.54 (C = 1)
Tree-based ensemble method (five signal bands + power spectral of ICs)	0.65 (50 learners, LR 1)

Tuned parameters present best prediction accuracy

Table 3 lists the average prediction accuracy (0.81) of all combinations. Only the result using shallow LSTM networks with power spectra of ICs as features is reported because it outperformed other methods.

Table 3.

Average prediction accuracies of 10 combinations

Combination number	1	2	3	4	5	6	7	8	9	10	Averaged prediction accuracy/Std.
Shallow LSTM networks	0.84	0.78	0.84	0.75	0.77	0.84	0.8	0.83	0.85	0.84	0.81/0.04

For network configurations, we compared the prediction accuracies of various deep layers between LSTM networks and ConvNets, both with five signal bands and power spectra of ICs. Deeper network configurations did not demonstrate improved prediction accuracy (Table 4). This result may reflect information loss in deeper layers and model overfitting.

Table 4.

Prediction accuracies of LSTM networks and ConvNets with various layers

# of layers	Prediction accuracy
	LSTM	ConvNets
1 block	0.84	0.65
2 blocks	0.78	0.60
3 blocks	0.73	0.73
4 blocks	0.62	0.69
5 blocks	0.54	0.64

DAG networks were also examined in various pipelines of ConvNets; the results revealed lower prediction accuracy compared with shallow LSTM networks (Table 5). Furthermore, we examined mixed network configurations (i.e. concatenating ConvNets and LSTM networks), but the prediction accuracy (0.55) was still lower than that of shallow LSTM networks.

Table 5.

Prediction accuracies of ConvNets with a configuration of DAG networks

Numbers of DAG pipelines	2	3	4
ConvNets	0.74	0.64	0.72

Discussion

Using a deep learning approach, this study examined the predictive effect of design exploration on designer creativity and identified the specific network architecture and features used in deep learning that can augment prediction accuracy. Our results were as follows. (i) Design exploration effectively predicted designer creativity. (ii) Shallow architectures had higher prediction accuracy than deeper architectures. (iii) Shallow LSTM networks had higher prediction accuracy than convolution neural networks. (iv) Bandpower features enhanced prediction accuracy, and shallow LSTM networks with the power spectra of ICs outperformed other deep learning methods. (v) DAG networks did not increase prediction accuracy. These results may offer insights into the neural basis of design activities and creative cognitive ability and provide a practical method for screening and identifying highly creative designers.

Design exploration’s synthetic and proactive nature spurs designers to innovate products and services (Fallman 2008). Studies have suggested the vital role of design exploration for inspiring creative performances and outcomes (Banerjee et al. 2008; Daly et al. 2018; Gero 2004; Molina-Castillo et al. 2011; Tabeau et al. 2017). However, little evidence has elucidated the relationship between design exploration and designer creativity (Goldschmidt 2011; Studer et al. 2018). The current study adopted an innovative methodology of deep learning and filled this research gap by illuminating the cause–effect relationship between design exploration and designer creativity. This finding has crucial implications for design educators planning instructional strategies and design managers identifying and developing talent.

In this study, deep network configurations did not result in high prediction accuracy. This result may derive from information loss in deep layers and model overfitting, which is typically caused by the complex structures of deep neural networks and insufficient data (Kim et al. 2017; Srivastava et al. 2014). Although we have applied dropout and batch normalisation to avoid overfitting problems, refraining from using complex system structures in this study was difficult. In addition, parallel or redundant explorations are not typically performed across strategies in the real world because of limited project resources (Clevenger et al. 2010). Most experienced designers readily adhere to a limited number of possible solutions (Ericsson 2006) and tend to create a few core ideas within a refined goal using tacit and flexible rules and a well-developed design rationale (Goldschmidt 2005). The limited alternatives and subjective performance experiences designers possess (Ball et al. 2004) may explain why shallow architectures had higher prediction accuracies than deep architectures.

Each design move repositions designers in their exploration. Goldschmidt (2005) indicated that each move is coupled with preceding moves to establish links between them and gradually form optimised designs. Frequent modifications between design moves is the strategy designers use to ensure focused processes. The nature of design exploration may explain why shallow LSTM networks outperformed ConvNets in terms of prediction accuracy. LSTM networks are applied with series network configurations, whereas ConvNets are applied with configurations of both DAG and series networks. A series network has a single input layer and a single output layer. The layers in between are arranged sequentially, generally in a particular order. In comparison, a DAG network has a more complex architecture, with inputs from multiple layers and outputs to multiple layers. The layers in between are arranged as a DAG, connected in terms of the given connections. The characteristics of connections in series network configurations appear more adaptable to exploration activities. This also explains why DAG networks did not improve prediction accuracy in this study.

Lotte et al. (2018) indicated that bandpower features signify the power of EEG signals for given frequency bands in channels, averaged over time windows. Bandpower features can be computed in diverse ways and are widely used in oscillatory activity (i.e. changes in EEG rhythm amplitudes). As such, bandpower features are perceived as gold standard features for brain–computer interaction and mental imagery analyses (Lotte et al. 2018). This explains why shallow LSTM networks with the power spectra of ICs outperformed other deep learning methods. Studies have confirmed the diverse influences of 3-D objects, text, and musical stimuli on designer attention, association, and imagination (Liang et al. 2017; Liang and Liu 2018; Liu et al. 2018; Yao et al. 2017). This study provides new evidence to support the predictive power of design exploration on designer creativity through the application of a deep learning approach.

Research limitations and follow-up

This study contributes insights into literature on both design and neurocognition, but three limitations must be acknowledged to facilitate further investigation into this novel area of research. First, the relatively small number of participants (20 out of 40) may be insufficient for a design experiment, despite being representative. Future studies with larger samples and those comparing graphic designers with designers of other fields would be beneficial. Second, the stimuli used in this study were limited to Miró’s artwork. Investigating a variety of forms of stimulation (e.g. films, texts, music, sound effects, or even smells) is a practically relevant direction for future studies. Third, deep learning methods are currently only partially able to deal with hierarchical structures, but exploration is a part of the process of inquiry that designers undertake in which multilevel experiments with diverse hypotheses are simultaneously tested. Continually adopting emerging techniques to enrich relevant research is crucial.

Closing remarks

To summarise, our results extended previous research on design exploration by providing neuroscientific evidence of causality between design exploration and designer creativity. This contribution clarified the role of design exploration in creative cognition and provided critical insights into how and when particular activities interact with design goals to support the generation of novel and useful ideas. Additional studies that design other exploration activities and use advanced artificial intelligence techniques are necessary to test and confirm the relationship between design exploration and designer creativity and to classify discrete creativity levels.

Further research should have several iterative phases of goal-directed and spontaneous exploration. Creative designers embrace the creative potential of exploration and are capable of integrating it into their work. Future empirical research on the interplay among design exploration, designer creativity, and projected outcomes may allow us to fully appreciate the intricate cognitive mechanisms involved in creative cognition. Such research may also elucidate the cognitive capacity and training that enhances design performance. Together, these implications suggest potential for more development of the exploration activity domain in design theory by bringing together insight from literature of various fields.

Compliance with ethical standards

Conflict of interest

The authors declare no conflict of interest.