The boundaries of design are constantly being pushed forward by the integration of cutting-edge technologies. As computational methods advance, designers are equipped with new tools that enhance their creative potential, allowing for unprecedented exploration of complex design spaces. One such innovative approach is Feature Space Exploration, a method that transcends traditional parametric design limitations by harnessing the power of deep learning. In my past research at Carnegie Mellon University, I focused on developing a novel design exploration framework that leverages deep learning to create feature spaces. These spaces offer a more intuitive and comprehensive way to explore design solutions, moving beyond the restrictive nature of parametric design. The project presented here marks a significant step forward in how designers can interact with complex datasets and uncover new possibilities in their work. From Parameters to Features: A New Approach to Design Space Parametric design has long been a cornerstone of computational design, enabling designers to generate multiple variations of a concept by tweaking a set of predefined parameters. However, while this approach offers flexibility, it also imposes limitations. The parametric space often restricts the designer to a narrow set of possibilities, defined by the parameters themselves. This can hinder the exploration of more complex relationships between different design elements. To address this limitation, I developed an alternative approach centered around what I call the "feature space." Instead of being confined to the parametric variables, the feature space is created by extracting and analyzing design features using deep learning models. This shift allows creative practitioners to explore a richer, more interconnected design space, where relationships between features are naturally expressed. The Experiment: Constructing the Design Spaces This study involved the creation of a synthetic dataset consisting of 15,000 3D models, each generated through a parametric algorithm with five key parameters. These parameters included the height of the vessel, the width of the base, the width of the top opening, and the coordinates of control points that define the vessel's shape. Each design variant is represented as a vector, corresponding to a specific 3D model. Once the dataset was generated, I employed a Variational Autoencoder (VAE) to construct the feature space. VAEs are a type of generative deep neural network that abstracts input data into smaller, more manageable dimensions—known as the latent space. This latent space serves as the foundation for the feature space, capturing the complex relationships between different design features. Visualization: Bridging the Gap Between High-Dimensional Data and Design One of the key challenges in exploring both the parametric and feature spaces is visualization. High-dimensional data can be difficult to comprehend and navigate. A five-dimensional design space makes it hard for designers to compare models and to visualize and compare the characteristics, I employed a dimensionality reduction process to reduce the space to two dimensions and enable the objects to be plotted and compared to one another. The image below shows the overall process of visualizing the space using the t-Distributed Stochastic Neighbour Embedding (t-SNE) algorithm, a popular dimensionality-reduction algorithm for visualizing high-dimensional data. Once the dataset was generated, I employed a Variational Autoencoder (VAE) to construct the feature space. VAEs are a type of generative deep neural network that abstracts input data into smaller, more manageable dimensions—known as the latent space. This latent space serves as the foundation for the feature space, capturing the complex relationships between different design features. Once the VAE was trained, the encoder was used to extract the features of each vessel in the test dataset from 32,768 dimensions, the size of each voxelized vessel, into 128-dimensional vectors, the latent vectors. Consequently, the entire test dataset of the vessels is represented into vectors whose total shape is [3,000, 128]. This visualization is not just a technical achievement; it is a crucial tool for designers. By reducing the complexity of the data into a visual format, designers can more easily identify patterns, clusters, and relationships within the design space. This allows for more informed decision-making and opens up new avenues for creative exploration. Comparison: Parametric Space vs. Feature Space The analysis of the design spaces highlights significant differences between how parametric and feature spaces represent and organize design solutions. Figure 6 presents a 2D visualization of the feature design space generated by the Variational Autoencoder (VAE) model for the vessel dataset. In this figure, we observe that vessels with similar morphological characteristics are naturally clustered together. For instance, thinner vessels are predominantly located at the top right of the image, while larger, bulkier vessels occupy the lower left corner. This clustering pattern illustrates the VAE model’s capacity to understand and map the complex relationships between the design parameters and their resulting influence on the vessel's shape. Conversely, when examining the parametric space in Figure 7, we notice a different organizational structure. Although concave vessels are grouped at the bottom of the image, the clustering does not fully consider other critical parameters such as the height of the vessels. This limitation is inherent in the parametric design approach, which tends to treat each parameter independently rather than exploring the intricate relationships between them. As a result, the parametric design space often fails to capture the full complexity of the vessel forms, leading to an incomplete representation of the possible design outcomes. In contrast, the feature space (as shown in above) allows for a more nuanced and gradual transition in shape, concavity, height, and width, providing a more comprehensive understanding of how these features interact and evolve. To deepen this comparison, a clustering algorithm—Density-Based Spatial Clustering of Applications with Noise (DBSCAN) — is applied to both the parametric and feature spaces. Figure 8 illustrates the outcomes of this clustering process. In the parametric design space, I identified a total of seven clusters: three large and four small. However, this clustering reveals a significant drawback of the parametric space—it does not provide sufficient information for intuitively comparing design variants on a local scale. Even within the same cluster, the parametric space shows extreme variations in vessel forms, indicating a lack of cohesion and continuity in how the designs are grouped. On the other hand, the feature design space presents a more refined clustering structure with nine distinct clusters: six major clusters and three smaller ones. In the feature space, the transitions between different clusters are smoother, reflecting gradual changes in the forms as we move through the space (local changes) and across the entire visualization (global changes). For example, shorter vessels are consistently located at the top, while taller vessels occupy the bottom section. Moreover, moving horizontally across the space, we observe a shift from concave to convex shapes, offering a clear and intuitive representation of how the vessel forms change relative to their features. This cohesive clustering in the feature space allows designers to make more informed and localized comparisons between similar design alternatives, significantly enhancing the design exploration process. The Future of Design Exploration This work demonstrates that deep learning can significantly enhance the design exploration process, offering a new paradigm that goes beyond the limitations of parametric modeling. By shifting the focus from individual parameters to complex feature relationships, I developed a method that allows for a more comprehensive and intuitive exploration of design spaces. Feature Space Exploration represents a significant advancement in the field of computational design. By moving beyond the constraints of parametric modeling, new levels of creativity and innovation are unlocked, providing designers with tools that offer greater freedom and precision in their work. This approach not only changes how designers interact with their tools but also opens up new possibilities for what can be achieved. As technology continues to evolve, the integration of deep learning and computational design will undoubtedly play a crucial role in shaping the future of creativity. https://tcabezon.github.io/3Dexploration/?embedable=true For more details or to explore the resulting feature space, refer to the project website: https://tcabezon.github.io/3Dexploration/ or the research article: "Feature Space Exploration as an Alternative for Design Space Exploration Beyond the Parametric Space" which includes the whole research process that is behind this project. The boundaries of design are constantly being pushed forward by the integration of cutting-edge technologies. As computational methods advance, designers are equipped with new tools that enhance their creative potential, allowing for unprecedented exploration of complex design spaces. One such innovative approach is Feature Space Exploration, a method that transcends traditional parametric design limitations by harnessing the power of deep learning. In my past research at Carnegie Mellon University, I focused on developing a novel design exploration framework that leverages deep learning to create feature spaces. These spaces offer a more intuitive and comprehensive way to explore design solutions, moving beyond the restrictive nature of parametric design. The project presented here marks a significant step forward in how designers can interact with complex datasets and uncover new possibilities in their work. From Parameters to Features: A New Approach to Design Space From Parameters to Features: A New Approach to Design Space Parametric design has long been a cornerstone of computational design, enabling designers to generate multiple variations of a concept by tweaking a set of predefined parameters. However, while this approach offers flexibility, it also imposes limitations. The parametric space often restricts the designer to a narrow set of possibilities, defined by the parameters themselves. This can hinder the exploration of more complex relationships between different design elements. To address this limitation, I developed an alternative approach centered around what I call the "feature space." Instead of being confined to the parametric variables, the feature space is created by extracting and analyzing design features using deep learning models. This shift allows creative practitioners to explore a richer, more interconnected design space, where relationships between features are naturally expressed. The Experiment: Constructing the Design Spaces The Experiment: Constructing the Design Spaces This study involved the creation of a synthetic dataset consisting of 15,000 3D models, each generated through a parametric algorithm with five key parameters. These parameters included the height of the vessel, the width of the base, the width of the top opening, and the coordinates of control points that define the vessel's shape. Each design variant is represented as a vector, corresponding to a specific 3D model. Once the dataset was generated, I employed a Variational Autoencoder (VAE) to construct the feature space. VAEs are a type of generative deep neural network that abstracts input data into smaller, more manageable dimensions—known as the latent space. This latent space serves as the foundation for the feature space, capturing the complex relationships between different design features. Visualization: Bridging the Gap Between High-Dimensional Data and Design Visualization: Bridging the Gap Between High-Dimensional Data and Design One of the key challenges in exploring both the parametric and feature spaces is visualization. High-dimensional data can be difficult to comprehend and navigate. A five-dimensional design space makes it hard for designers to compare models and to visualize and compare the characteristics, I employed a dimensionality reduction process to reduce the space to two dimensions and enable the objects to be plotted and compared to one another. The image below shows the overall process of visualizing the space using the t-Distributed Stochastic Neighbour Embedding (t-SNE) algorithm, a popular dimensionality-reduction algorithm for visualizing high-dimensional data. Once the dataset was generated, I employed a Variational Autoencoder (VAE) to construct the feature space. VAEs are a type of generative deep neural network that abstracts input data into smaller, more manageable dimensions—known as the latent space. This latent space serves as the foundation for the feature space, capturing the complex relationships between different design features. Once the VAE was trained, the encoder was used to extract the features of each vessel in the test dataset from 32,768 dimensions, the size of each voxelized vessel, into 128-dimensional vectors, the latent vectors. Consequently, the entire test dataset of the vessels is represented into vectors whose total shape is [3,000, 128]. This visualization is not just a technical achievement; it is a crucial tool for designers. By reducing the complexity of the data into a visual format, designers can more easily identify patterns, clusters, and relationships within the design space. This allows for more informed decision-making and opens up new avenues for creative exploration. Comparison: Parametric Space vs. Feature Space Comparison: Parametric Space vs. Feature Space The analysis of the design spaces highlights significant differences between how parametric and feature spaces represent and organize design solutions. Figure 6 presents a 2D visualization of the feature design space generated by the Variational Autoencoder (VAE) model for the vessel dataset. In this figure, we observe that vessels with similar morphological characteristics are naturally clustered together. For instance, thinner vessels are predominantly located at the top right of the image, while larger, bulkier vessels occupy the lower left corner. This clustering pattern illustrates the VAE model’s capacity to understand and map the complex relationships between the design parameters and their resulting influence on the vessel's shape. Conversely, when examining the parametric space in Figure 7, we notice a different organizational structure. Although concave vessels are grouped at the bottom of the image, the clustering does not fully consider other critical parameters such as the height of the vessels. This limitation is inherent in the parametric design approach, which tends to treat each parameter independently rather than exploring the intricate relationships between them. As a result, the parametric design space often fails to capture the full complexity of the vessel forms, leading to an incomplete representation of the possible design outcomes. In contrast, the feature space (as shown in above) allows for a more nuanced and gradual transition in shape, concavity, height, and width, providing a more comprehensive understanding of how these features interact and evolve. To deepen this comparison, a clustering algorithm—Density-Based Spatial Clustering of Applications with Noise (DBSCAN) — is applied to both the parametric and feature spaces. Figure 8 illustrates the outcomes of this clustering process. In the parametric design space, I identified a total of seven clusters: three large and four small. However, this clustering reveals a significant drawback of the parametric space—it does not provide sufficient information for intuitively comparing design variants on a local scale. Even within the same cluster, the parametric space shows extreme variations in vessel forms, indicating a lack of cohesion and continuity in how the designs are grouped. On the other hand, the feature design space presents a more refined clustering structure with nine distinct clusters: six major clusters and three smaller ones. In the feature space, the transitions between different clusters are smoother, reflecting gradual changes in the forms as we move through the space (local changes) and across the entire visualization (global changes). For example, shorter vessels are consistently located at the top, while taller vessels occupy the bottom section. Moreover, moving horizontally across the space, we observe a shift from concave to convex shapes, offering a clear and intuitive representation of how the vessel forms change relative to their features. This cohesive clustering in the feature space allows designers to make more informed and localized comparisons between similar design alternatives, significantly enhancing the design exploration process. The Future of Design Exploration The Future of Design Exploration This work demonstrates that deep learning can significantly enhance the design exploration process, offering a new paradigm that goes beyond the limitations of parametric modeling. By shifting the focus from individual parameters to complex feature relationships, I developed a method that allows for a more comprehensive and intuitive exploration of design spaces. Feature Space Exploration represents a significant advancement in the field of computational design. By moving beyond the constraints of parametric modeling, new levels of creativity and innovation are unlocked, providing designers with tools that offer greater freedom and precision in their work. This approach not only changes how designers interact with their tools but also opens up new possibilities for what can be achieved. As technology continues to evolve, the integration of deep learning and computational design will undoubtedly play a crucial role in shaping the future of creativity. https://tcabezon.github.io/3Dexploration/?embedable=true https://tcabezon.github.io/3Dexploration/?embedable=true For more details or to explore the resulting feature space, refer to the project website: https://tcabezon.github.io/3Dexploration/ or the research article: "Feature Space Exploration as an Alternative for Design Space Exploration Beyond the Parametric Space" which includes the whole research process that is behind this project. https://tcabezon.github.io/3Dexploration/ "Feature Space Exploration as an Alternative for Design Space Exploration Beyond the Parametric Space"

This story contains new, firsthand information uncovered by the writer.

Unlocking Creative Potential: Exploring Latent Spaces in 3D Generative Design

Read My Stories

Too Long; Didn't Read

Make resilience your competitive advantage

Expanding Design Exploration: Exploring Feature Spaces Beyond Parametric Boundaries

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

Untitled Story

Unlocking Creative Potential: Exploring Latent Spaces in 3D Generative Design

The 💩 Poop Emoji's 10th Anniversary

10 Year Challenge — How Popular Websites Have Changed In China.

10 ways how designing a chat bot is different from designing an app

10 Ways To Make Your Innovative, Game-Changing, Synergistic Website Stand Out In 2018

Unlocking Creative Potential: Exploring Latent Spaces in 3D Generative Design

The 💩 Poop Emoji's 10th Anniversary

10 Year Challenge — How Popular Websites Have Changed In China.

10 ways how designing a chat bot is different from designing an app

10 Ways To Make Your Innovative, Game-Changing, Synergistic Website Stand Out In 2018

Light-Mode

Classic

Newspaper

Dark-Mode

Neon Noir

Minty

HN StartUps