Sage Journals: Discover world-class research

Abstract

Designers in the built-environment disciplines increasingly solve problems using generative design methods, which promise novel and performant solutions to design problems but produce large design spaces that are challenging to explore. Design Space Exploration (DSE) interfaces have been used to understand, refine and narrow design spaces. Still, a critical analysis of current DSE interfaces shows a gap between their features and how designers explore and make decisions. We conducted a design study with domain experts to develop a DSE interface (DesignSense) that tightly integrates and adds to several features found separately in current DSE systems. We present a formative focus group evaluation, which suggested more areas for improvement and highlighted the need to distinguish designers from scientists as two user groups of DSE systems with varying needs, amongst other findings.

Keywords

Architectural design design analytics design space exploration generative design visual analytics user-centered design

Introduction

Designers in the built-environment disciplines increasingly solve problems using generative design methods. Generative design (GD) refers to computationally creating design forms (e.g., building envelopes or spatial compositions) informed by advanced performance analysis (e.g., daylight simulations, sustainability analysis, cost estimation) to explore alternative solutions to a design problem. GD can amplify designers’ exploration and help find designs that satisfice¹ design performance goals, preferences and constraints,² potentially in ways they did not anticipate.³ For example, GD allows the generation of spatial compositions for a building while providing performance data, such as view quality, energy load, solar gain, and daylight.⁴ A premise of GD is to allow designers to consider a large set of alternative solutions (hundreds, if not thousands) simultaneously for the same time needed to create solutions manually. However, this is challenged by the cognitive choice overload which designers experience when presented with an overwhelming amount of information and a large design space of alternatives.^5–7

Any complex design problem has multiple conflicting quantitative and qualitative objectives, which requires designers to evaluate their trade-offs. In a typical generative design workflow, the design alternatives are sifted through to find the ones that meet the quantitative and qualitative design criteria,^8,9 such as aesthetics, composition, usable area, cost, and energy load. The qualitative criteria are computationally hard to assess, and mostly relies on design expertise. To support designers in managing the design alternatives, several Design Space Exploration (DSE) tools have been proposed (e.g., see the Design Explorer¹⁰). A characteristic feature of these interfaces is the combination of abstract data visualizations (data views) and visuals of design geometries (form views such as 2D images and 3D models), which can be used to evaluate and select generated designs.

As an example of a generative design project, in a real-world case exploring massing design solutions for an urban rezoning project in Vancouver, Canada, our industry partner’s designers generated more than 100 alternatives (Figure 1 shows a subset of solutions). Following that, designers search for alternatives that balance design goals, which can be driven by clients’ preferences, city regulations, and aesthetic, economic, or communal considerations. Each alternative is evaluated by considering performance metrics, such as setback limits, floor-space ratio, building height, site coverage in residential zones and annual hours of sunshine. Another objective is maximizing views from floors to nearby parks. Furthermore, building masses and the space between them are subject to qualitative considerations such as ease of movement, space for greeneries, and ease of construction.

Figure 1.

A sample set of design alternatives and focusing on one in a close-up view. The exploration studies solar gain and shadow factors (red means more radiation). Gradient colours on the surfaces represent the solar radiation received.

When presented with a collection of alternatives, how can designers search for the designs that best meet their goals and balance their design criteria? How can DSE tools support this kind of search, and what are the characteristics of these tools? And how to create these tools in ways that do not overwhelm designers or interrupt their creative flows?

Our close study of Design Space Exploration (DSE) approaches and tools in the literature revealed an opportunity to focus on a systematic exploration of tasks and corresponding tool features for DSE systems. This study critically reviews current DSE tools followed by identifying a domain characterization of DSE, i.e. the abstract goals and tasks to which DSE can be broken down. We present DesignSense as a novel DSE tool that learns from and expands on several features found separately in current DSE systems. We also based the interface design on domain task characterization and lessons adopted from information visualization. Finally, we present a formative focus group evaluation that suggests several areas for improvement.

Methodology for DSE problem characterization

Design studies¹¹ are a commonly used methodology in the field of information visualization to develop bespoke data visualization solutions for targeted domain problems. Design studies are suitable for designing and developing design space exploration tools as they tackle a domain problem (exploring the outputs of generative design) supported by data visualizations. Through a design study, we aimed to characterize the domain problem of design exploration of generative design through multiple iterations of understanding and abstracting DSE as well as designing, prototyping and evaluating interactive visualizations for supporting DSE tasks.

We followed an iterative and agile user-centred design approach for system prototyping and solicited feedback on multiple system iterations from architects and computational designers (as domain experts) from two global design consulting firms. Besides continuous feedback from domain experts, we analyzed previous DSE interfaces through multiple critical lenses that were developed iteratively throughout our research, namely the degree of coupling between design forms and data on the interface, the utilization of simplification techniques such as data clustering, the richness of exploration-support features such as grouping, ranking, selecting and zooming on design alternatives. This critical analysis informed our approach to designing DSE interfaces. Finally, previous studies influenced how we frame the problem of DSE for generative design and how we approached designing DSE interfaces. Specifically, we built on a study by Shireen et al.,¹² which investigated design exploration under choice overload conditions, and relied on models of DSE from domain researchers.^8,13 Shreen et al.¹² identified that designers explore design spaces through criteria identification, construction, evaluation, and application when searching for solutions in a generative design task environment. Finally, we further evaluate the developed system through a formative focus group with domain experts to receive initial feedback on how well it supports DSE of generative design, with a task-based evaluation to follow in the future, featuring think-aloud protocols during design exploration and open-ended interviews.

Background and related work

Before critically reviewing the literature, we introduce the key concepts we borrowed from the information visualization field (c.f. Munzner²¹ for details) to help the readers contextualize the arguments presented.

• A view is a visual representation of some data (e.g. a scatterplot is a view representing two quantitative variables).

• Parallel Coordinates Plot (PCPs) is a common visualization choice with datasets that include many variables (>3). Each data variable is represented by a set of parallel vertical axes, one for each variable, and each data item is represented as a series of line segments crossing the vertical axes at points that correspond to their values.

• A visual encoding denotes how data variables are mapped into visual attributes. For example, colour saturation can encode a country’s population on a world map.

• Filtering is showing only the relevant aspects of the data on a view.

• Brushing is the act of highlighting elements on an interactive view & Linking or Coordination is when the items highlighted on one view are also brushed on other views.

• Multiple Coordinated Views are interactive views whose visual representations change in response to interactions (e.g. brushing, filtering) on any view.

• Information Seeking is the process of finding relevant information in a dataset. A commonly accepted guideline for designing visual information-seeking tools is to: “Overview first, zoom and filter, then details-on-demand”.⁶³

Data generation process and data abstraction

Generative design systems in architectural design increasingly use parametric data-flow models.¹⁴ A typical combination uses a parametric model to generate 3D forms, such as Grasshopper and Rhino3D,¹⁵ where each combination of values for the model’s parameters is a design alternative, and all possible combinations form a design space.¹⁴ Designers generate designs satisficing the project’s goals—e.g., low energy usage and high utilization of daylight—by following a performance-informed process¹⁶ where parametric models combined with performance simulation tools are used, such as Ladybug and Honeybee.¹⁷ Alternatively, designers can generate design geometry by sampling from the design space through automated methods (e.g., genetic algorithms).¹⁸

Each design alternative in DSE includes the input parameters used to generate it and the output performance metrics resulting from evaluating it. Each design alternative also has an associated design form represented with 2D images, 3D geometry or both. Images can be used to visualize the results of performance simulations, such as colouring based on how much annual sunlight they receive (e.g., see Figure 1).

Design space visualization

Using information visualizations for DSE has many precedents in different fields. For example, the ARL Trade Space Visualizer,²² modeFrontier,²³ RAVE,²⁴ and LIVE²⁵ are widely used systems in engineering to integrate design visualization and optimization. They use feature-rich support for various abstract visualizations such as parallel coordinate plots, scatterplot matrices, and histograms. Users of these systems can shop for alternatives^26,27 considering desired performance criteria.

In architectural design, selecting designs requires judging performance and form criteria together, which is necessary for a holistic evaluation. Designers seek to understand their design problem and the performance-to-form relationships in their generated design space (i.e., make sense of it).^12,13,41 Therefore, data visualizations should support identifying patterns in data and revealing relations between data attributes. Users facing shopping interfaces place visuals of ‘merchandise’ at the forefront but only support simple widgets and controls that do not sufficiently contribute to understanding the problem. Finally, unlike passive shoppers, designers use any insight gained towards generating new solutions (refine). In the middle between data-focused and visually-oriented shopping interfaces, we find DSE interfaces that couple abstract data visualizations with form visualization in 2D or 3D views. The early examples of such DSE interfaces from the visualization community are FilmFinder,²⁸ Design Galleries,²⁹ and EZChooser,³⁰ among others from different design domains.^10,31–47

Analysis of DSE interfaces

Our survey covered about 30 interfaces and focused on issues such as the relation between data and forms during DSE and the management of alternatives. The DSE interfaces we identified are complex; we analyze them through multiple lenses developed iteratively. The recent work by Garcia and Leitao⁴⁸ presents a similar analysis of DSE interfaces. Their study included nine interfaces and covered how design parameters and forms are visualized and the available means for managing alternatives through searching, filtering and ranking. Our analysis makes similar (and different) observations and expands them by employing concepts from information visualization and human-computer interaction. The study by Garcia and Leitao covers other areas not discussed here, such as the use of multidimensional sliders and the relation between Algorithmic Design and DSE (c.f. Garcia and Leitao⁴⁸ for details).

Data and form coupling

The literature on DSE argues for coupling qualitative and quantitative characteristics for evaluating design alternatives.^{35,38,40,42,43,49,50} Our analysis distinguishes the interfaces representing design forms (form views) and data like performance metrics and input parameters (data views).

Data views and input manipulation

DSE interfaces use abstract data visualizations for selecting alternatives based on their performance or input values. For example, parallel coordinates plots (PCPs)^{10,32,34,38,41,46} or scatterplots^40,44,51 are used separately or together for supporting multidimensional data visualization.^35,37,45,47 When considering the manipulation of input parameters, as opposed to visualizing their values, Garcia and Leitao⁴⁸ observed that most DSE interfaces included 1-dimensional sliders only to manipulate inputs. The alternatives to 1-dimensional sliders, namely 2D/3D sliders and 2D/3D parallel sliders, are rare, except for the D.Star,⁴¹ which supports manipulating plotlines on a PCP.⁵²

Form views, overview and details

Overviews of design forms can be seen in multiple DSE interfaces, including interactive (e.g., scrollable, sortable) grids of thumbnail images^10,35 or 3D geometries.⁴⁰ Other visual presentation approaches rely on preprocessing with clustering or dimensionality reduction. Aside from overviews, detailed form views for an alternative can be found as a zoomed 2D image and an interactive 3D view.^{10,31,38,40,47} The detailed views can have multiple images^10,38 to provide different perspectives on forms (e.g. side, top or perspective views).

Compositing form and data views

Multiple patterns exist for compositing⁵³ form and performance data,⁵³ for example, juxtaposition (e.g., Dream Lens⁴⁰), superimposition (e.g., colouring surfaces by the amount of sunlight they receive in Figure 1), nesting (e.g., visual thumbnails inside Self-Organized Maps³⁸), overloading (e.g., forms’ details window inside Parallel Coordinates Plots in Cupid³²), or integration (e.g., Design Galleries²⁹).

Interactions between data and form views

Most DSE interfaces support selecting alternatives based on their attributes only. Form views often only react to data-driven queries, and users relate forms with data attributes through brushing and linking combined with a distinctive and common visual encoding^34,40,41,45 or through explicit visual links in others.²⁹ A typical example of data-driven queries is to filter design forms based on selections on a scatterplot ^40,44,51 or a parallel coordinates plot.^10,34,35 On the other hand, form-driven queries can be seen in only a few systems, including the chisel tool in DreamLens⁴⁰ and the interactive cluster displays in Cupid.³² A similar observation by Garcia and Leitao⁴⁸ showed that DSE interfaces supported visualizing one or more designs. However, they did not support superimposing or editing designs through direct manipulation, as opposed to tweaking sliders. We haven’t seen any DSE interface with a tight coupling of form and data views (through form-driven exploration), albeit arguing its importance in the literature.

Design space simplification

Multiple researchers^5,54 argued that simplifying design spaces can alleviate the choice overload created by working with a large set of alternatives. Abuzuraiq and Erhan⁵⁵ showed that simplification, whether through clustering or dimensionality reduction, has been used in DSE systems to create overviews of the design space and to set the stage for data-form coupling (e.g., embedding displays of forms on a self-organizing map^38,42 or in a hierarchical clusters tree embedded in a parallel coordinates plot³²).

Knowledge externalization

Designers make sense of a large set of alternatives by organizing them into collections that embody and dynamically change with their design criteria.^12,41 The use of user-created collection(s) is rare in DSE systems. We find examples of supporting single collections under the name of a “cart”,³⁴ or “bookmarks”⁴² or a “privileged collection”.³¹ A single collection helps accumulate exploration progress but does not enable multiple criteria reviews, such as needed by intelligence analysts²⁰ or designers.¹² A few DSE systems allow working with multiple collections, whether by creating multiple data-flow streams³⁸ or first-class spatial containers.⁵⁶ On the other hand, general-purpose visual analytics systems support multiple collections in abstract representations such as markings (e.g., Spotfire⁵⁷) and sets (e.g., Tableau⁵⁸).

DSE domain goals and tasks

Based on the critical literature review, the tools used for DSE, and our interaction with design practitioners, we proposed the following DSE goals and tasks.

DSE goals

In generative design, interactive data visualizations can serve several overarching goals: making sense of (understanding), refining, and narrowing the design space.

Understand

Understanding the design space involves uncovering patterns, relationships, and performance trade-offs. Furthermore, designers can apply or identify new criteria by comparing multiple alternatives and assessing their strengths and weaknesses.

Refine

The generative design process is iterative. Examining alternatives can lead to changing the generative model (e.g., parametric model) to produce different forms or updating the performance metrics. Designers continually improve their understanding of the problem.

Narrow

Designers seek alternatives that satisfice¹ the existing or emerging criteria. An alternative is a proposal, not merely a data point; even poorly performing options can lead to better ones.¹³ Finally, the selected designs are not treated as final solutions. Instead, they are refined or used to guide the subsequent exploration.⁸

DSE tasks

We propose the following five high-level DSE tasks inspired by the visualization tasks taxonomy by Brehmer and Munzner.¹⁹

Find potential designs (T1)

Designers explore to identify satisficing designs that meet their evolving criteria. This exploration involves browsing and exploring when the problem is ill-defined and turning into looking up and locating designs when criteria are stable and understood. Designers may shift their focus on different parts of the design space through filtering, sorting or zooming.

Relate inputs to performance and form data (T2)

Designers analyze how input parameters influence performance and form to identify dependencies and correlations by sensitivity analysis.

Understand performance trade-offs (T3)

Designers consider output-to-output relationships to find effective designs or trade-offs. When trade-offs are present, designers may find, e.g., the Pareto-frontier among conflicting performance objectives. To arbitrate between designs with performance trade-offs, designers compare design forms or weigh performance metrics based on various factors. Designers also connect outputs to design forms to identify the features contributing to a performance portfolio.

Couple form and data (T4)

Designers may compare design forms to arbitrate between designs with similar or trading-off performances. Overall, the analysis of designs’ performance and form is interwoven and has no specific order. Designers can inspect an individual alternative (performance or form) to learn about it in detail, compare multiple alternatives, or overview many alternatives as a step towards understanding and narrowing them.¹²

Externalize insight (T5)

In a study by Shireen et al.,¹² designers were asked to explore a large set of generated alternatives printed on cards. They observed several search patterns, including putting designs into groups according to different criteria or hypotheses (e.g., the group of buildings with circular shapes) and marking, tagging, ranking, and discarding alternatives. The above patterns suggest that designers externalize their knowledge during DSE akin to the sensemaking cycle reported by Pirolli and Card.²⁰

DesignSense’s system design for DSE

We developed DesignSense as a fully functional DSE prototype featuring coordinated and linked views,⁵⁹ representing and tightly coupling views of the design forms and performance data (https://www.designsense.app/). We proposed mechanisms for expressively selecting from and navigating through design alternatives (Figure 2). DesignSense is implemented as a front-end web interface using Vue.js and D3.js. It enables the storage and retrieval of design data, similar to Design Explorer.¹⁰

Figure 2.

DesignSense includes the following coordinated views: (1) A parallel coordinate plot. (2) a scatterplot with lasso and rectangular brushes. (3) a gallery of sorted and selectable image thumbnails. In addition to (4) the top bar where system-wide operations are accessed. (5) a panel where alternatives saved as a set are shown. (6) a panel where results from a background clustering algorithm are presented.

DesignSense facilitates the exploration and selection of numerous alternatives (T1) while enabling both qualitative and quantitative assessment (T4) to support sensemaking and problem understanding (T2, T3). Interactive form and data views facilitate combined evaluation, fostering bidirectional selection and design inspection, facilitating a continuous review of qualitative and quantitative criteria. Operations such as grouping, information-seeking, and annotating aid in organizing the design space (T5) and adding subjective semantics to groups, enabling users to build on previous exploration progress. Complex selection queries are visually and incrementally made using various logical operations, with support for clustering and Pareto analysis throughout the exploration process. Users can seamlessly switch between overview, comparison, and inspection levels of design analysis without mode switching. Interface components are customizable to accommodate changing project needs, leveraging large displays effectively. Below, we describe the system details, justifying our design decisions where needed.

System features and composition

DesignSense proposes novel features while adapting others that appeared separately in the literature but not in the same combination and degree of integration in DesignSense. Its first contribution addresses the need for coupling form and data through form-driven selections and linking form and data views. Second, it tackles the choice overload problem⁵⁴ by introducing integrated clustering features. Third, DesignSense aims to support the sensemaking cycle in DSE through knowledge externalization, allowing users to create, manipulate, and annotate sets of designs. Additionally, it offers minor yet important contributions, including incremental expressive selection (disjunctive queries), workspace customization, and seamless switching between overview, comparison, and inspection modes.

Incremental and Expressive Selection

Most DSE systems we analyzed (except^38,61) supported only conjunctive queries that are made by finding the intersection (logical AND operations) between sets, e.g., users can narrow the design space by repeatedly brushing on scatterplots or by picking ranges on widgets’ sliders. However, we argue that disjunctive queries should also be supported since they afford a higher degree of expressiveness and flexibility and allow users to select non-contiguous regions in the design space or select values from dimensions that are not ordered. Kazemi⁶² argues for disjunctive queries in design exploration on interactive and visually guided queries. In DesignSense, the users can choose alternatives from any view, and the selection is coordinated with all views through brushing and linking, using a consistent visual encoding scheme. Selections occur via various brush types tailored to each view, and they are shared. They can be refined incrementally (similar to Wills⁶⁰) through actions like replacement, addition, intersection, subtraction, and toggling or negating. The users can choose an operation, invert and clear selection by key modifiers or on fixed or floating menus (Figure 3).

Figure 3.

(1) set-based operations applied on selections. (2) toggling inspection (left), selection modes, creating sets, clearing and inverting selection (middle), information-seeking functions, e.g. zoom, filter, and resetting dataset (right). (3) On-demand context menu with frequently needed functions, where colour-coded icons distinguish their category (inspection, selection or information-seeking).

Information Seeking

DesignSense enables exploring a dataset shared between all or on one view. At any time, the shared dataset can be restricted to the selected design alternatives (zoom), or users may exclude the selection from the dataset (filter). We implemented undo to return to a previous state(s) and reset operations to reload the initial dataset to provide control and avoid premature commitment.

Inspection

While reviewing alternatives, users may inspect an alternative or visually link its representation to different views. When inspecting, the encoding on all the views changes to emphasize the data in focus. Users can initiate inspection by hovering (along with a modifier key) over an alternative’s representation in any view, such as its thumbnails in a gallery, lines in the parallel coordinates plot, or dots in the scatterplot (Figure 4).

Figure 4.

(1) Rectangular brushing on the PCP view. Tooltips show the precise brush range. (2) The lines of an inspected alternative are thicker, while the lines of currently selected and unselected alternatives are given a lower visual priority. Finally, inspection is enhanced with tooltips that show the exact values for all dimensions. (3) and (4) inspection on the scatterplot with a circle outline and on the AG with thick dashed outlines around the thumbnail.

Data views

To support identifying the relationships (T2, T3) and narrowing the design space with data-driven queries (T1), we provide two data views, a PCP view⁶⁴ and a scatterplot view. Both visualizations are common in the DSE interfaces. In practice, PCPs are used to provide an overview of multidimensional data. They are also commonly used for filtering and identifying outliers. Since the order of axes in a PCP impacts whether a correlation is discernible, we support reordering and inverting axes. A rectangular brush along any axes can select an interval of values; once dragging is over, the alternatives brushed are integrated into the current selection depending on the current selection operation. To enhance the precision of selection on the PCP, we show a tooltip with the exact range of the currently drawn brush (Figure 4).

In DesignSense, users can select alternatives by rectangular and lasso brushes on the scatter plot. They can use the Pareto Frontier function to decide on the trade-offs when performance objectives conflict. The Pareto Frontier consists of designs that cannot be modified for one objective without compromising the others. Users can visualize and select frontiers for the two dimensions on the scatterplot (Figure 4 - (3)).

Form-driven selection in Augmented Gallery (AG)

DesignSense offers filtering, comparing, and inspecting alternatives based on qualitative criteria on the forms visualized in an interactive gallery of thumbnails, referred to as the Augmented Gallery (AG) (Figure 5). The bottom panel shows a grid of image thumbnails, each representing a different alternative. The thumbnails can be resized to increase the details and visibility, scrolled through, or sorted based on a data dimension. Also, the AG enables form-driven selection by dragging rectangular brushes on the thumbnails, which is most effective when combined with a sorting order and an image size large enough to discern the qualities of the design alternatives. It is also possible to toggle selections on their thumbnails.

Figure 5.

The Augmented Gallery (AG) shows sorted thumbnail images for all alternatives. Users can select alternatives by brushing on thumbnails or toggling, choosing to offer only the alternatives currently selected (eye icon) and seeing an enlarged image or 3D view of the alternative under inspection.

Users can hide unselected alternatives from view (through the eye icon) on the AG and continue narrowing the selection. Further selections are shown in this mode with dashed lines (Figure 5). The AG facilitates an in-depth comparison of a few alternatives through on-demand expandable panels containing a vertical list of data key-value pairs (Figure 6-Bottom). For closer examination, the AG offers a detailed inspection panel showing an enlarged image of the selected alternative, with the option to cycle through multiple images if available. The AG also has a 3D form view that allows panning, rotating, and zooming to enable focusing on desired design aspects.

Figure 6.

Top: Brushing is denoted with dashed lines when the AG only shows the current selection. Bottom: Direct comparison is supported via an expandable panel below each thumbnail listing design data as key-value pairs. This feature becomes available when the size of the icons is set to the largest.

Coupling Form and data

Coupling the qualitative and quantitative aspects of alternatives can be achieved by coordinating and moving back and forth between form and data views in DesignSense. For example, selecting alternatives in the gallery also visualizes the selection on the scatterplot and the parallel coordinates plot (or vice versa). This allows for the visual and numerical filtering of alternatives (e.g., a designer can first select a set of alternatives that fall within a certain threshold of sunshine utilization on the PCP by followed or preceded by selecting the alternatives with acceptable open-space utilization on the gallery). Coordination can also happen on one alternative through inspection (e.g., a designer reviews the form behind a specific alternative with a high area-to-volume ratio, as in Figure 4).

User-created sets

During exploration may want to save their progress to revisit it later or combine the results of this inquiry with earlier results. DesignSense enables that by supporting saving and retrieving selections into sets (Figures 7-(1)). Replace, Add, Intersect, Subtract, and Toggle operations apply when retrieving a set, which is treated like making a new selection. Users indirectly operate on and compose new sets from existing ones by systematically retrieving sets. This process is visually informed⁶⁵ and distinguishable in all views since recovering a set updates the current selection. Users can delete sets or add comments to them.

Figure 7.

(1) The Sets view contains a list of cards, each representing a set. The number of alternatives in each is shown. Sets can be deleted, retrieved and updated. (2) In the Clusters view the thumbnails show the clusters’ representatives. (3) the trophy symbol on the scatterplot view computes and highlights the Pareto-frontier for the chosen dimensions.

Clustering

Clustering is supported through the Clusters view to integrate simplification in DSE. Using the K-means algorithm, alternatives are clustered based on user-chosen quantitative dimensions after normalization. The clustering algorithm is run every time the visualized data is transformed, e.g., through filtering or zooming operations. We use the alternatives closest to the clusters’ center as representatives to visualize each cluster. Representatives in Clusters view (Figures 7-(2)) provide a visual overview of the design space and can be a starting point for further exploration. Selecting any cluster can be followed by any set operation of its members against the selection as the other cluster.

Flexible Layout and supporting multiple displays

All views in DesignSense can be resized and moved to support working on multiple display setups. Furthermore, an on-demand context menu (Figure 3-(3)) is provided to reduce travelling to the top menu bar for most system operations.

How DesignSense addresses gaps in the DSE literature

DesignSense addresses the multiple gaps found by surveying around 20 DSE systems:

1. Form-driven Selection. Although coordinating form and data views can be a common DSE system feature, only two systems offer form-first interaction, i.e. filtering designs based on forms alone.^32,40 The specific form-driven selection implemented in DesignSense, i.e. brushing/clicking on the images gallery, is not seen in other DSE systems.

2. Support for Multiple Collections. Only five DSE systems out of the systems surveyed allowed designers to create collections or sets of designs. Three out of five only enabled putting designs into a single collection^31,34,42 (i.e. as a shopping cart³⁴), while two others enabled creating multiple collections^38,56 at the same time. With a support for multiple groups, each group can represent a different search criteria or the result of a separate line on inquiry.

3. Support for Expressive Selection. Except for three systems,^38,61,62 most DSE systems only supported Intersection (AND queries) between subsequent selections on the interface. In contrast, DesignSense and a few others,^38,61,63 support additional selection mechanisms such as forming Unions and Subtraction boolean operations between groups of designs, which results in a more expressive query language for selecting designs.

4. Simple Integration of Simplification into DSE. Out of the surveyed systems, some used simplification (i.e., clustering or dimensionality reduction), but only two systems^32,42 (other than DesignSense) integrated simplification into design exploration such that it is coordinated with form or data views. However, we argue that DesignSense presents the results of clustering/simplification in a simpler manner, i.e. thumbnails of clusters’ representatives.

In summary, DesignSense tightly links and adds to several features found separately in current DSE systems.

Evaluation of DesignSense for DSE of generative design outputs

We evaluated DesignSense in a formative focus group study to receive initial feedback on how it can support DSE goals and tasks. Figure 8 shows DesignSense at the time of the evaluation featuring the PCP, the scatterplot, and the Augmented Gallery (without the inspection panel), coordinated brushing and linking, information-seeking, user-created sets, clustering, and moving and resizing views. After the evaluation, we added the inspection, side-by-side comparison, lasso selection, Pareto-frontiers computation, or context menus features.

Figure 8.

The interface of DesignSense at the time of evaluation.

Focus group details

The focus group consisted of eight architectural and computational designers from the same firm, all experienced in generative design and DSE using Design Explorer.¹⁰ Four were expert computational designers, and two specialized in building performance. Participants met online. Two researchers led the session: one presented DesignSense by screen sharing, and the other moderated the discussion on the DSE tasks and tool features. The aim was to gather feedback on if and how DesignSense could be used in practice and to receive suggestions for improvement. The participants were asked to focus on the tasks and features rather than the project.

The two-hour session was recorded, and meeting notes were taken for data analysis. In the first 30 min, we demonstrated the DesignSense features using a project from the participants’ firm. An open-ended discussion with frequent references to the interface followed the demonstration in the rest of the session. The focus group resulted in an engaging discussion about the DSE and DesignSense centring on idea generation rather than confirming features.

Evaluation results

One researcher reviewed the video recording and the meeting notes to encode data in a thematic analysis to group the topics that emerged. Generally, the participants agreed that the features for supporting DSE tasks are comparable to or enhanced from other DSE interfaces, such as Design Explorer.¹⁰ Below is a summary of the themes that arose during the discussion:

Visualization and shifts in design concerns

A noticeable theme emerged around distinguishing and prioritizing the data visualized based on the two concerns: design form and building performance. Both involve using different suites of tools while working closely with each other on a design project. For example, performance analysis results are reported to designers to inform their design decision-making; changes in the forms imply changes in the performance measures. The participants interested in building performance expressed a potential integration of their usual analysis process, such as creating a value function akin to the Design Space Construction framework.⁹ This method involves weighting and summing normalized performance values based on project-specific preferences, to help in multi-criteria decision-making. Additionally, they showed interest in sensitivity analysis, descriptive statistics, trend lines, data normalization, prediction via surrogate modelling of building performance, and ANOVA tests.

The participants who work closely with form asked for the ability to inspect a design’s geometry in more detail, e.g., for “...bigger images or 3D views,” and found that “image thumbnails alone to be too small.” One participant proposed a balanced emphasis on form and performance data by noting “…links back to form and performance and how you balance between them and how you get a form to be part of your decision making.” Regarding screen space, the participants preferred more form views rather than data views.

Scatterplot

The participants found brushing and highlighting data between scatterplot and other visualizations useful. They asked for additional demonstrations of brushing on various views multiple times with emerging design questions.

Clustering

The participants found that clustering, combining clusters, creating sets and brushing contributed to improved filtering of design alternatives. They noted that these features complemented each other and had the potential for “..making the decision-making [process] easier.” One participant stated, “I see… from the designer perspective, I truly believe that the clustering, grouping, and filtering...basically most of them belonged in my interpretation to the filtering family of operations.” The participant focused on the details of post-processing in performance analysis.

Set-based exploration

The focus group appreciated using sets to store and retrieve alternatives based on preferences or criteria and user notes. And they recommended a persistent repository for such sets. One participant noted that “…design does not happen in a day...how do you save the clusters [sets] if you are done?” and added,”...this would enable designers to continue where they left off”. One participant further stated,”…we would like to [retrieve and] evaluate a specific alternative only when needed.” There was an agreement on the importance of both automated and manual composition of clusters, which does not exist in the current tools they used. Clustering techniques, such as K-means analysis, were recognized as beneficial for grouping design families based on selected criteria.

Derived data

One participant suggested new features that can enable designers to derive additional data from available data, such as through pre-determined or custom formulas. At least descriptive statistics functions (e.g., mean, min, max, std) can be useful for comparing alternatives. They noted that “…organizing big data sets is one thing and [creating] results out of it another…for example how to normalize the results…”, which followed by “…we try to avoid picking one winner [and instead] we try to find the characteristics of winning designs [by looking at data in general]”. While PCPs visualize the dimensions with the same length as a “normalized” form, this is insufficient to make inferences on overall performance. Such functions can be of great utility when integrated into decision workflows.

One participant highlighted the difficulty in “subjective intervention” when working with derived data, such as weight preferences assigned to each objective in a value function. They mentioned that assigning “…weights to different concerns, such as energy [use], daylighting and view” can allow balancing the effect of some objectives—e.g., by assigning higher weights—albeit at the cost of transparency and implicit bias.

Diagonal sorting

The sorting alternatives in the AG view provoked ideas for additional sorting strategies. The participants shared a printed form of “diagonally sorted” 100 design alternatives to “scan the thumbnails from one corner of the grid to the other,” showing transitioning continuously. This was different from row-by-row, left-to-right, and top-to-bottom sorting, which requires going back and forth between the beginning and end of the rows.

Comparison

The participants showed an interest in comparing a small number of design alternatives in detail. They pointed at a different interface for comparing designs side-by-side. This feature is now implemented in the prototype but was not part of the evaluated version.

Multiple ways of visualizing form

Participants requested supporting multiple images for each design alternative to be able to view different performance metrics. For example, they mentioned using data mapped on the views to show “energy, daylight, solar radiation.”

Flexible layout

The participants appreciated the ability to resize and move the different panels freely in the DesignSense interface. One participant pointed out that “...the way people do design reviews until today is they print things up so that they can get a broad view and they can make comparisons,” but that with current tools “...you are stuck in a screen”.

Discussion

Matches vs. Mismatches

Three categories of features resonated the most with participants, judging from the positive feedback they received or the discussion they provoked. The first category relates to the coupling of form and data, a major driver for this work. The second category is set-based exploration, which includes all the features for creating and manipulating sets like coordinated selection, disjunctive queries, information seeking, user-created sets, and clustering. The interest in these features can be seen as a response to the large design spaces of generative design, which motivate systematic set-based exploration. The third category contains more specific features that were received positively, like the scatterplot view and the flexible interface layout change.

The participants also proposed multiple features highlighting some system mismatches during evaluation. These include assigning more screen space to form views, integrating more data analysis into design exploration, and linking the representations of individual alternatives on data and form views. The participants in the evaluation were silent about disjunctive queries (c.f. section on Incremental and Expressive Selection), which extend the conjunctive (narrowing) queries prevalent in DSE tools.

Designers vs. Scientists

The evaluation highlighted the different concerns designers and building scientists had when using a DSE interface like DesignSense. Matejka et al.⁴⁰ made a similar distinction between the ‘engineer’ type of users and the ‘artists’. They also argue that accommodating both groups (and those in between) can encourage them to appreciate each other’s concerns. Each group contributed different features to our discussions, e.g., asking for a larger form view versus more data analysis. In general, the evaluated version of DesignSense appealed more to the scientists among the participants. It highlighted the need to tailor the interface to each group through different features.

Exploration vs. Analytics

The evaluation revealed the need to support the design decision-making with design data visualizations and the demand for integrating exploration with data analysis (i.e., design analytics). This can be accomplished through visually-assisted support for techniques such as sensitivity analysis and value function elicitation, among others.

Future work

Our future research will continue, based on the initial formative evaluation, to focus on the following areas: (1) attuning to different user groups, (2) improving the system’s scalability and the exploration session’s longevity, and (3) conducting a task-based evaluation.

As evident from the focus group, the features of DesignSense currently appeal more to building scientists and designers with building performance expertise. An important area of future research is exploring how to make DesignSense easier to learn and use by designers without a background in data analysis. This can be achieved by introducing interface presets. For example, a designer’s preset prioritizes form views and simple visualizations, which will differ from an engineer’s preset, focusing on advanced data analysis. Furthermore, a combination of tooltips, tutorials, and walkthroughs could help make DSE interfaces more approachable to designers.

DesignSense is currently implemented as a front-end web page. We plan to implement a back-end system that allows users to maintain a persistent record of their exploration and design sessions. Furthermore, loading images and 3D models is currently a performance bottleneck that prevents DesignSense from scaling to large datasets. A back-end server can optimize form data processing by pre-compressing before serving to reduce loading time. Finally, a back-end server makes it possible to connect DesignSense to parametric modellers so alternatives can be generated on demand without leaving the interface, as well as offloading heavy building performance analysis tasks to the server.

Finally, the two-hour focus group session was not enough to review the system in detail and leaned towards suggestions rather than testing the choices made on DesignSense. Next, a follow-up task-based evaluation should directly assess DesignSense concerning the revisited domain tasks and the flow between them within the context of a realistic design project and the diverse needs of the users (i.e., designers or scientists).

Conclusion

Unlocking the full potential of generative design hinges on many factors that span the entire generative design process and extend beyond it since the generative design is often one part of a larger design process.⁸ These factors include the variation⁶⁶ and validity⁶⁷ of the underlying generative model and the ability to sample the design space that the model endows for diverse⁶⁸ and optimal alternatives. Generative design also relies on the ability to explore this generated set in a way that is unhindered by the size of this set^5,54 and in a manner that reflects design judgment,^{35,38,40,42,43,49,50} with its quantitative and qualitative dimensions. We argue that designers may avoid generative design systems if they lack proper support in exploring and utilizing the generated outputs. Therefore, we view this work as a step towards reducing the barriers to using generative design in practice by designing better DSE interfaces than are currently available. In particular, DesignSense does so by enabling designers to explore the outputs of generative design without being overwhelmed by the overload of design choices and to be data-informed in their exploration without forgoing their design sense. To that end, the positive feedback from the focus group on DesignSense shows that we are heading in that direction, and the suggestions made hint at some concrete next steps to better achieve our goal.

Although we focused our design study on the architectural design domain, the same need for coupling form and data and the same challenge of choice overload exists in other domains where forms are culturally important such as product design, mechanical design, urban planning and game design⁶⁹ to name a few. The demand for coupling ’form’ and data is not exclusive to design domains either, the design study by McCurdy and Meyer⁷⁰ with astronomers and astrophysicists showed a lack of support for the coordinated statistical and visual analysis of galaxy observation. Finally, design space exploration and shopping share many features, but whereas designers can create new alternatives on demand, consumers cannot. Nevertheless, future work can explore the data visualization and analysis features that could enhance consumer-facing shopping interfaces. Figure 9 shows a preliminary exploration on that front where we load and explore a dataset of used cars into DesignSense.

Figure 9.

Using DesignSense to shop for used cars using the posts retrieved from a popular community website.

We presented a domain characterization of DSE for generative design and a critical analysis of available DSE systems. We introduced DesignSense as a novel DSE system that integrates features found separately in existing systems and adds new ones. We based our design reasoning on domain task analysis and lessons from the information visualization field. We share our initial findings from a formative focus group evaluation of DesignSense that suggested several areas for improvement and highlighted the need to design DSE interfaces that accommodate the differing needs of both designers and scientists.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research,authorship,and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research,authorship,and/or publication of this article: This research was funded by NSERC Discovery (RGPIN-2022-03108) with SFU ethics approval #20180335.

ORCID iDs

Ahmed M. Abuzuraiq

Halil Erhan

References

Simon

. The sciences of the artificial. Cambridge: MIT Press, 1969.

Caetano

Santos

Leitao

. ˜Computational design in architecture: defining parametric, generative, and algorithmic design. Frontiers of Architectural Research 2020; 9(2): 287.

Chaszar

Joyce

. Generating freedom: questions of flexibility in digital design and architectural computation. Int J Architect Comput 2016; 14(2): 167–181.

Nagy

Lau

Locke

, et al. Project discover: an application of generative design for architectural space planning. In: Proceedings of the symposium on simulation for architecture and urban design. Society for Computer Simulation International, p. 7.

Brown

Mueller

. Designing with data: moving beyond the design space catalog. In: Nagakura

Mueller

Tibbits

, et al. (eds). Disciplines and disruption - Proceedings catalog of the 37th annual conference of ACADIA, 2017, pp. 154–163.

Erhan

Chan

Fung

, et al. Understanding cognitive overload in generative design: an epistemic action analysis. In: Proceedings of the 22nd International conference of the association for computer-aided architectural design research in Asia, pp. 127–136.

Davis

. Generative design is doomed to fail. URL https://www.danieldavis.com/generative-design-doomed-to-fail/, accessed 12 Oct 2020.

Bradner

Iorio

Davis

. Parameters tell the design story: ideation and abstraction in design optimization. In: Proceedings of the Symposium on Simulation for Architecture & Urban Design. SimAUD ’14. San Diego, CA, USA: Society for Computer Simulation International.

Haymaker

Bernal

Marshall

, et al. Design space construction: a framework to support collaborative, parametric decision making. J Inf Technol Construct 2018; 23(8): 157–178.

10.

Thornton Tomasetti

. Design explorer 2, https://ttacmgithubio/DesignExplorer/2020 (accessed 29 Aug 2020 from).

11.

Sedlmair

Meyer

Munzner

. Design study methodology: reflections from the trenches and the stacks. IEEE Trans Vis Comput Graph 2012; 18(12): 2431–2440.

12.

Shireen

Erhan

Woodbury

, et al. Making sense of design space. In: C¸agdas

Ozkar

G¨ ul

, et al. (ed). Computer-aided architectural design. Future trajectories. Singapore: Springer. ISBN 978-981-10-5197-5, pp. 191–211.

13.

Woodbury

Burrow

. Whither design space? AI EDAM (Artif Intell Eng Des Anal Manuf) 2006; 20(2): 63–82.

14.

Woodbury

. Elements of parametric design. London: Routledge, 2010.

15.

Robert

. Grasshopper: algorithmic modeling for Rhino, https://wwwgrasshopper3dcom (2022, accessed 29 Nov 2022 from).

16.

Shi

. Performance-based and performance-driven architectural design and optimization. Front Architect Civ Eng China 2010; 4(4): 512–518.

17.

Sadeghipour Roudsari

Pak

. Ladybug: a parametric environmental plugin for grasshopper to help designers create an environmentally-conscious design. In: Proceedings of BS 2013: 13th Conference of the International Building Performance Simulation Association, 2013, pp. 3128–3135.

18.

Wortmann

Nannicini

. Introduction to architectural design optimization. Cham: Springer International Publishing, 2017, ISBN 978-3-319-65338-9. pp. 259–278.

19.

Brehmer

Munzner

. A multi-level typology of abstract visualization tasks. IEEE Trans Vis Comput Graph 2013; 19(12): 2376–2385.

20.

Pirolli

Card

. The sensemaking process and leverage points for analyst technology as identified through cognitive task analysis. In: Proceedings of International Conference on Intelligence Analysis, 2005, pp. 2–4.

21.

Munzner

. Visualization analysis and design. A K Peters Visualization Series. Florida, US: CRC Press, 2014.

22.

Stump

Yukish

Martin

, et al. The ARL trade space visualizer: an engineering decision-making tool. In: 10th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference. DOI: 10.2514/6.2004-4568.

23.

Clarich

Geremia

Parashar

, et al. Use of multivariate-data-analysis techniques in modefrontier for efficient optimization and decision making. In: 13th AIAA/ISSMO Multidisciplinary Analysis Optimization Conference.

24.

Daskilewicz

Rave

GBJ

. A computational framework to facilitate research in design decision support. J Comput Inf Sci Eng 2012; 12(021005).

25.

Yan

Qiao

, et al. A work-centered visual analytics model to support engineering design with interactive visualization and data-mining. In: 2012, the 45th Hawaii international conference on system sciences. IEEE, pp. 1845–1854.

26.

Balling

. City is planning via a multiobjective genetic algorithm and a Pareto set scanner. Optimization in Industry; 2: 103–112.

27.

Stump

Yukish

Simpson

, et al. Design space visualization and its application to a design by shopping paradigm. In: International design engineering technical conferences and computers and information in engineering conference, volume 2: 29th design automation conference, parts A and B, pp. 795–804.

28.

Ahlberg

Truve

. Tight coupling: guiding user actions in a direct manipulation retrieval system. In: Proceedings of the HCI’95 Conference on People and Computers X. HCI ’95. ISBN 0521567297. USA. Cambridge University Press, pp. 305–321.

29.

Marks

Andalman

Beardsley

, et al. Design galleries: a general approach to setting parameters for computer graphics and animation. In: Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pp. 389–400.

30.

Wittenburg

Lanning

Heinrichs

, et al. Parallel bargrams for consumer-based information exploration and choice. In: Proceedings of the 14th annual ACM symposium on User interface software and technology, pp. 51–60.

31.

Marin

Marsault

Mavromatidis

, et al. EC-CO-GEN: an evolutionary simulation assisted design tool for energy rating of buildings in early design stages to optimize the building form. In: Building simulation. New York: Springer, 2013.

32.

Beham

Herzner

Groller

, et al. Cupid: cluster-based exploration of geometry¨ generators with parallel coordinates and radial trees. IEEE Trans Vis Comput Graph 2014; 20(12): 1693–1702.

33.

Chen

Janssen

Schlueter

. Analysing populations of design variants using clustering and archetypal analysis. In: Real time-extending the reach of computation: eCAADe 2015; proceedings of the 33. International conference on education and research in computer aided architectural design in Europe, Vienna, Austria, 16-18 September 2015. eCAADe, 2015, vol 1, pp. 251–260.

34.

Doraiswamy

Ferreira

Lage

, et al. Topology-based catalogue exploration framework for identifying view-enhanced tower designs. ACM Trans Graph 2015; 34(6): 1–13.

35.

Chaszar

von Buelow

Turrin

. Multivariate interactive visualization of data in generative design. In: Attar

Chronis

Hanna

, et al. (eds). Proceedings of the symposium on simulation for architecture and urban design (simAUD 2016). ISBN 978-1365-05872-1. simAUD, pp. 223–230.

36.

Harding

. Dimensionality reduction for parametric design exploration. In: Adriaenssens

Gramazio

Kohler

, et al. (eds). Advances in Architectural Geometry 2016. Zurich, Switzerland: vdf Hochschulverlag AG an der ETH Zurich. ISBN 978-3-7281-3777-7, 2016, pp. 274–287.

37.

Goguelin

Flynn

Essink

, et al. A data visualization dashboard for exploring the additive manufacturing solution space. Procedia CIRP 2017; 60: 193–198.

38.

Fuchkina

Schneider

Bertel

, et al. Design Space Exploration Framework-A modular approach to flexibly explore large sets of design variants of parametric models within a single environment. In: Computing for a better tomorrow - Proceedings of the 36th eCAADe Conference, Computing for a better tomorrow - Proceedings of the 36th eCAADe Conference. CUMINCAD, vol 2, pp. 367–376.

39.

Harding

Brandt-Olsen

. Biomorpher: interactive evolution for parametric design. Int J Architect Comput 2018; 16: 144–163.

40.

Matejka

Glueck

Bradner

et al. Dream Lens: exploration and visualization of large-scale generative design datasets. In Proceedings of the 2018 CHI Conference on human factors in computing systems. CHI ’18, New York, NY, USA. ACM. ISBN 978-1-4503-5620-6, pp. 369:1–369:12.

41.

Mohiuddin

Ashtari

Woodbury

. Directly interactive design gallery systems: interaction terms and concepts. In: Computational studies on cultural variation and heredity. Springer, 2018, pp. 15–32.

42.

van Kastel

. Visual Analytics for Generative Design Exploration: an interactive 3D data environment for a computational design system facilitating the performance-driven design process of a nearly Zero-Energy sports hall. Master’s Thesis. TUDelft, 2018. URL https://resolver.tudelft.nl/uuid:ad6f454b-0e67-4664-88d4-87d2132a1f71

43.

Pan

Sun

Turrin

, et al. Design exploration of architectural geometries and structural performance for sports arenas based on som-clustering and structural performance simulation. In: Structures and Architecture-Bridging the Gap and Crossing Borders: Proceedings of the Fourth International Conference on Structures and Architecture (ICSA 2019), Lisbon, Portugal, 24-26 July, 2019. CRC Press, p. 342.

44.

Autodesk . Fusion 360. https://wwwautodeskca/en/products/fusion-360 (2020, accessed 29, Aug 2020 from).

45.

Autodesk . Project refinery beta. https://wwwautodeskcom/campaigns/refinery-beta (2020, Accessed 29 Aug 2020 from).

46.

Thornton Tomasetti

. Asterix. https://asteriskthorntontomasetticom/ (2020, accessed 29 Aug 2020 from).

47.

Thornton Tomasetti

. Thread. https://threadthorntontomasetticom/ (2020, accessed 29 Aug 2020 from).

48.

Garcia

Leitao

. Interfaces for design space exploration. In: Co-Creating the future: inclusion in and through design - Proceedings of the 40th conference on education and research in computer aided architectural design in Europe (eCAADe 2022) - volume 1, Ghent, 13-16 September 2022, pp 331–340 2022.

49.

Flager

Welle

Bansal

, et al. Multidisciplinary process integration and design optimization of a classroom building. J Inf Technol Construct 2009; 14(38): 595–612.

50.

Attia

Beltran

De Herde

, et al. “Architect friendly”: a comparison of ten different building performance simulation tools. In: Proceedings of 11th international building performance simulation association conference and exhibition. Tex.organization: International Building Performance Simulation Association (IBPSA), pp. 204–211.

51.

Joyce

. Web based data visualisation applied to creative decision making in parametric structural design. In: Proceedings of IASS annual symposia, 2015, pp. 1–12. Number: 20 tex.organization: International Association for Shell and Spatial Structures (IASS).

52.

Plant

Schaefer

. Visualisation and browsing of image databases. In: Multimedia analysis, processing and communications. Springer, 2011, pp. 3–57.

53.

Javed

Elmqvist

. Exploring the design space of composite visualization. In: IEEE pacific visualization symposium. IEEE, 2012, pp. 1–8.

54.

Erhan

Wang

Shireen

. Harnessing design space: a similarity-based exploration method for generative design. Int J Architect Comput 2015; 13(2): 217–236.

55.

Abuzuraiq

Erhan

. The many faces of similarity: a visual analytics approach for design space simplification. In: 25th International conference on computer-aided architectural design research in Asia: design in the age of humans, CAADRIA 2020. The Association for Computer-Aided Architectural Design Research in Asia.

56.

Woodbury

Mohiuddin

Cichy

, et al. Interactive design galleries: a general approach to interacting with design alternatives. Des Stud 2017; 52: 40–72.

57.

Inc TS . Spotfire, https://spotfiretibcocom/ (2020, accessed 29 Aug 2020 from).

58.

Software

. Tableau desktop, https://wwwtableaucom/products/deskto (2020, accessed 29 Aug 2020 from).

59.

Roberts

. State of the art: coordinated & multiple views in exploratory visualization. In: Fifth international conference on coordinated and multiple views in exploratory visualization (CMV 2007). IEEE, pp. 61–71.

60.

Wills

. Selection: 524,288 ways to say” this is interesting. In: Proceedings IEEE Symposium on Information Visualization’96. IEEE, pp. 54–60.

61.

Von Buelow

. ParaGen: performative exploration of generative systems. Journal of the International Association for Shell and Spatial Structures 2012; 53(4): 271–284.

62.

Kazemi

. Query: a prototype for visual search in interactive design galleries. Master’s Thesis. Communication, Art & Technology: School of Interactive Arts and Technology, 2021.

63.

Shneiderman

. The eyes have it: a task by data type taxonomy for information visualizations. In: Proceedings 1996 IEEE Sym. on visual languages. IEEE, pp. 336–343.

64.

Inselberg

. Multidimensional detective. In: Proceedings of VIZ’97: Visualization conference, information visualization symposium and parallel rendering symposium. Tex.organization: IEEE, pp. 100–107.

65.

Shneiderman

. Direct manipulation: a step beyond programming languages. In: Proceedings of the joint conference on easier and more productive use of computer systems. (Part-II): Human interface and the user interface-volume, 1981, p. 143.

66.

Nagy

Villaggi

Zhao

, et al. Beyond heuristics: a novel design space model for generative space planning in architecture. In: Disciplines and disruption - Proceedings catalog of the 37th annual conference of the association for computer aided design in architecture. ACADIA, 2017.

67.

Erhan

Abuzuraiq

Zarei

, et al. What do design data say about your model? A case study on reliability and validity. In: 25th international conference on computer-aided architectural design research in Asia: Design in the age of humans. CAADRIA, 2020.

68.

Yousif

Yan

Culp

. Incorporating form diversity into architectural design optimization. In: Disciplines and disruption - Proceedings catalog of the 37th annual conference of the association for computer aided design in architecture. ACADIA, 2017.

69.

Abuzuraiq

Alsalman

Erhan

. Shopping for game levels: a visual analytics approach to exploring procedurally generated content. In: Proceedings of the 15th international conference on the foundations of digital games, 15, Sep, 2020, pp. 1–4.

70.

McCurdy

Meyer

. GalStamps: analyzing real and simulated galaxy observations. In: 2019 IEEE visualization conference (VIS), 20 Oct 2019. IEEE, pp. 276–280.