A review of AI-based product shape generation technologies: trends, challenges, and future directions

Data identification and acquisition phase

Product shape data recognition algorithms can be classified into various categories based on their characteristics and implementation strategies. Shape-based methods include geometric invariants (e.g., corner features), descriptors (Gaussian, Fourier, wavelet), skeletonization, moment invariants, wavelet moments, and ICA. Performance benchmarks show that geometric invariant methods achieve 78–85% accuracy on standard datasets, while wavelet-based descriptors demonstrate 82–89% accuracy with 25% faster processing speeds. The implementation involves image preprocessing (grayscale conversion, filtering, binarization), feature extraction (edges, contours, corners, shape descriptors), classifiers (support vector machine (SVM), neural networks (NN), decision tree (DT)), and deep learning techniques. Algorithm selection depends on task type (classification, regression, clustering), data type (text, image), data scale and quality, algorithmic properties (stability, correctness, efficiency, scalability), and application requirements (interpretability, resource demands, bias-variance tradeoff).

Shape generation algorithm literature was categorized by algorithm type and visualized as a circular diagram using Illustrator cc2019 (Fig. 6), illustrating the application of product shape requirement extraction techniques in natural language processing (NLP). Studies by Yuan, Marion & Moghaddam (2022), Cong et al. (2023), Huang, Zhu & Huang (2024) and Yang, Liu & Chen (2023) provide effective methods for product shape requirement extraction using deep learning models (convolution neural networks (CNNs), bidirectional encoder representations from transformers (BERT), long short term memory (LSTM), vision transformers).

In text mining, Lai et al. (2022) used bidirectional long short-term memory (BiLSTM), conditional random field (CRFs), multilayer perceptron (MLP) and sequential perceptual engineering for text analysis and Apriori+ and SEM to explore relationships in user comments. Liu et al. (2022) proposed a tf-epa-based strategy for screening sentiment word pairs. In time series analysis, Zhang, Li & Zheng (2024) introduced Dynamic Mode Decomposition with Conditional Correlation (DMDCC) based on Decomposition with Conditional Correlation (DCC) theory, and Jing et al. (2022) proposed an Electroencephalogram (EEG)-driven programming (DP) prediction approach using BP neural networks. Classification algorithms include MF-SVM with EC model, graph-based encoding with SVMs, spherical harmonics with conditional variational autoencoders (CVAEs), and Riemannian manifold mapping with Binary Particle Swarm Optimization (BPSO).

Dimensionality reduction algorithms, such as user knowledge networks with Girvan-Newman algorithm and linear discriminant analysis (LDA), were used by Li, Wang & Sha (2023) to quantify product evolution. Sequence pattern mining techniques, like pattern tree mining (PTM) algorithm, were introduced by Tucker & Kim (2011). In regression analysis, Bushra et al. (2023) utilized GPR surrogate models.

Shape data extraction is constrained by data source characteristics, extraction methods, computational resources, and environment. Extraction methods include filter methods (median filtering, Wiener filtering, non-local means (NLM), BM3D) and learning-based models (convolutional neural networks (CNNs), recurrent neural networks (RNNs), generative autoencoders (GANs)). Feature extraction algorithms for image data (Local Binary Patterns (LBP), Histogram of Oriented Gradients (HOG)) and 3D model data (surface reconstruction, voxelization) are also utilized. Rule-based reasoning algorithms, such as expert systems, are applied as well.

Li, Roy & Saltz (2017) established an evolution model integrating user knowledge network, Girvan-Newman algorithm, LDA, information axiom, and evolutionary graphs. Additionally, Shao et al. (2020) employed wadaptiveos-elm for dynamic simulation data mining.

Current technologies enhance designer efficiency by automatically extracting requirements from user reviews via deep learning (ResNet-50+BERT) and incorporating user feedback into the design process, promoting a shift from “designer-led” to “user co-creation.”

Data analysis phase

We utilized MATLAB 2023b (The MathWorks, Natick, MA, USA) to develop cyclic network graph code, encompassing parameter definition, adjacency matrix creation, network visualization, node setup, cyclic graph definition, and label addition. Keywords from selected literature were compiled into an Excel file and imported into the code to generate the effect depicted in Fig. 7. Shape generation data analysis is a complex process that relies on various tools and methods to achieve final design goals. The process varies with design objectives (e.g., bionic design, particle swarm optimization), data types (predictive, inferential, descriptive), and analytical tools (simulation, visualization, statistical analysis). According to Abiodun et al. (2018), deep learning approaches in shape analysis demonstrate superior performance with classification accuracy ranging from 85% to 97% depending on dataset complexity and size.

Current product shape analysis algorithms are diverse, including constraint-based solutions (Ren et al., 2024). comparative analysis, edge detection, Hough transform, contour tracking, shape matching, and deep learning. Real-time feedback and adjustment mechanisms are crucial for off-standard inspection to ensure process stability and exception handling. Diagnostic analysis delves into data causes and correlations. Perceptual engineering maps user perception to product design features, and combining it with neural style migration enhances model efficiency. Grey correlation and hierarchical analysis clarify relationships between shape elements and consumer perception (Chen, Song & Ge, 2023).

Lin et al. (2013) proposed a probabilistic factor graph model using Markov chain Monte Carlo (MCMC) sampling for 2D pattern coloring. Exploratory data analysis employs visualization and statistical methods to uncover dataset patterns, with factor analysis aiding dimensionality reduction. TOPSIS determines demand weights for neural network evaluation. Yuan, Marion & Moghaddam (2022) developed dynamic data envelopment analysis (DDE) and dynamic multi-objective data envelopment analysis (DMDE) models, while Sharma (2023) used DSE for design scenario analysis. Chen, Song & Ge (2023) introduced ‘Mutational Reasoning on Graphs’ with probabilistic self-attention. Yang & Xiao (2011) created a 3D model feature similarity tool, and Yang et al. (2023) proposed a second-order method for Fisher matrix computation. Wang et al. (2022) presented a shape topology design method based on Quantization Theory I. For predictive analysis, neural networks and regression models are used, including BP, GNN, Bayesian, higher-order, and linear networks. Hayakawa, Noji & Kato (2025) applied regression models for design space and tyre tread prediction. Decision trees, such as DTM classifiers, also offer predictive capabilities.

Statistical inferential analysis often uses SEM. Choi et al.’s (2024) Spnr framework incorporates semantic mask diffusion models, and Ghosh et al.’s (2016) network-centered design uses SEM for parameter estimation. Common methods for design data relationships include correlation analysis, multiple linear regression, NA-IPA, and IGDT-MFP.

Data transformation phase

Data mapping is a key process designed to ensure compliance, eliminate data redundancy and facilitate analysis by depicting and visualising correlations between data component fields and integrating them into a unified schema or database (Zhao et al., 2023). In the product design domain, data mapping is particularly important for shape generation, involving the transformation (Wu, 2021), integration and standardisation of data from different sources, formats and structures to support various attributes of the product design. After selecting and categorising the relevant techniques, we have used Illustrator CC 2019 to map the shape generation algorithms and models applied in the current data transformation phase (shown in Fig. 8).

The product shape data mapping mechanism consists of a series of algorithms designed to transform shape data into a form that facilitates analysis (Khosravi et al., 2024), optimisation and design generation. These algorithms can be classified into four broad categories based on the type of method: (1) Feature-based mapping methods aim to extract geometric features (e.g., contour lines, axes of symmetry), topological features (e.g., number of holes, connectivity relationships), or physical features (e.g., material properties, stress distributions) from the data, and construct a feature vector space for mapping. Typical methods include morphological graph methods (for cross-domain mapping through graph matching) and fuzzy set theory (for dealing with the ambiguity of feature boundaries). (2) Model-driven mapping methods, which realise data transformation through a library of predefined mapping rules (e.g., ontologies, patterns or templates). For example, both Theory of Inventive Problem Solving (TRIZ) contradiction matrix and quality function deployment (QFD) quality house map functional requirements to structural or design parameters. Integrated QFD-TRIZ methodologies demonstrate significant improvements in innovative product design, with studies showing 30–40% reduction in design cycle time and 25% improvement in design solution quality compared to traditional methods (Sharma & Mishra, 2020). (3) Learning-based mapping methods, which use training data to learn input-output mapping relationships, typical algorithms include neural networks, decision trees, random forests and graph neural networks. These methods require a large amount of labelled data and have poor interpretability due to their ‘black box’ nature. (4) Theory-guided mapping methods such as FBS theory, which establishes a hierarchical mapping from function to behaviour to structure, and Effect-based System Method (ESM) effect theory, which optimises shape parameters based on physical effects (e.g., thermal expansion and contraction). It can be seen that traditional design data conversion methods are still necessary, and that relying only on AI for design is not yet a complete substitute for the applicability of traditional models, especially in highly creative projects or projects involving human subjects, where traditional designer-led methods are still advantageous (Hu, Fu & Zhao, 2024).

Currently, there are three types of mapping relationships: function-to-structure mapping, feature-to-image mapping, and requirement-to-function relationship mapping. Function-to-structure mapping focuses on three types of mapping relationships: (1) function-to-structure mapping, such as relying on theoretical frameworks such as TRIZ, QFD, and FBS. Liu et al. (2025) constructed a biologically-inspired design model for innovation based on TRIZ. Mettas (2010) utilised a hierarchical design structure matrix (HDSM) to enhance device performance. Geiger & Sarakakis (2016) set reliability through design for reliability (DfR). Endress, Rieser & Zimmermann (2023) uses effects-based system approach (ESM) to optimise component geometry. QFD applications in product design show significant improvements in converting customer requirements into engineering characteristics, with studies reporting 20–35% reduction in development time and 15–30% improvement in customer satisfaction scores (Shin, Shin & Kang, 2023). (2) The feature-to-image mapping aspect covers morphology map approach, symbolic semantics theory (Liu & Yang, 2022), text-to-shape transformation, fuzzy set theory and KJ method. (3) Requirements-to-function relationship mapping includes functional structure concept network (FSCN), rapid application development (RAD), generative axiomatic design (PCGA-DLKE), and axiomatic design extended version, which provide strong support for requirements-to-function mapping (Wang, Liu & Yang, 2022).

Current technologies can break through the limitations of the designer’s personal experience, e.g., learning mapping (e.g., GNN) to obtain style characteristics from case study learning, or function-structure mapping (e.g., TRIZ contradiction matrices) to transform abstract requirements into concrete design parameters, e.g., ‘lightweighting’ to automatically correlate with material selection and topology optimisation.

Shape extraction phase

Currently, product shape extraction methods are profoundly influenced by data sources and tools. To systematically analyze these methods, we first clustered the keywords in the literature and categorized them into five major groups based on differences in objectives and methodologies: edge-based, region-based, contour-based, data reduction-based shape extraction methods, and other shape extraction methods, encoded as 1 to 5 respectively. Subsequently, we counted the names and literature quantities of specific shape extraction methods within each category, distinguishing between standalone and combined uses, and visualized this information through a circular heatmap (as shown in Fig. 9). The implementation steps of the circular heatmap include: converting the data matrix into an annular distribution in the polar coordinate system, determining the position of each data block using polar coordinate formulas, mapping numerical values to color spaces, creating concentric rings to represent different data layers, and adding annotations to the angular and radial axes.

Edge-based shape extraction: Edge detection, a fundamental technique in image processing, is widely used to identify significant changes in images. The Harris corner detection algorithm demonstrates superior performance compared to SUSAN corner detection algorithm in comprehensive evaluations, with improved accuracy in distinguishing between edges and corners through differential corner score analysis (Emmert-Streib et al., 2021). Ji et al. (2020) utilized the Geometric Decontouring Network (GDCNet) to effectively eliminate false contours. Ma & Kim (2014) proposed the predictive data-driven product family design (PDPF) model, incorporating the Harris corner detection algorithm. Stoler, Lorusso & Capodieci (2006), based on design-based metrology (DBM), employed the Roberts operator to extract patterns, enabling an automated process. Yang et al. (2024) introduced a mesh denoising (EMD) and real mesh noise generation (RMNG) model based on the rotated squared model, addressing the challenge of feature-preserving mesh denoising.

Region-based shape extraction: Research on region-based shape extraction is relatively scarce and primarily combined with other methods for data acquisition. Rios et al. (2021) validated autoencoder representations, enhancing optimizer performance by providing potentially complementary degrees of freedom.

Contour-based shape extraction: Image segmentation is a key contour extraction technique. Bae, Tai & Zhu (2017) proposed an image segmentation model using the L-1 variant of Euler’s elastica energy as boundary regularization, effectively representing curvature. In feature selection, Cdna Microarray adopted multiple features to avoid treating noise as individual pixels and processed poor-quality spots through an adaptive adjustment algorithm. Zhao, Tang & Gong (2024) introduced a novel curvature-driven multi-stream graph convolutional neural network (CDMS-Net) architecture. Huang et al. (2021) utilized VGG16 for feature extraction, enhancing the efficiency and generation speed of local image style transfer. VGG16 neural network achieves 92.7% top-5 test accuracy on ImageNet dataset containing over 14 million images belonging to 1,000 classes. VGG16-XGBoost hybrid models demonstrate superior performance with accuracy of 0.97 and weighted F1-score of 0.97 in medical image analysis tasks (Howland et al., 2023).

Data reduction-based shape extraction: Extraction methods based on data dimensionality reduction are more of a way of thinking, utilizing mathematical or statistical methods to reduce the number or dimensions of data features while preserving important information from the original data.

Other innovative methods: Furthermore, scholars have innovated various methods and models tailored to different extraction purposes. Ahmad, Hassan Amin & Khan (2010) extended the study of grammatical style definitions using Scale-Invariant Feature Transform (SIFT). Chen et al. (2023) proposed a novel deep learning module called “Variant Reasoning on Graphs” to effectively utilize variational knowledge. He et al. (2016) applied Gaussian filtering, segmentation, and Gaussian-shaped Fast Fourier Transform (FFT) to fMRI image reprocessing. Fotopoulou, Oikonomou & Economou (2019) designed linear matrix operators through optimal projection, proposed a graph-based encoding technique, and utilized Support Vector Machines (SVM) for category prediction.

Technological advancements and applications: The technology at this stage can assist designers in working away from high-performance workstations. For example, algorithms like GDCNet can reduce noise while preserving sharp features (e.g., product corners), solving the problem of excessive smoothing caused by traditional denoising. Additionally, the automatic conversion from point cloud to parametric model (e.g., L − 1 Eulerian elastic model) supports the direct use of 3D scanning data in CAD modeling.

Shape generation phase

In the shape generation process, data transformation and mapping play a crucial role in bridging the gap between abstract design concepts and concrete geometric representations (Fazeli & Peng, 2022) We reviewed and defined shape generation algorithms in literature, organized these algorithms using Excel, and created a stacked circular diagram (Fig. 10). The algorithms are categorized into three main types: graph-based generation algorithms, technical feature-based algorithms, and other algorithms. Graph-based mapping models aim to represent shapes and their components using graph structures, where nodes represent geometric primitives (e.g., points, lines, arcs) and edges represent topological relationships (e.g., adjacency, connectivity). Transformations such as scan conversion and region filling are applied to these graphs to generate final shapes. Recent studies show that these transformations achieve computational complexity ranging from O(n) for simple line algorithms to O(n²) for complex polygon filling operations, with modern implementations achieving processing speeds of 10⁴−10⁵ operations per second (Alzubaidi et al., 2021). Technical feature-based modeling focuses on capturing and processing technical features such as dimensions, tolerances, material properties, and functional requirements. They typically employ parametric modeling and optimization techniques to generate shapes satisfying specific engineering constraints. Examples include Range-GAN, CycleGAN, FCGAN for fault diagnosis, and CGAN-CNN hybrids (combining conditional GANs with CNNs for topology optimization and performance prediction).

Figure 10 shows that graph-based generation algorithms are primarily used for generating basic shapes, including line scan conversion, arc generation, polygon scan conversion, and region filling methods. Techniques such as Bresenham’s algorithm, polar coordinate algorithms, and boundary clipping are involved. Empirical evaluations demonstrate that Bresenham’s algorithm achieves pixel-level accuracy with computational efficiency improvements of up to 40% compared to traditional floating-point methods (Foley, 1996). Technical feature-based algorithms are widely applied in 3D data generation and parametric shape generation. With rapid advancements in deep learning algorithms, shape generation methods based on latent variables and generative adversarial networks (GANs) have gained significant attention. In latent variable models, the conditional deep generative model Range-GAN proposed by Heyrani Nobari (2022) addresses sparse condition challenges in data-driven inverse design, enabling automated design synthesis with range constraints. GAN modeling methods are often combined with other approaches, such as CycleGAN proposed by Cabezon Pedroso, Ser & Díaz-Rodríguez (2022) (a GAN architecture integrated with other methods for style transfer or domain adaptation in shape generation). Additionally, Wang & Xue (2024) utilized the fuzzy clustering generative adversarial network (FCGAN) model (enhancing GAN capabilities in pattern recognition tasks related to shape analysis through fuzzy clustering) for fault diagnosis. Herath & Haputhanthri (2021) combined conditional generative adversarial networks (CGANs) with convolutional neural networks (CNNs) for topology optimization and performance prediction. Comparative studies demonstrate structural optimization efficiency gains of 25–35% compared to conventional finite element analysis methods, with solution convergence achieved in 60–80% fewer iterations. Other algorithms include parametric shape generation as a research hotspot. Parametric mapping models use coordinate systems (e.g., spherical, Euclidean, polar) and mathematical functions to parameterize shapes. Various shapes can be generated and modified by adjusting parameters (McKay & de Pennington, 2022). Main methods involve spherical, Euclidean, and polar coordinate parameterization. Hybrid modeling combines elements from multiple categories to leverage their respective advantages. For instance, integrating graph-based representations with technical feature-based optimization. Another example combines FCGAN with parametric models to enhance shape diagnosis and generation capabilities.

General evaluation factors for shape generation algorithms include: Quantitative and qualitative metrics: Depending on application scenarios, either or both may be combined. For example, quantitative metrics for accuracy and efficiency, and qualitative metrics for aesthetics and design innovation. Benchmark testing: Comparing performance against existing algorithms or industry standards. User feedback: Incorporating feedback from designers and engineers to assess practical usability and value. Computational resources: Considering the trade-off between model complexity and computational demands (Ma et al., 2024).

The technologies in this stage can assist designers to focus on creative expression rather than technical realization, such as CycleGAN to realize automatic conversion of sketches to high-precision models, Range-GAN to generate diverse solutions under constraints, expanding design possibilities, and helping enterprises to deploy an AI-assisted design pipeline to achieve a 15–30% reduction in load.

Decision-making assessment phase

Shape generation decision-making algorithms represent a core component of design automation and intelligence (Arbabi et al., 2022). Their primary value lies in assisting designers to efficiently generate product shapes that meet functional requirements and constraints through algorithmic means. The development of this field has always been closely integrated with optimization algorithms, forming a multi-layered and multi-paradigm technical system.

Core algorithms include bio-inspired algorithms and data-driven algorithms. Bio-inspired algorithms encompass genetic algorithms (GA) (which simulate biological evolution through mechanisms of population initialization, crossover, mutation, and selection to perform global searches in shape spaces, progressively approaching optimal solutions), Empirical studies demonstrate that GA-based shape optimization achieves convergence rates of 85–95% within 200–500 generations, with parallel implementations reducing computational time by 40–60% compared to sequential approaches (Gad, 2022). Ant colony optimization (ACO) (which draws on ant foraging behavior to construct path optimization models for shape combinations, suitable for multi-objective shape design problems), particle swarm optimization (PSO) (based on swarm intelligence particle collaboration mechanisms, balancing exploration and exploitation through information sharing and local search to enhance global optimization efficiency), Performance benchmarks show that PSO algorithms typically converge 30–50% faster than traditional genetic algorithms, with multi-objective PSO variants achieving Pareto front coverage of 90–95% in engineering design problems. Simulated annealing (SA) (which introduces thermodynamic annealing processes, accepting probabilistic inferior solutions to break through local optima and enhance global exploration capabilities in shape design). Data-driven algorithms include generative adversarial networks (GANs), such as those developed by scholars like (Goodfellow, Bengio & Courville, 2016) which can create novel designs by learning from existing design data (Wang et al., 2023).

During the decision-making phase, we reviewed the application history of relevant algorithms and models. As shown in Fig. 11, since 2002, active learning frameworks and real-coded genetic algorithms have become widespread, driving the transition of shape generation from rule-based to data-driven approaches. After 2012, academic research shifted its focus to the mapping relationships between functional and physical domains, proposing multi-criteria decision-making methods (such as quality function deployment (QFD), ideal final result (IFR), artificial neural networks (ANN), axiomatic design (AD), etc.). Notable achievements include the decision-making model proposed by Song et al. (2024), which integrates behavioral analysis, failure mode and effects analysis (FMEA), and TRIZ theory. In terms of evaluation method innovation, Yu et al. (2025) employed the Multinomial Logit (MNL) model to quantify consumer acceptance of new designs, providing market feedback for design iteration (Chen & Xu, 2024; Chen & Bian, 2024). Ranscombe, Kinsella & Blijlevens (2017) introduced the holistic shape analysis (HSA) method, combining 3D geometric comparison tools to extract shape difference features and support design optimization decisions.

Since Adeli (2002) first applied neural networks in the field of structural engineering research in 1989, the role of neural networks in the design industry has become increasingly prominent. Currently, neural network-based shape generation decision-making algorithms are more comprehensive and scientific, examples include the niching artificial fish swarm algorithm, multi-level artificial neural networks, and multi-layer perceptron genetic algorithm neural networks. In actual decision-making processes, decision-making algorithms generally form closed-loop interactions with evaluation methods, innovating around decision-making objectives and evaluation criteria. There are three main approaches: the generate-evaluate cycle (e.g., GANs generate candidate shapes—HSA/MNL evaluation—feedback to optimize the generator), multi-criteria decision-making (QFD defines requirement weights—PSO/GA multi-objective optimization—ANN verifies performance), and interdisciplinary integration (Game Theory (e.g., Nash equilibrium) optimizes multi-stakeholder design objectives, Markov Decision Processes model design state transitions). Techniques at this stage can assist designers in avoiding unidimensional decision-making pitfalls, such as PSO + QFD frameworks to simultaneously optimize cost, performance and user satisfaction. Another example is the combination of FMEA and behavioral analysis to identify potential failure modes early in the design.

Optimisation iteration phase

Product shape optimization algorithms constitute a complex, multidisciplinary field focused on achieving functional, efficient, and aesthetic balance in product geometry via mathematical methods (Sheikh et al., 2022). By judiciously selecting algorithm combinations, effective product shape design generation and shape data analysis can be conducted, enhancing design efficiency and product quality (Ebrahimi & Jahangirian, 2017). We analyzed algorithm usage frequency during the optimization iteration phase using mountain landscape visualization code, primarily employing the surf function to render landscapes with color, axis, and label adjustments (Fig. 12). The technological evolution and cross-disciplinary integration of this field follow these trajectories:

Algorithmic system innovation: As cross-disciplinary design research deepens, shape optimization algorithms become increasingly critical in the design generation process. Optimization frameworks based on Poisson equations, coupling Boundary Element Method (BEM), FEM, and LSM, enable high-precision modeling of physical field distributions and structural characteristics, demonstrating advantages in fluid mechanics optimization (Moosavi, Jablonka & Smit, 2020). Computational fluid dynamics optimization using coupled BEM-FEM approaches achieves accuracy improvements of 15–25% while reducing computational time by 30–45% compared to traditional single-method approaches (Mirjalili et al., 2023). Mirjalili & Lewis (2016) proposed novel swarm intelligence algorithms like WOA and SOA, offering new paradigms for multi-objective optimization. These algorithms simulate biological swarm behaviors, requiring precise parameter tuning but exhibiting strong global search capabilities for high-dimensional complex problems. WOA demonstrates superior performance on 29 benchmark functions with convergence rates 20–35% faster than traditional genetic algorithms, while achieving solution accuracy within 98% of global optima for engineering design problems.

Integrated algorithm applications: When applying optimization algorithms, factors such as data quality, algorithmic complexity, and computational resources must be considered. Integrated algorithm applications are now mainstream, including deep fusion of GA with neural networks for data-driven optimization efficiency and combined applications of PSO with Adam algorithms, performing well in performance prediction and hybrid optimization. Notable studies include Li et al.’s (2021) PSO-SVM prediction model and Tan et al.’s (2023) PSO-Adam hybrid optimization strategy.

Topology optimization technologies: Since 2019, deep learning-accelerated topology optimization has emerged as a hotspot, with three directions: preprocessing enhancement, system-level optimization, and numerical method innovation (e.g., Regenwetter Lyle’s solver, increasing efficiency by >40%).

Emerging research areas: Uncertainty Optimization: Kim (2023) modeled uncertainty in product degradation signals, while Ao et al. (2023) proposed a degradation-based multi-objective image optimization framework to enhance design robustness.

Human-computer collaborative design: Zhao, Sharudin & Lv (2024) combined Kansei Engineering with WOA algorithms, pioneering emotion-driven shape generation and bridging affective computing and optimization.

Evaluation system: Product shape optimization algorithms adopt a multi-dimensional evaluation system, including objective function formulation (balancing mechanical performance, material use, and costs), constraint handling mechanisms, convergence metrics, computational complexity, and robustness verification.

Current technologies promote a shift from deterministic optimization to resilient design, e.g., the electronics industry adopting impact optimization frameworks. They solve NP problems using group intelligence algorithms like Whale Optimization Algorithm (WOA) and enhance human-computer interaction (Kansei+WOA), incorporating emotional data into optimization goals to boost product emotional value.

Design concept evaluation phase

Design concept evaluation has become a critical phase in product development, systematically assessing early-stage creative concepts across feasibility, innovation, user preferences, functionality, and emotional appeal (Xue & Wu, 2022). Traditionally dominated by qualitative methods (e.g., focus groups, expert reviews), the field now integrates multi-criteria decision-making (MCDM) approaches like Analytic Hierarchy Process (AHP), Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS), and fuzzy logic to enhance objectivity.

Recent advancements focus on multimodal evaluation, incorporating neurophysiological data (eye-tracking, EEG, functional near-infrared spectroscopy (fNIRS), galvanic skin response (GSR)) to model implicit cognitive-emotional responses. This fusion of physiological signals and subjective feedback enables data-driven, user-centered refinement of design solutions.

As a pivotal bridge between ideation and realization, design concept evaluation supports evidence-based decision-making (Sharma, 2023), reduces development risks, and informs preference modeling. This article positions it as a foundational prerequisite for modeling strategies, integrating multimodal metrics to quantitatively evaluate concepts and guide iterative design optimization.

Performance prediction: AI-driven innovative form design

AI-assisted design technology enables the generation of product forms that meet specific stylistic or functional requirements, Wang et al. (2024b) assisting designers in precisely controlling key parameters such as geometric features, Miao et al. (2024) material properties, and loading conditions. Additionally, by integrating dimensionality reduction algorithms (e.g., t-Distributed Stochastic Neighbor Embedding (t-SNE), Uniform Manifold Approximation and Projection (UMAP)) with Bayesian optimization, it effectively addresses the complexities of high-dimensional design spaces, accelerating design iteration (Arjomandi Rad, Cenanovic & Salomonsson, 2023). In terms of real-time performance prediction, the deployment of edge computing nodes for distributed inference enables real-time performance prediction and optimization, meeting the demands of instant feedback scenarios such as additive manufacturing and Augmented Reality/Virtual Reality (AR/VR). At the level of design paradigm innovation, generative adversarial networks (GANs) mine potential design patterns by analyzing vast amounts of form data, further enhancing immersive design experiences. Furthermore, while traditional trial-and-error-based design is gradually being replaced by AI-driven methods, human-centric innovation projects (e.g., QFD, FBS models) still rely on traditional design logic. In the future, it is recommended to establish a cross-domain database that integrates multi-source data such as materials, processes, and user behavior, utilizing transfer learning to address data sparsity issues. Moreover, developing explainable AI modules will provide transparent decision-making support for safety-critical designs (e.g., medical devices, aerospace) (Chen et al., 2023).

Reverse design: multimodal data-driven technology upgrade path

Reverse design technology, centered on data-driven approaches and intelligent algorithms, enables precise transformation and innovative optimization from physical objects to digital models (Zhang & Chen, 2024). In terms of multimodal data acquisition and digital reconstruction, it leverages computer vision-based high-precision 3D scanning, point cloud processing, and deep learning algorithms to construct an integrated “scan-design-manufacture” platform.

Regarding multimodal data fusion, between 2018 and 2020, the field of physiological signal acquisition encountered interference issues in the fusion of EEG/ECG/EMG modalities. However, after 2023, the adoption of parameter-free modal modulation modules eliminated such interference, resulting in a 40% improvement in reconstruction accuracy (Hou, Tuerhong & Wushouer, 2023).

In terms of application innovation in vertical fields, a library of parametric models and a matrix of style transfer algorithms have been established to support topology optimization and innovative form generation based on 3D reconstruction (Xu et al., 2023). For instance, Tesla achieved a 15% weight reduction in battery packs and a 30% improvement in thermal management efficiency through topology optimization and generative design.

In the future, by combining multi-objective optimization algorithms and integrating “top-down” system modeling with "bottom-up" generative design, reverse design technology will promote the co-evolution of form, structure, and function.

Structure optimization: deep integration of algorithms and manufacturing techniques

The field of structural optimization is currently undergoing a paradigm shift driven by the deep integration of algorithms and additive manufacturing technologies. By combining topological optimization frameworks with the global search capabilities of genetic algorithms and the nonlinear mapping advantages of neural networks, and taking into account the constraints of additive manufacturing processes such as overhang angle limitations, it achieves collaborative optimization of lightweighting and load-bearing performance. The algorithms in this field have evolved from early homogenization methods that relied on manual input (prior to 2015) to the current stage where, after 2020, the Bi-directional Evolutionary Structural Optimization (BESO) method combined with AI tools enables form-force collaborative design (reducing iterations by 70%). In 2023, the Time-Varying Concurrent Topology Optimization (TVCTO) framework supports parallel optimization of three variables (macroscopic density, microscopic structure, and deformation parameters), resulting in a 60% reduction in computational load (Chen et al., 2023). Future developments in this area will focus on the construction of cloud-based collaborative platforms, integrating multi-physics simulations to create a closed-loop process of “digital twin-algorithmic optimization—virtual verification.” This will significantly enhance structural efficiency and shorten the R&D cycle in fields such as aerospace and automotive. The core advantages of this approach lie in multi-scale collaborative design, AI-driven intelligent iteration, and manufacturing integration.

Technological revolution and multidimensional challenges

As the core driving force of modern technology, artificial intelligence (AI) and machine learning (ML) have enabled intelligent product form generation design. Algorithmic automated optimization of design parameters significantly enhances design efficiency and precision. AI-driven design methodologies can generate personalized forms based on user needs and market trends, satisfying customized product demands. However, practical applications face multiple challenges: dataset quality defects, such as noise and incompleteness, compromise model reliability; the “black-box” nature of ML models lacks decision transparency, posing challenges for safety-critical designs; high-dimensional design spaces and complex algorithmic models increase computational burdens and resource requirements.

Technology empowerment and cross-domain innovation

AI and ML demonstrate vast potential in the field of form generation (Ren & Xiong, 2022; Xiong, Yue & Wu, 2023) driving synergistic advancements in automated intelligent design, personalized customization, performance prediction and optimization, sustainable design, and additive manufacturing. In automated design, hybrid models integrating expert knowledge graphs with generative adversarial networks (GANs) need to be constructed to simulate the creative thinking of human designers. Combining digital twin technology with additive manufacturing establishes a real-time feedback loop for “design-validation-production.” For performance prediction and optimization, multiphysics coupled simulation models should be developed to analyze the mapping relationships between product form parameters and material properties, with reinforcement learning enabling dynamic optimization. In the dimension of sustainable design, a life-cycle carbon footprint assessment model must be built (Zhao, Tang & Gong, 2024). Through topological optimization and materials genome technology, redundant structures and material consumption can be reduced, promoting green design and circular manufacturing.

Ecosystem co-construction and ethical governance

While advancing algorithm-manufacturing collaborative design, ethical issues raised by AI must be considered. First, vigilance is required against bias amplification effects in data-driven design—if training data contains cultural, gender, or economic biases, algorithms may generate solutions that reinforce inequalities in social resources. Second, the a mbiguous intellectual property ownership of AI-generated designs necessitates establishing rights allocation mechanisms to balance the interests of algorithm developers, data providers, and users. Furthermore, the “black-box” issue in the design process demands enhanced algorithmic transparency, especially in critical product areas such as medical implants, where decision interpretability and traceability must be ensured. Additionally, from an environmental perspective, excessive design tendencies must be evaluated to avoid material waste and electronic waste issues. Finally, a multi-stakeholder governance framework involving ethicists, designers, and engineers, along with relevant policies, needs to be constructed. Through technical standards and certification systems, technology can be guided toward responsible innovation.