2.1. PRISMA Flow

We applied the PRISMA methodology to track the search criteria and obtain articles for BA and SLR in a clear, transparent, and detailed manner following ¹⁴.

2.2. Inclusion/Exclusion Criteria

The flowchart in Fig. 2 and the inclusion and exclusion criteria listed in Table 1 were used to screen the retrieved records. For BA, only English-language documents published within the last 10 years were included, resulting in 1,816 records. For the SLR, we selected articles published within the last five years in the subject areas of computer science, engineering, mathematics, and social sciences following the recommendations from the BA results. The filtering criteria that are illustrated in Table 1 were then applied. A total of 121 articles were identified. After a manual review, we identified 51 relevant articles for our study. For more details, please refer to Section 3.

To ensure methodological rigor and reproducibility, explicit inclusion and exclusion criteria were defined and applied consistently across all screening stages.

Inclusion criteria: Studies explicitly addressing the integration of AI and MDD; research focusing on adaptive, intelligent, or context-aware UIs; peer-reviewed journal articles; studies published between 2017 and 2024 (for SLR); and articles written in English.

Exclusion criteria: Studies focusing solely on AI or MDD without a combination; papers unrelated to UIs or human-computer interaction; short papers, editorials, or non-peer-reviewed content; duplicate or inaccessible full-text articles; and studies lacking sufficient methodological or experimental detail.

These criteria were iteratively refined during the screening process to ensure consistency and relevance.

The filtering criteria summarized in Table 1 were designed to ensure both the relevance and methodological rigor of the selected corpus. In particular, different criteria were applied at the BA and SLR stages to balance breadth and depth of the analysis. The BA phase employed a broader time window (last 10 years) to capture the evolution of the research field, identify publication trends, and detect dominant themes and subject areas. In contrast, the SLR phase focused on a more recent subset (the last five years) to analyze state-of-the-art approaches and current research directions in greater detail. The restriction on journal articles and peer-reviewed publications was intended to ensure scientific quality and reliability while limiting the corpus to open-access sources to improve the transparency and reproducibility of the review process. Furthermore, the selection of subject areas was guided by the BA results, enabling the SLR phase to focus on the most relevant and representative disciplines rather than relying on predefined or arbitrary domain boundaries. Overall, these criteria collectively support a two-stage review strategy in which BA provides a global mapping of the research landscape, and the subsequent SLR enables a focused and in-depth examination of methodologies, applications, and research gaps.

2.3. Content Analysis and Thematic Synthesis

Beyond the bibliometric filtering process, the final set of eligible studies was subjected to a structured qualitative content analysis to ensure that the findings of this survey were grounded in the substantive contributions of prior work rather than in bibliographic metadata alone. Although BA was used to identify publication trends, dominant keywords, and disciplinary distributions, it did not provide sufficient depth for extracting methodological insights, research gaps, or future directions. The full texts of the selected articles were systematically examined following the PRISMA-based selection process. Each study was analyzed using a predefined coding framework designed to capture both the technical and conceptual dimensions of the research. Specifically, the analysis considered (i) the type of UI or interaction environment (e.g., AR, VR, VHs, or metaverse applications); (ii) the interaction paradigm and modalities employed; (iii) the context-awareness mechanisms and contextual data sources; (iv) AI techniques applied (e.g., ML, DL, and reinforcement learning); (v) MDD methodology or modeling strategy adopted (e.g., PIM–PSM transformations, domain-specific modeling, and runtime adaptation); (vi) the application domain; (vii) the reported limitations and challenges; and (viii) the future research directions identified by the authors. Accordingly, the methodological workflow of this study combines the strengths of the BA and SLR: the former provides a high-level mapping of the research landscape, whereas the latter enables an in-depth interpretation of methodologies, applications, and unresolved challenges in the integration of AI and MDD for adaptive UI generation.

The outcomes of this content analysis are reflected in the thematic organization and synthesis presented in Sections 3 and 4.

2.4. Screening and Selection Process

The screening process was conducted in multiple stages following PRISMA guidelines. First, title and abstract screening were performed to eliminate irrelevant studies. Subsequently, full-text screening was conducted to assess the eligibility based on the defined inclusion and exclusion criteria. To improve the reliability, the screening process was independently performed by two reviewers. Disagreements were resolved through discussion, and when necessary, a third opinion was considered. This process ensured consistency and reduced subjective bias during study selection. The transition from 121 to 51 studies resulted from the application of stricter relevance criteria during the full-text analysis, particularly focusing on studies that explicitly combined AI and MDD for adaptive UI generation. Studies that partially addressed only one of these aspects were also excluded.

The reduction from 3,335 initial records to 51 final studies reflects a progressive filtering strategy aimed at ensuring both relevance and quality. While early filtering stages relied on automated criteria (e.g., year, subject area, document type), the final reduction was driven by conceptual relevance to the AI_MDD integration problem, ensuring that only studies directly addressing adaptive interface generation were retained.

3.1. Publication Trends Over the Past 10 Years

Referring to RQ1 for BA, to address this, an analysis of the documents was performed; Fig. 3 illustrates that 1,648 out of 1,816 publications were published in the last five years, with a publication ratio of approximately 91% of the total publications. This indicates that the authors and researchers have shown increased interest in AI_MDD approaches and that these technologies are increasingly applied across various fields. This result suggests focusing on studies from the last five years, as this represents the largest proportion of published research. This result was adopted as a criterion for filtering articles in the Scopus database for the SLR. In addition to reporting publication counts, this indicator is important because it reflects the historical acceleration of research at the intersection of AI, MDD, and UIs. The concentration of publications in recent years suggests that this is not a mature or stable field but an emerging and rapidly expanding field. This historical pattern also justifies the decision to focus on the SLR phase in recent studies because they are more likely to capture current methodological developments than older, more generic MDD or HCI studies.

3.2. Primary Research Keywords

Referring to RQ2 for BA, co-occurrence was selected as the analysis type, and “Author’s keywords” was designated as the unit for the bibliometric study of the most popular keywords. In this regard, Fig. 4 and Table 2 present the most frequent keywords found in the dataset, totaling 682. As shown in Table 2, the keywords utilized in the research were as follows: AI 957, ML 525, human 147, and human-computer interface 206. Further, other prolific keywords in the research topic field are presented in Table 2. In addition, Table 2 also shows the total link strength (TLS) for each keyword in the field of research topic. The results regarding keywords specific to a research topic can be used as criteria when conducting research in this field. The use of keyword co-occurrence was motivated by the need to identify how the field is conceptually organized rather than simply the most frequently appearing terms. This analysis reveals the thematic evolution of the literature and shows that the field is shaped not only by technical AI terms but also by concepts related to human-computer interaction, user experience, and adaptation. In contrast to previous survey studies that have often examined MDD or intelligent interfaces separately, the present study used a keyword structure to demonstrate the hybrid and convergent nature of the research landscape.

3.3. Most Common Subject Areas

Referring to RQ3 for BA, according to the Scopus database analyzer, an analysis was conducted on the publication subject areas of the articles over the previous 10 years; the first subject area is “computer science,” which accounts for approximately 39%. This is followed by engineering, mathematics, and social sciences, as illustrated in Fig. 4. This result was selected as a criterion for filtering articles in the Scopus database for the SLR. Subject area distribution was selected as a bibliometric indicator because it helps to explain the disciplinary breadth of the field and its historical positioning across different research communities. The prominence of computer science, engineering, mathematics, and social sciences confirms that AI_MDD for adaptive UIs cannot be understood as a purely technical problem. Rather, it is an interdisciplinary research area that combines software engineering, intelligent systems, and human-centered design. This supports one of the central premises of the present survey, namely that a meaningful review must bridge these domains rather than analyze them in isolation.

3.4. Findings from BA

After showing the results of the most popular and reliable analysis applications (Scopus analyzer and VOSViewer) conducted on the research topic, these results were discussed and summarized as follows. Publication trends: Growth in publications over the last five years, with 91% of the research output occurring in the last 10 years. Primary keywords: “AI,” “ML,” “human,” and “human computer interface” dominate the field. Subject areas: this research is interdisciplinary and involves computer science, engineering, mathematics, and social sciences. In the next section, we use these results as the criteria to conduct a more focused and accurate SLR. These criteria were used to filter 1,816 records following the steps shown in Fig. 2 and Table 1. The bibliometric findings also help to clarify how this study differs from previous survey studies. Earlier reviews often concentrated on a single research stream, such as MDE, adaptive UIs, intelligent tutoring systems, or XR-based interaction technologies. By contrast, the present study used BA to map a broader research landscape before conducting a systematic review of the intersection between these streams. In this sense, the contribution of this review lies not only in summarizing prior work but also in providing an integrated perspective on how AI methods, MDD methodologies, and emerging interaction environments are evolving together as part of a unified research direction.

4.1. Types of Interfaces, Their Uses and Interactions

Referring to RQ4 (of SLR), the types of interfaces, their uses, and interactions are described as follows.

4.1.3. Methodological Evolution and Future Research Directions

The analysis included 51 studies. However, several of these are systematic reviews and mapping studies that collectively cover more than 300 primary studies.

To move beyond descriptive coverage of prior studies, we synthesized the uploaded literature from both quantitative and methodological perspective. As shown in Fig. 5, the evidence base was substantial, spanning major review studies with 55, 46, 63, 85, and 93 analyzed primary studies, respectively. This confirms that the field is no longer defined by isolated efforts but by a growing and methodologically diverse body of work ⁵⁴,⁵⁵,⁵⁶,⁵⁷. Fig. 6 further shows that the empirical evaluations vary considerably in scale, ranging from exploratory usability studies with a few dozen participants to large-scale questionnaire-based studies with more than 1,000 responses. This variation reflects the transition from proof-of-concept approaches toward more mature evaluation practices ³⁹,⁵⁸,⁵⁹,⁶⁰,⁶¹ . Most importantly, Fig. 7 shows a clear methodological profile of MDE4ML research, which is still largely academic (89%), strongly centered on PIM-level abstraction (93%), heavily dependent on model-to-text transformations (91%), and predominantly fully automated (85%), whereas tool availability remains limited (50%) ⁶². Taken together, these findings suggest that the field has evolved from classic transformation-driven MDD to hybrid AI_MDD approaches. However, industrial maturity, tooling, and scalable deployment remain important future research directions.

4.2. Approaches and Synergies

Referring to RQ6 (of SLR), AI and MDD are two paradigms that shape the way in which software is designed, developed, and deployed. While AI employs data-driven approaches to enable intelligent decision-making and automation, MDD focuses on abstracting the system design by relying on high-level models. Table 8 presents the MDD approaches in distinct methodologies, highlighting their key characteristics and application domains. Classic MDD using UML/SysML focuses on code generation from platform-independent models for UI and ML component design ⁶²,⁵⁰, whereas agile MDD incorporates iterative refinement with large language models (LLMs) assistance for rapid prototyping ⁶³,⁶⁴. Domain-specific modeling (DSM) employs custom languages tailored for AR/VR and internet of things (IoT) systems, enabling specialized applications such as educational AR tools and industrial IoT solutions ⁶⁵,⁶⁶. AI-enhanced MDD leverages LLMs and ML for automated model-to-code transformations, particularly in generating ML pipelines ⁶²,⁶⁷. Self-adaptive MDD supports runtime model updates for context-aware UIs and adaptive systems ⁶⁸,⁶⁹. Multi-experience MDD enables cross-platform development of AR/VR and web interfaces from a single model ⁷⁰.

Table 8 shows the trend toward integrating AI, particularly LLMs, to automate and enhance MDD processes, with domain-specific approaches dominating AR/VR and IoT applications. However, gaps remain in standardizing MDD for generative AI and addressing the ethical implications, underscoring the need for further research in these areas. AI approaches can be broadly classified into several categories based on their application areas, methodologies, and underlying principles. Table 9 organizes these approaches into key methodologies and their respective application domains and highlights several dominant trends. Supervised and unsupervised learning techniques are commonly used for predictive modeling and pattern detection, particularly in educational and institutional adoption studies ⁷¹,⁷²,⁷³. DL drives applications in computer vision and natural language processing (NLP), such as AR scene reconstruction and interactions with VHs ⁴⁷,³⁵. Generative AI, particularly through LLMs and generative adversarial networks (GANs), has become a transformative approach to dynamic content generation and AI-assisted development, particularly in adaptive AR and model-driven systems ³⁸,⁶³,³³,⁴⁷. Reinforcement learning facilitates the creation of personalized learning pathways ⁷⁴,⁴⁸, whereas model-driven AI automates code generation for IoT and ML systems ⁶²,⁶⁶. Explainable AI (XAI) focuses on enhancing transparency in healthcare and ensuring ethical design practices ⁴⁹,⁷⁵, whereas edge AI supports real-time processing of AR and VR applications ⁴⁸,¹⁷. Computer vision enhances spatially aware AR experiences and facilitates the creation of lifelike avatars ³⁰,³³, whereas conversational AI enables responsive, human-like interactions in the virtual tutoring and service domains ³⁵,³⁴. Table 9 highlights the prevalence of generative learning and DL in recent research, as well as synergies such as LLMs enhancing MDD, but notes gaps in privacy-preserving federated learning and neurosymbolic AI integration. These approaches collectively demonstrate the role of AI in creating adaptive, intelligent systems across education, healthcare, AR/VR, and industrial applications, with methodological choices heavily influenced by domain-specific requirements for real-time processing, personalization, and transparency. The combination of MDD and AI is a promising advancement in software engineering. MDD emphasizes abstraction and automation through models, thereby facilitating more efficient software development. In contrast, AI provides intelligent capabilities that enhance decision-making and adaptability within software systems. The integration of these approaches could lead to more efficient, adaptive, and intelligent software solutions ⁶⁶.

Furthermore, the article “Using AI and MDD to Generate Intelligent and Adaptive UIs” explores how MDD and AI complement each other, highlighting their potential to transform software development by improving productivity and adaptability.

Table 10 outlines the six core methodologies for integrating AI with MDD to develop an adaptive and intelligent UI. Each method has its distinct advantages:

1. 1.

   AI-assisted code generation: This approach leverages LLMs to automate the translation of high-level models (e.g., UML) into executable code ⁶³,⁶². Although it significantly reduces manual effort, it can result in a code of variable quality and limited maintainability.

2. 2.

   Model@runtime adaptation: This technique uses AI to dynamically adjust UI models in real-time based on contextual data, such as user behavior or environmental sensors. This enables the creation of personalized interfaces, although it may incur computational latency and energy inefficiency ⁶⁹,⁶⁸.

3. 3.

   DSM enhanced with AI: This approach tailors solutions for niche applications such as AR in education or the industrial IoT by combining custom modeling languages with ML components. However, they encounter challenges related to scalability and interoperability across various domains ⁶⁵,⁶⁶.

4. 4.

   Self-learning UIs: These interfaces use reinforcement learning to iteratively optimize the UI layouts based on user interactions. Although effective, they require extensive training data and may exhibit slow convergence ⁷⁴,⁴⁸.

5. 5.

   Generative UI design: This methodology utilizes generative AI techniques (e.g., GANs and LLMs) to create UI prototypes based on abstract requirements, enabling rapid design. However, this introduces the risk of biased or unpredictable outputs ³⁸,³³.

6. 6.

   XAI and MDD combination: This approach embeds interpretability techniques, such as decision logs, into MDD workflows. This combination builds trust and aids in debugging, although it adds complexity to the process ⁴⁹,⁷⁵.

Emerging Architectural Patterns

Beyond the individual techniques, the reviewed literature reveals several recurring architectural patterns. The first pattern is a model-to-code pipeline in which platform-independent models are transformed into executable interface implementations. The second pattern is model@runtime adaptation, in which the interface models are continuously updated during execution in response to user and contextual signals. The third pattern combines domain-specific models with embedded AI components, particularly in AR/VR and IoT-oriented systems. More recent studies have pointed toward a fourth pattern, namely generative AI-assisted interface engineering, in which LLMs or other generative methods participate in model refinement, transformation, or interface synthesis. These recurring patterns provide a more integrative understanding of the field than an isolated categorization of techniques.

Comparative Analysis of AI_MDD Approaches

A comparative reading of the reviewed studies suggests that AI_MDD approaches differ not only in methodological orientation, but also in their practical effectiveness for adaptive interface generation. Classic MDD approaches remain valuable for structural consistency and code generation but provide limited support for runtime adaptation. DSM is particularly effective in AR/VR contexts because it enables tailored abstractions for specialized interaction environments; however, its applicability is often constrained by interoperability limitations. In contrast, self-adaptive and model@runtime approaches offer stronger support for personalization and context-aware behavior; however, they face challenges related to latency, computational cost, and system complexity. Generative AI-based methods accelerate prototyping and interface production; however, they often raise concerns regarding maintainability, predictability, and transparency. Taken together, these findings indicate that no single AI_MDD approach dominates all scenarios; instead, effectiveness depends on the balance between structural rigor, adaptive intelligence, and platform heterogeneity.

Table 11 provides a comparative overview of the key AI_MDD approaches and highlights their strengths, limitations, and suitable application environments. Classical MDD approaches remain effective in ensuring structural rigor and portability, particularly in the traditional web and UI engineering contexts; however, they offer limited support for runtime adaptation. In contrast, domain-specific modeling combined with AI enables tailored solutions for specialized domains, such as AR/VR and IoT. However, this advantage comes at the cost of reduced interoperability across platforms. Model@runtime approaches introduce dynamic personalization capabilities by adapting interfaces based on contextual information. However, they often face challenges related to latency and system complexity. Recent approaches that integrate generative AI within MDD pipelines facilitate rapid prototyping and design exploration. However, they raise concerns regarding the predictability and explainability of the generated outputs. Finally, multi-experience MDD approaches aim to ensure cross-platform consistency across heterogeneous environments, including the web, AR, VR, and metaverse systems, although their adoption is still constrained by the limited maturity of the supporting tools. Overall, the comparison suggests that no single approach fully satisfies the requirements for adaptive UI generation, thereby reinforcing the need for hybrid solutions that balance structural control, adaptability, and scalability.

The reviewed studies indicate that AI and MDD are not independent components but complementary elements in the generation of adaptive interfaces. Although the MDD provides a structural backbone for modeling and transforming UIs, AI methods introduce intelligence through learning, prediction, and runtime adaptation. This combination is particularly important in complex interaction environments, where interface requirements are dynamic, context-dependent, and heterogeneous.

4.3. Gaps and Challenges

To better contextualize the identified challenges, it is important to clarify how emerging interaction technologies (AR, VR, VH, and MV) relate to MDD. In most existing literature, these technologies are primarily discussed as application domains. However, from a model-driven perspective, these can be understood more accurately as environments that impose specific constraints on the representation, transformation, and execution. For instance, AR systems require interface elements to be modeled in relation to the physical space and sensor data, whereas VR environments introduce immersive interaction flows that depend on user movement and multimodal inputs. Similarly, VHs involve conversational and behavioral modeling, and metaverse platforms emphasize interoperability and persistent multi-user interaction. These characteristics extend beyond the assumptions of traditional MDD, which have largely focused on screen-based and relatively static interfaces. This perspective suggests that interaction technologies are not external to MDD but instead redefine their requirements. They highlighted the need for richer modeling abstractions, more flexible transformation mechanisms, and stronger support for runtime adaptation within model-driven frameworks. While the relationship between interaction technologies and MDD highlights evolving modeling requirements, the role of AI introduces an additional dimension that enables these requirements to be operationalized. In recent studies, AI techniques (particularly ML and DL) have been increasingly used to support dynamic adaptation, user modeling, and decision-making within interface generation processes. In this context, MDD provides a structural foundation through model abstraction and transformation, whereas AI contributes to the capability of learning from user behavior and the environmental context. Interaction technologies define the conditions under which these mechanisms must operate, often requiring real-time responsiveness, multimodal interaction, and cross-platform consistency. Rather than functioning independently, these three dimensions form an interdependent system in which modeling, adaptation, and interaction are tightly coupled. Despite this convergence, current approaches often integrate AI into MDD pipelines in a limited or ad hoc manner, without a clear alignment between learned models and model-driven representations. This lack of combination restricts the scalability and generalizability of the adaptive interface solutions. Therefore, a more cohesive combination of AI and MDD is required, particularly in complex environments. Future research is likely to focus on frameworks in which models are continuously updated through AI-driven insights, enabling runtime adaptation and supporting multi-experience interfaces across heterogeneous platforms. Such an approach would allow MDD to evolve from a static design paradigm into a dynamic, learning-enabled system capable of responding to diverse interaction contexts.

Referring to RQ7 (of SLR), while combining AI with MDD to generate an adaptive UI offers benefits, gaps and challenges exist in its implementation and use. Table 12 systematically organizes the key gaps in combining AI with MDD to generate an adaptive UI into six critical categories, each with its distinct manifestations and underlying causes. Technical limitations (TL) stem from immature AI code generation tools that produce poorly structured codes and edge computing constraints that introduce latency in dynamic UI adaptation, fundamentally restricting real-time responsiveness ⁶³,⁶⁹. Explainability challenges (EC) arise from insufficient combinations of XAI techniques with MDD workflows, resulting in opaque decision-making that erodes user trust and complicates the debugging of complex model transformations ⁴⁹,⁷⁵.

Data dependencies (DD) reveal how biased training datasets perpetuate discrimination in UI outputs, while reinforcement learning approaches demand impractical volumes of user interactions owing to cold-start problems and inadequate data diversity ⁷⁴,⁴⁸. Tool chain gaps (TG) highlight the absence of unified development platforms caused by proprietary formats and lack of standardization, particularly hindering AI integration with domain-specific modeling tools ⁶⁶,⁶⁵. Ethical risks (ER) expose vulnerabilities in current AI_MDD pipelines, where missing governance frameworks allow for accessibility violations and intellectual property ambiguities in AI-generated UI components ⁷⁶,³⁸. Finally, human factors (HF) demonstrates how excessive automation diminishes designer agency and how complex hybrid workflows create adoption barriers, both rooted in inadequate human-in-the-loop mechanisms ⁶²,⁵⁰. These challenges reveal a multidimensional problem space in which technical immaturity intersects methodological gaps, ethical blind spots, and human-centered design shortcomings. This situation demands coordinated advances in tools, standards, and frameworks to realize the full potential of AI_MDD synergies for next-generation UIs. There exists critical unresolved issues, such as real-time adaptation; most AI_MDD systems struggle with sub-second UI updates due to model inference delays ⁶⁹,⁶⁸, cross-platform generalization (interoperability); generated UIs often fail to port seamlessly across AR/VR/web ⁷⁰, and energy efficiency; no studies address the high computation costs of continuous AI_MDD co-processing ⁴⁸. Although the identified challenges are categorized into distinct groups, they are inherently interconnected and often reinforce each other. For instance, DD, such as limited or biased datasets, directly affect the performance of AI-driven adaptation, which, in turn exacerbates TL, particularly in terms of real-time responsiveness. Similarly, the lack of a standardized TG hinders the integration of explainable AI mechanisms, thereby intensifying EC and reducing system transparency. Further, ER and HF issues are not isolated concerns but are closely linked to the level of automation introduced by AI_MDD systems. Increased reliance on automated decision-making without sufficient human-in-the-loop mechanisms can reduce user trust and limit designer control. These interdependencies suggest that addressing individual challenges in isolation is insufficient, and that a holistic perspective is required to effectively advance AI_MDD-based adaptive UI generation.