3.1. What AI methods are primarily used in design innovation?

In this section, we present a design-centric AI typology identified through a comprehensive literature review. A total of 46 documents were reviewed and analysed to develop an AI typology for design and innovation. With the aim of supporting a deeper understanding of the various AI technologies applied in design practices, our novel AI classification framework is design-centric, distinguishing itself from existing technology-focused AI classifications, such as AI technologies across STEM fields (Reference Berman, Chubb and WilliamsBerman et al., 2023), AI methods in industry 4.0 (Reference Alenizi, Abbasi, Hussein Mohammed and Masoud RahmaniAlenizi et al., 2023), and data-type-based generative AI classifications (Reference Gozalo-Brizuela and Garrido-MerchanGozalo-Brizuela and Garrido-Merchan, 2023). Traditional AI categorizations, which focus on algorithmic perspective- such as how models learn data patterns and the types of training data used (e.g., supervised and unsupervised learning, deep learning, NLP, and CV) (Reference Mukhamediev, Popova, Kuchin, Zaitseva, Kalimoldayev, Symagulov, Levashenko, Abdoldina, Gopejenko, Yakunin, Muhamedijeva and YelisMukhamediev et al., 2022) fail to account for design practices. To address this gap, we reorient AI classification toward how models support design practice, treating data as a design material. The six types of design-centric AI methods are subcategories of three widely recognised computational methods- machine learning, deep learning, and generative AI (Reference Rane, Mallick, Kaya and RaneRane et al., 2024)- as follows: machine learning is divided into traditional machine learning (ML), and reinforcement learning (RL); deep learning is categorised into deep models in natural language processing, and deep models in computer vision; and generative AI is classified into large generative AI and generative AI (Figure 1). The detailed description of each AI method can be found in Table 1.

Table 1. Six types of design-centric AI methods for design innovation

First two types of design-centric AI methods are the subcategories of machine learning technologies: traditional machine learning (ML) and reinforcement learning (RL). Traditional ML models heavily rely on statistical techniques to explicitly learn data distributions and patterns, thereby improving accuracy (Reference Sharifani and AminiSharifani and Amini, 2023). These models are particularly suited for predicting values based on extensive input training data and features in design practices (Bodendorf and Franke, Reference Bodendorf and Franke2021; Golkarnarenji et al., Reference Golkarnarenji, Naebe, Badii, Milani, Jazar and Khayyam2019). In contrast, RL is based on a principled mathematical framework for reward-driven autonomous learning (Reference Arulkumaran, Deisenroth, Brundage and BharathArulkumaran et al., 2017) and is primarily adopted for decision-support tasks, such as predicting user activities, and engagement (Reference Wang and HuWang and Hu, 2024). Although some studies applied Traditional ML models for decision making tasks (Reference Zhang, Yang, Jiang, Nigam, Tomotake, Kenji, Levent and KaraZhang, 2019), these applications were often constrained to specific, predefined, and selected features extracted from complex design resources, often requiring significant input from data engineers. The primary applications of machine learning (ML) include recommendation systems and classification, which are used to suggest design artifacts and tag them, enabling the implementation of image search engines. The use of traditional ML in design practice is limited due to their reliance on labelled and structured data for supervised learning. As a result, these models have been primarily used for simpler tasks, such as recommendations, data analytics, feature engineering, and data aggregation in design document management (Reference Huo, Liu, Xiong, Xiao and ZhaoHuo et al., 2022). On the other hand, RL does not require a large, well-organised training dataset or extensive feature engineering performed by data engineers (Reference Wang and HuWang and Hu, 2024). However, its application remains limited to predicting user activities at a high abstract level due to computational complexity in certain scenarios and low generalisation performance.

Deep learning technologies are adopted differently in design practice depending on whether it is image-based symbolic reasoning or reasoning with language-model. Consequently, they are subcategorised into deep models in natural language processing (NLP) and deep models in computer vision (CV). Deep models in NLP aim to derive effective feature representations for symbolic and sequential data from large corpora to predict the next words (Reference Otter, Medina and KalitaOtter et al., 2021), often used for analysing user product reviews and extracting design insights from design databases (Reference Gammack, Akay, Ceylan and KimGammack et al., 2022; Reference Wang, Li and TsungWang et al., 2020). These language model-based works rely on the symbolic reasoning capabilities of these models, limiting their assistance to generating knowledge, such as text summarisation or annotating tags and labels. Therefore, fine-tuning for design-specific tasks often requires extensive datasets. On the other hand, deep models in CV aim to extract latent and abstract features from image collections to identify informative feature maps from sub-pixel data (Reference Chai, Zeng, Li and NgaiChai et al., 2021). Image-based symbolic reasoning is applied in design ideation, visual aspect recommendations, and rapid prototyping based on product specifications and requirements (Reference Zhou, Sun, Mu, Wu, Zhou, Wu, Zhang, Xi, Gunes and SongZhou et al., 2022). While Convolutional Neural Networks (CNNs) have transformed the conventional design practices (Reference Wu, Xing, Si, Dou, Wang, Zhu and LiuWu et al., 2020), deep models in CV remain somewhat restricted to specific design tasks, such as decision support and design feature analysis. Numerous off-the shelf AI adopt deep learning technologies to support designers. Adobe’s Neural Filters, for instance, leverage convolutional neural network (CNN)-based AI models to assist with photo editing tasks, including color correction, denoising, filter recommendations, and content-aware fill. Canva, on the other hand, has launched a CLIP-inspired deep learning model to recommend multilingual keywords for template labelling to assist creators.

The final two types of design-centric AI methods are subcategories of generative AI technologies: Generative AI (GenAI) and large generative AI. Distinguished by their underlying architectures, GenAI refers to early deep generative models, such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) (Reference Ruthotto and HaberRuthotto and Haber, 2021), whereas large GenAI models are based on transformer models (Reference Bandi, Adapa and KuchiBandi et al., 2023), including DALL-E, GPT-3, and Stable Diffusion. The two methods also differ in the scale of trainable parameters and training datasets, which directly impact generalisation and knowledge abstraction capabilities. Large GenAI can reproduce design artifacts from text prompts, whereas GenAI typically generates specific data samples closely tied to its training data patterns, with limited human-AI interaction. Unlike traditional models that predict class labels or dependent variables, GenAI creates plausible data points by mimicking data samples or distributions. For instance, GenAI has been employed to generate manufacturable components (Reference Zhang, Yang, Jiang, Nigam, Tomotake, Kenji, Levent and KaraZhang et al., 2019), or characteristic-conditioned 3D objects (Reference Krahe, Bräunche, Jacob, Stricker and LanzaKrahe et al., 2020). On the other hand, large GenAI is widely utilised for generating captions, images, and designs, ranging from rough sketches to replicas of masterpieces; (Jeon et al., Reference Jeon, Hong, Chen, Murakami, Li and Klenk2024 (Reference Gallega and SumiGallega and Sumi, 2023).

This is attributed to the ability of large language models (LLMs), which can process and generate a wide range of text, from design brief, social media posts to technical reports (Reference Goel, Shaer, Gu, Delcourt and CooperGoel et al., 2023). An interesting observation is that GenAI is predominantly applied within industrial design contexts, whereas large GenAI is mainly utilised in UI/UX design applications such as Figma. Despite the capability of both GenAI and large GenAI, their application in design practices remains limited. A well-known limitation of generative AI models is the instability of model training when limited datasets are used. Although large generative AI models can mitigate this issue through various knowledge transfer techniques, understanding new datasets remains a critical challenge, particularly in creative work. This limitation arises from insufficient training data, and it is unrealistic to assume that high-quality data will always be available for creative tasks.

In this section, six distinct design-centric AI methods are identified with specific design applications and challenges. The performance of traditional ML and RL heavily relies on specific features and statistical methods defined by data analysts. The design challenges are associated with adopting traditional ML and RL include labour-intensive data works and limited generalisation capabilities. Deep models in NLP and CV enhance the efficiency of design work by enabling the extraction of design insights, design classification, and recommendation. However, these methods bring design challenges, such as requiring extensive datasets, generating limited knowledge, and restricting the flexibility of deep models due to the integration of various types of multimodal design data. GenAI and large GenAI generate design outputs such as documents, and artifacts, thereby expanding their role beyond merely interpreting complex design data. The adoption of Gen AI is limited due to training bias and instability in optimisation whereas large GenAI has challenges, including high computational complexity.

While the six design-centric AI methods enable value creation, there are inevitable trade-offs, potential harms, and unintended consequences- such as destruction of other forms of value, such as social, moral, and environmental. It is because when data is selected, combined, contextualised and interpreted for AI methods, it represents particular versions of reality assembled for specific purposes- whether business plans, or political objectives (Reference JasanoffJasanoff, 2018). Moreover, due to the lack of dataset controllability and outcome explainability, designers must be more mindful of potential ethical issues (e.g., copyright, discrimination) when adopting large GenAI models. The next section will describe where and how each design-centric AI method is adopted.

3.2. How do AI methods support design practices?

This section explains how AI supports design practices, ranging from knowledge augmentation, knowledge generation, design generation, prediction (symbolic reasoning and numerical reasoning), decision making, and coordination. The detailed descriptions of each type are as follows:

1. Knowledge augmentation: Large GenAI enhance designers’ knowledge by providing quick access to extensive information, including design documents, user reviews, and market trends. Access to large datasets enables designers to make more informed decisions and stay updated with the latest developments in their field (Salminen et al., 2024).

2. Knowledge generation: Traditional ML, deep models in NLP, and large GenAI contribute to generating new knowledge by identifying patterns and correlations within complex datasets that may not be immediately apparent to designers. For instance, user interaction data can discover emerging behaviors, inspiring innovative design solutions that address unmet needs (Akay and Kim, 2021; Huang et al., 2023).

3. Design generation: GenAI and large GenAI assist in creating design outcomes by automatically generating a wide range of design alternatives (Solórzano Requejo et al., 2024). This accelerates the ideation process and facilitates the exploration of unconventional solutions that might otherwise remain unexplored.

4. Prediction (symbolic reasoning and numerical reasoning): Traditional ML and deep models in CV are employed to anticipate outcomes and simulate scenarios. Symbolic reasoning involves identifying design materials by predicting their labels. Image classification methods are adopted for symbolic reasoning in complex design datasets (Krahe et al., 2019). Numerical reasoning, on the other hand, predicts specific values or properties within design specifications (Reference Golkarnarenji, Naebe, Badii, Milani, Jazar and KhayyamGolkarnarenji et al., 2019) or forecast production costs (Mandolini et al., 2024).

5. Decision making: RL, Traditional ML, and deep models in CV facilitate decision-making by evaluating design options based on criteria such as cost, feasibility, and user impact. These methods assist designers in making more objective and efficient decisions (Bodendorf and Franke, Reference Bodendorf and Franke2021; Wang, X. and Hu, Reference Wang and Hu2024).

6. Coordination: Coordination focuses on enhancing communication and collaboration between team members. Deep models in CV and NLP, as well as large GenAI, support managing project timelines, assigning tasks, and monitoring progress to ensure alignment with project goals (Song et al., 2022). These methods also promote knowledge sharing within teams by organizing and retrieving relevant information when needed (Reference Gammack, Akay, Ceylan and KimGammack et al., 2022).

These assistance types, linked to six AI methods, illustrate how AI supports various design practices across the four design stages of the Double Diamond design process (Table 2, Figure 2). During the Discovery phase, AI is used to provide well-defined information and predictions. Reinforcement learning (e.g., DRL) and traditional ML approaches help identify patterns and guide strategic choices. Deep learning models for NLP (e.g., BiLSTM, BERT) provide insights from large textual datasets, supporting teams in understanding user needs, market trends, and constraints. In the Define phase, AI methods focus on coordination and expanding the design knowledge for effective collaboration. Deep learning models in computer vision (CNNs) and NLP (BERT) facilitate clearer communication within teams and assist in identifying requirements from user feedback and reference materials. Large-scale generative models (e.g., ChatGPT, DALL·E, Midjourney) can aid in creating personas—both text and visual—helping teams refine their design concepts and establish a shared vision more efficiently.

Figure 2. Design-centric AI Typology

Table 2. AI Methods and Assistance Types Across Design Phases

During the Develop phase, AI methods support the generation of detailed concepts and the refinement of designs. Deep learning models for computer vision, Gen AI (e.g., GANs), and large GenAI (e.g., Stable Diffusion) enable the creation of multiple design alternatives through iterative processes. Techniques like RBF-ANN assist with symbolic reasoning, while multimodal models (StyleCLIP, DALL·E, LLMs) combine text and images to facilitate rapid design exploration. These tools allow designers to experiment with various aesthetics, functionalities, and colours before finalizing the design outcomes. Finally, in the Deliver phase, AI adoption focuses on decision-making and prediction as the product moves toward implementation. Traditional machine learning methods (SVM, KNN, Kalman filters) and deep learning vision models (e.g., CNNs, ResNet-50) help predict performance, verify quality, and ensure seamless integration. Generative models (Stable Diffusion) and large language models (e.g., GPT-3) assist with final documentation, adjustments, and minor refinements before launch.

Figure 3. Mapping Design-centric AI Methods on Double-Diamond design process

AI methods are increasingly adopted to support designers’ divergent thinking in discovering and managing design knowledge during the develop and discover phases, followed by supporting convergent thinking in the deliver and define phases. Traditional ML and RL are evenly used across four design phases, whereas other methods are applied to specific phases. Deep models in CV are used in the discover and define phases, while deep models in NLP and GenAI are utilised in the develop and deliver phases. Large genAI is primarily used for divergent thinking in the discover and develop phases. This section demonstrates how AI methods evolve to address the different needs of each design stage, beginning with data-driven exploration and progressing toward high-fidelity design generation and final optimisation. By employing six design-centric AI methods—traditional ML, RL, deep models in NLP and CV, and large-scale generative models—design teams can accelerate their process, enhance decision-making processes, deliver more refined solutions, ultimately driving innovation in design.