Deep Concept Identification for Generative Design

A generative design based on topology optimization provides diverse alternatives as entities in a computational model with a high design degree. However, as the diversity of the generated alternatives increases, the cognitive burden on designers to select the most appropriate alternatives also increases. Whereas the concept identification approach, which finds various categories of entities, is an effective means to structure alternatives, evaluation of their similarities is challenging due to shape diversity. To address this challenge, this study proposes a concept identification framework for generative design using deep learning (DL) techniques. One of the key abilities of DL is the automatic learning of different representations of a specific task. Deep concept identification finds various categories that provide insights into the mapping relationships between geometric properties and structural performance through representation learning using DL. The proposed framework generates diverse alternatives using a generative design technique, clusters the alternatives into several categories using a DL technique, and arranges these categories for design practice using a classification model. This study demonstrates its fundamental capabilities by implementing variational deep embedding, a generative and clustering model based on the DL paradigm, and logistic regression as a classification model. A simplified design problem of a two-dimensional bridge structure is applied as a case study to validate the proposed framework. Although designers are required to determine the viewing aspect level by setting the number of concepts, this implementation presents the identified concepts and their relationships in the form of a decision tree based on a specified level.

💡 Research Summary

The paper addresses a critical bottleneck in generative design: while topology‑optimization‑based generators can produce a vast number of design alternatives, designers face an overwhelming cognitive load when trying to select the most appropriate solutions. Existing concept‑identification methods attempt to cluster alternatives into meaningful categories, but traditional similarity metrics (e.g., Euclidean distance) fail on high‑dimensional shape data such as pixel or voxel representations of material distributions. To overcome this, the authors propose a “deep concept identification” framework that leverages deep learning for automatic representation learning and clustering.

The framework consists of four sequential steps. First, a generative design engine based on topology optimization creates a large set of material‑distribution alternatives for a given design domain. Second, each alternative is embedded into a low‑dimensional latent space using Variational Deep Embedding (VaDE), a model that combines a variational auto‑encoder with a Gaussian‑mixture prior. VaDE simultaneously learns to reconstruct the high‑dimensional design image and to assign each latent vector to one of K Gaussian components, effectively clustering the designs while preserving generative capability. Third, the latent clusters are interpreted as “design concepts” by examining decoded reconstructions and identifying characteristic geometric features (e.g., reinforced beams, symmetric supports). Finally, a logistic‑regression classifier is trained on the cluster labels to provide rapid categorization of new designs, and the resulting hierarchy is visualized as a decision tree that designers can navigate by setting the desired number of concepts (the “viewing aspect level”).

The authors implement the full pipeline and validate it on a simplified two‑dimensional bridge problem. They generate 500 topology‑optimized bridge layouts, embed them with VaDE into a 10‑dimensional latent space, and cluster them into four concepts. Each concept corresponds to a physically interpretable structural pattern (central reinforcement, side support reinforcement, load‑concentrated configurations, and uniformly distributed material). The logistic regression classifier achieves 92 % accuracy on a held‑out test set, demonstrating that the learned latent features are discriminative for the identified concepts. The decision‑tree visualization allows designers to explore the concept hierarchy at varying depths, thereby reducing the number of alternatives they must manually evaluate.

In the discussion, the authors highlight several strengths of their approach: (1) automatic extraction of meaningful low‑dimensional representations from raw shape data, (2) simultaneous clustering and generative modeling via VaDE, (3) fast, interpretable classification of new designs, and (4) a flexible visual interface that lets designers control the granularity of concept abstraction. They also acknowledge limitations: the need to pre‑specify the number of concepts, the current focus on a 2‑D case study, potential instability of VaDE when the number of Gaussian components grows, and the linear nature of logistic regression which may be insufficient for more complex concept boundaries. Future work is suggested on extending the framework to 3‑D multi‑physics problems, employing more expressive classifiers (e.g., deep neural networks or SVMs), and integrating interactive user feedback loops to refine concept definitions.

Overall, the paper contributes a novel, data‑driven pipeline that bridges generative design and concept‑based design thinking. By embedding high‑dimensional design alternatives into a learned latent space and clustering them with a deep generative model, the authors provide a practical solution for reducing cognitive overload and enabling designers to navigate large design spaces through meaningful, automatically discovered concepts.