Synthetic Data in a Data-Intensive World: A Literature Review to Address an Application-Driven Analysis

1. Introduction

In a context of increasing digitalisation and of a consequent data proliferation, synthetic data [1] is increasing in popularity in different disciplines and applications. This is only an apparent contradiction, which actually reflects established technological trends relying on huge amounts of data. Indeed, synthetic data are often adopted for different reasons.

Primarily, synthetic data may play an effective protective role [2] in presence of privacy issues as a response to actual societal needs. For instance, re-identification of data is not that uncommon, as in a recent case in South Korea [3]. Some disclosure strategies are based on multiple synthetic datasets [4]. In addition, synthetic data are extensively adopted in presence of restrictions to access real data, to deal with a lack of specific data (e.g. simulation of rare events), as well as to establish data augmentation and hybrid solutions [5].

Multiple synthetic data generators have been developed in the past two decades, aiming at the translation and operationalization of the "look like real" concept [6]. In the last years, the different techniques have often been object of study and comparative analysis (e.g. [7,8]). Such consolidated methods are still evolving to scale [9] and, more recently, have been further enhanced by LLMs [10] and GenAI [11].

Despite their relative popularity and largely discussed potentiality, the utility of synthetic data is still a subject of research. The inherent complexity of requirements needs to be approached at different levels, including general and specific utility [12].

This work explicitly addresses an application perspective of synthetic data by providing an inductive overview of the state of the art from the different application domains. Indeed, by selecting recent research papers with an application focus, this review provides a specific view, which is complementary, rather than overlapping, to the more consolidated analysis of techniques/methods to produce syntetic data. While holistically addressing issues and solutions in a cross-domain context, the fine grained analysis within the specific application domains contributes to consider specific characteristics and peculiarities in context.

• Structure of the paper

The paper is structured in a relatively traditional way, with methodological aspects preliminarily discussed (Section 2), followed by the core part of the paper. The latter is composed of 3 different sections, dealing respectively with a holistic analysis (Section 3), an AI-driven topic modeling (Section 4), and a discussion part (Section 5), including implications at different levels, research gaps and current limitations. Finally, the selected papers arranged by category are reported in the appendix.

2. Methodology and Approach

Synthetic data and their application have been object of significant research in the last few years. However, a fine-grained cross-domain identification of the different contributions may be challenging, given the variety of situations, domain peculiarities and the fragmented nature of pertinent literature.

In order to maximize the value of this review, we have adopted hybrid methods for the retrieval/selection as well as for the analysis. Details are addressed in the following subsections.

2.1. Paper Selection

The extent and intent of this research have suggested the combination of a systematic approach adopting a more traditional keyword-based search with snowballing from retrieved papers (Figure 1).

In such a context, selection criteria are a determinant for a proper and effective operationalization of the aimed application perspective.

• Retrieval process

The retrieval process is structured in three sequential phases (Figure 1). The main process assumes a keyword-based basic search, where a given keyword $i\in RootKeywords$. The latter set is understood as a very generic collection of terms that aims at (i) retrieving a first corpus of relevant papers in scope, and (ii) identifying the main application domains and their specific keywords. Our heuristic experimentation has proven an approach based on a very reduced set of root keywords ($RootKeywords=\left\{SynteticData\right\}$) to be more effective than extended ones (e.g. including keywords like "application").

The second phase aims at a consolidation in the different domains by performing an additional search, where keywords include domain specific generic terminology. More formally, a domain j is approached by querying on heterogeneous keywords, including also domain-specific terms as in eq. 1, while domain keywords are reported in Table 1.

$$(i,k),wherei\in RootKeywords,k\in {DomainKeywords}_j$$

(1)

As mentioned previously and represented in Figure 1, a last consolidation step is a snowball search, which relies on the following citation trails.

• Selection Criteria

A second methodological pillar for effectively identifying a body of knowledge is the definition of selection criteria which consistently apply throughout the process (Figure 1).

In this specific case, selection criteria are crucial in constructing an application perspective for the target review. A retrieved paper is finally selected if it falls at least into one of the following categories:

• Discussion or review of specific applications of synthetic data.
• Synthetic data (including also hybrid) solution for a specific domain or application.
• Description of a development or of an empirical work in the context of a specified application domain.
• Any other application-related contribution, for instance papers discussing issues, challenges or potentialities related to the application of synthetic data.

The mentioned criteria clearly exclude generic contributions on synthetic data not explicitly related to applications as they do not contribute to an application-driven analysis, which is the main object of this paper.

2.2. Analysis Framework

Also the conducted analysis relies on a hybrid approach, which integrates a human-performed critical analysis with an AI-driven thematic structure. While intrinsically independent in concept, the two methods naturally converge to provide a consolidated synthesis from a significant and heterogeneous corpus of papers.

• Classification and Characterization

The primary analysis dimension considered in this work is the application domain. An overview of the conducted analysis is proposed in Section 3.

Seven main domains have been identified as a result of an iterative human-performed analysis of the retrieved papers: (i) Sciences, (ii) Healthcare (including also Health Economics), (iii) Economics, (iv) Engineering and Architecture, (v) Energy, (vi) Cybersecurity, and (vii) Technology (Computer Vision/Machine Learning/AI). In the way it is defined, the latter category is probably not a formal application domain. However, it reflects the contribution of synthetic data to the development of the technology itself, as well as it contributes to better understand relevant trends in computation, where a significant portion of research is conducted in a relatively generic context with potential applications in multiple domains.

Regardless of the associated domain, for each selected paper a type and a number of characterizing keywords are specified. The paper type is defined according to a simplified binary classification, where conceptual papers focus on review or discussion, while application papers present a specific solution or more empirical works beyond conceptualization. A very limited set of keywords (between 1 and 3) pretend to semantically characterize a given paper in a generic, yet specific, way. In order to prioritize conciseness and effective conceptualization, such labeling is human performed.

• AI-driven Topic Modeling

Topic modeling is a popular and sophisticated unsupervised machine learning technique that enables efficient exploration of unstructured data, contextual trend discovery and thematic structure modeling.

This study adopted BERTopic [13], with the title and abstract of the selected articles as input. The results are presented and discussed in Section 4. Although topic modeling is rapidly emerging as a primary analysis technique given its proven ability to provide insight, in this specific context its role is mainly integrative to further consolidate the analysis and mitigate bias towards a more exhaustive and consistent overview of the existing body of knowledge.

3. Application of Synthetic Data

Overall, 255 papers have been included in this review. In addition to the expected focus on the development of the technology itself ($\approx 28\%$ of the selected articles), healthcare is the most represented domain ($\approx 21\%$), followed by three emerging domains (Engineering, Economics and Energy), which contributes around 11-12% each; finally, applications in sciences is the topic of about the 9% of the considered papers, while the remaining part ($\approx 6\%$) is related to cybersecurity (clearly emergent domain).

A very first analysis can be deducted from a simple longitudinal analysis of such a distribution (Figure 2a). With some approximation, very recent contributions (2024-2026) leverage on the huge (and still partially unexplored) capability of LLMs, while a previous time segment (2021-2023) seems more related to the last generation of Machine Learning techniques; older contributions, ideally published before 2020, reflect a more traditional approach to synthetic data with a significant component in the broad context of Computer Vision or of consolidated applications, primarily in Healthcare and Sciences.

Such a simple overview clearly demonstrates a strong and increasing interest in synthetic data across all domains, mostly powered by the continuously growing AI capabilities. The relationship between synthetic data and AI should be considered according to a double perspective. Indeed, on one side, synthetic data are actively and consistently contributing to model training and, therefore, to the development and consolidation of AI technology; at the same time, AI technology is providing enhanced and unprecedented capabilities for synthetic data generation, enabling de facto new opportunities through a novel generation of techniques.

Last but not least, always as a response to recent technological advances in AI, an effective application of synthetic data is rising also in domains where the application of synthetic data was definitely less usual, such as economics, engineering and energy.

Looking at the type of the selected papers (Figure 2b), healthcare, a domain with a well-known historical interest in synthetic data, presents a fundamental balance between conceptual and empirical/methodological works. With a minor incidence, conceptual works characterize also economics, cybersecurity and sciences (38%, 31% and 23% of the papers respectively), while they are less popular in the other domains (in a range 6-11%).

3.1. Science

Because of the very broad scope of this category, which can potentially include all fields of science in their diversity, we have not aimed at exhaustiveness or fine-grained classification but, rather, at the identification of main trends and their recent evolution. The list of selected papers is reported in Table A1.

A simplified interpretation of the conducted analysis has identified a predominant corpus of papers directly related to Social Sciences, as well as evidence of application in other field of science.

• Social Sciences and Humanities. On one hand, this review confirms the historical interest of social sciences in synthetic data, especially looking at human behavior and related fields. That is mostly due to privacy concerns, which naturally establish critical trade-offs and a consequent scarcity of data. Such an interest is further highlighted by more recent contributions in the field (e.g. [14,15]), which leverage on LLMs and enhanced semantic capabilities.
  Another clear pattern in the specific field of Social Sciences points out a broader use of synthetic data, above all in the broad context of politics (e.g. [16,17,18,19]) and urban data (e.g. [20]). Once again, LLMs is enabling unprecedented capabilities and opportunities in a privacy sensitive space.
• Other fields of Science. Looking more holistically at the use of synthetic data in Sciences, the mainstream perception points out an increasing general interest amid opportunities and challenges [21].
  There is currently evidence of application in multiple fields beyond the social sciences, for instance in physics [22], chemistry [23,24], and astronomy [25]. Additionally, a significant amount of work in terms of mathematics/statistics is directly related to data science and its applications, mostly AI technology. The latter is partially reported in Table A7 as part of applications in technology development.

3.2. Healthcare

Healthcare is an intrinsically privacy-sensitive domain, characterized by a huge demand for data in contrast to accessibility constraints. In principle, healthcare presents an inherent complexity, with a core in medical aspects, health data and domain specific technology, in addition to the intersections with other dimensions, including social aspects and health economics. According to a previous analysis (Figure 2a), in general terms the domain can be considered somehow mature from a perspective of synthetic data adoption [26,27], with a further huge potential, mainly generated by recent advances in AI technology [18,28,29,30,31].

From a conceptual perspective, as probably expected, the review has confirmed a strong focus on privacy matters [32,33], legal/ethical implications [34], evaluation/validation [35,36,37,38], and governance [39].

The papers in the domain included in this review are reported in Table A2. From a strictly application-focused perspective, they may be classified as follows:

• Health Data. It includes clinical data [35,38], Health Record [40,41,42], data for education and training [43], behavioral data [18], profiling [34], as well as more specific aspects, including data sharing [44], metadata [45], bias [46], fairness [47], utility [48], access and augmentation [49], and Open Data [50].
• Medicine. Applications directly or indirectly related to medicine [51] are the most significant in quantitative terms. They range in a broad spectrum, including, among others, applications in Machine Learning [52] and Advanced Analytics [53], pharmaceutics [48], cancer research and clinical trials [54], drug development [55], sport medicine [56], hematology [57], early diagnosis and classification [58,59], and epidemiology [60].
• Biomedical Research. It may be considered as a well-identified sub-field of the previous one, with a specific approach [61,62,63].
• Health Economics. They may be critical, if not central, when looking at healthcare as a whole or from a managerial or economic perspective. There is evidence of direct [64,65] [18] and indirect (e.g. services [66]) application in the field.
• Specialized technology. Technology plays a primary role in modern healthcare. Such a relevance is reflected also in the adoption of synthetic data, for instance in the broad context of health informatics [67] and software [68], as well as in technology for medical imaging [59,63,69,70,71]. Synthetic data is also seen as an enabler for Machine Learning applications in medicine [72].

3.3. Economics and Business

The contributions related to the broad fields of economics are reported in Table A3. In quantitative terms, the review shows a significant amount of very recent work (Figure 2a) to confirm the rising interest for synthetic data in the domain. More qualitatively, it is possible to distinguish few main patterns with a certain degree of overlapping as follows:

• Opportunities, Challenges and Expectations. There is an extensive ongoing critical analysis of opportunities and challenges [73,74,75], resulting in huge expectations [76,77] amid significant concerns in terms of privacy and security [78,79,80,81,82].
• Financial Data. A consistent, yet diverse, corpus of works refers directly to financial data in a privacy-sensitive context. Concrete applications in addition to the previously mentioned are related to different aspects, including fraud [83,84,85,86], financial models (e.g. sequential [87] and decentralized [88]), Federated Learning [89], Governance [39], Risk Modeling [90], Predictive Models [91], Banking [92], and Trading [93,94].
• Other Contributions. The application of synthetic data in the broad domain is beyond Finance. The review includes for instance application in Industry 4.0 [95], Business Data [96] and Intelligence [97], AI governance in enterprise [98], commercial delivery [99], and agriculture economics [100].

3.4. Engineering and Architecture

Synthetic data are also emerging in the broad field of engineering (Table A4). Our analysis has pointed out the following main trends/patterns:

• Classic fields of Engineering. It includes a prevalence of applications in civil/mechanical engineering [101,102,103,104,105,106,107,108,109,110], as well as contributions in other fields, such as environmental [59,111], geo-technical [112,113] and electronic [114] engineering.
• Hybrid solutions. Applications in the domain often present an explicit focus on hybrid solutions [101,103,105,115,116,117].
• Modern Industrial Engineering. Other applications are related to the modern Industrial Engineering, for instance Advanced Manufacturing [118,119], autonomous construction systems [120], Industrial Plants [121], Industry 4.0 [95] and 5.0 [122].
• Architecture. While not numerically especially relevant, there is evidence of interesting applications in Architecture [123,124,125].
• BIM. Synthetic data in Building Information Modeling (BIM) [126] definitely deserves a mention, considering existing works [107,116,127,128] and, more holistically, the great potential of such a standard across different application contexts (e.g. sustainable design [129]).

3.5. Energy

Contributions related to energy systems are reported in Table A5 and evidently show an emerging interest in synthetic data. The retrieval process, which assumes just very generic keywords as an input, has allowed a domain breakdown as follows:

• Energy Data / Management. A significant cluster of contributions is associated with energy data and management. It is the case, for example, of demand forecasting[130,131] and consumer behavior [132,133,134].
• Smart Energy Systems. A similar sub-domain has a main focus on smart/intelligent systems. It includes Smart Grids [135,136,137], Smart Home [138], as well as more specific advanced functionalities, such as fault detection [139,140,141,142].
• Green Technology. As probably expected, sustainability and green energy are extremely relevant, if not central, when dealing with energy systems. Such a research popularity is proportionally reflected also in the application of synthetic data, for instance in the field of wind [140,143,144,145,146,147] and photovoltaic [148,149,150,151] power.
• Nuclear Energy. A very specific field with unexplored opportunities in the current technological climate [152], often characterized by a certain inter-disciplinarity [153,154].

3.6. Cybersecurity

While quantitatively still less relevant than other domains (Fig 2a), the application in cybersecurity has tangibly gained relevance in the last few years (Table A6). That is probably an indirect consequence of the strong and complex relationship between cybersecurity and AI.

Our analysis points out a diversified application, from big contexts (such as Smart City [155], Social Networks [156], IoT [157] and VANET [157]) to more specific aspects (for instance human factors [158]), along a number of more generalized contributions (e.g. Open Data [159] and Simulation [160]). Current research suggests future increasing trends in the field.

3.7. Technology: Computer Vision, Machine Learning and AI

Technology development became extremely data-voracious, in terms of both scale and complexity. Synthetic data has gained relevance in the field consequently. In general terms, the related work, reported in Table A7, can be summarized as follows:

• Conceptual & Methodological Foundations. A minor, yet significant, number of works provides a conceptual ground, in terms of systematic reviews/surveys, theoretical guidelines and discussion papers. More concretely, there are several papers providing a relatively generic value in the field, such as a broad survey on challenges and ethics [161], a data augmentation taxonomy [162], as well as more specific contributions (e.g. process modeling for industrial vision [163], deep learning fundamentals [164,165], and feature selection methodologies [166]).
• Machine Learning & LLMs. Research on synthetic data has evolved from early applications in specific tasks to a more comprehensive technological approach to address the gap between reality and synthesis [167,168]. Overall, such a trend resulted in more sophisticated solutions aimed, among others, at eliminating dataset bias [169], improving model fairness [170], and defining more dynamic learning pipelines [171,172].
  The impact of synthetic data is tangible in the evolution of Deep Learning systems [162,164,165] and, similarly, in LLMs [173,174,175]. Key sub-topics include mathematical and long-context reasoning [176,177,178,179,180,181], specialized fine-tuning [182], instruction tuning [183,184], information retrieval [185], low-resource translation [186], and reducing model sycophancy [187].
• Computer Vision. Synthetic data have been and are extensively used in the different applications within the widely diverse Computer Vision domain [188].
  A primary application area leverages synthetic imagery to overcome the scarcity gap, as well as to extend or enrich existing datasets. Research in this category spans general object detection [189,190], indoor scene[191], 3D workflow integration [172], semantic segmentation [192,193,194,195], as well as scene text recognition [191,196,197].
  A more narrowed application field is Face Recognition and Biometrics, where synthetic data is pivotal to mitigate bias and addressing severe privacy constraints [198,199,200,201,202,203]. Related applications also cover 3D face reconstruction [204] and visual similarity [205].

4. AI-Driven Topic Modeling

By processing with BERTopic the titles and abstracts of the selected papers (Table A1, Table A2, Table A3, Table A4, Table A5, Table A6 and Table A7), 8 main different topics have been identified (Figure 3).

The composition in terms of tokens (Figure 3) allows a clear characterization of the topics, which is clearly consistent with the human analysis previously proposed. Topic 0 can be clearly associated with applications in the health domain and its critical privacy requirements. Also Topic 2, Topic 3 and Topic 5 present similar generic patterns, with a focus on finance, energy and cybersecurity respectively. All mentioned topics reflect broad application fields, matching some of the application domains identified by human analysis.

The remaining topics presents a more narrowed focus. Topic 4 and Topic 6 are semantically correlated, by proposing a common technology-oriented character (mainly machine learning for the first topic and LLMs for the second one). Topic 1 and Topic 7 are in a strict symbiosis, as the former results from applications involving images, while the latter seems to be even more specific with a focus on face recognition.

From a strictly quantitative perspective, the identified topics may be arranged as a function of the size of the corresponding clusters (Figure 4a). According to such a metric (Figure 4b), Healthcare is the most relevant topic, followed image-related applications and finance.

The similarity matrix (Figure 5) highlights a conceptual similarity between healthcare (Topic 0) and finance (Topic 2), which share privacy concerns as per previous token analysis (Figure 4), and between applications involving images (Topic 1) and faces (Topic 7), reflecting a hierarchical relationship from generic to specific .

5. Discussion

The application of synthetic data is still object of discussion across the different domains, as a significant corpus of papers with a conceptual/critical focus confirm. This section aims at concisely addressing three main discussion points, including (i) implications at different levels (e.g. legal, ethical and political), (ii) key research gaps, and (iii) limitations of this research.

5.1. Legal, Ethical and Political implications amid Opportunities and Challenges

Unprecedented opportunities (and expectations) fostered by constantly increasing technical capabilities are clearly recognized, in general terms as well as in more specific contexts across the different application domains. Such opportunities are often, if not always, organically discussed looking at related challenges and rising trade-offs.

The affordability is clearly acknowledged in most analyses. However, it is not often considered a determinant for the adoption of synthetic data, which seems primarily aimed at overcoming the typical limitations and concerns associated with potentially sensitive data, and, secondarily, at bridging availability gaps.

Recent contributions (e.g. [206,207,208], among others) have explicitly focused on legal, ethical and political implications, as well as on key concepts in the field (e.g. Fairness [209]). Such a conceptual body has intrinsically enabled a more abstracted level of analysis and set the basis for a more holistic perspective. Resulting principled frameworks and guidelines are considered to have tangible practical implications and to be critical for operationalization.

5.2. Research Gaps: A Call for Synthetic Data Engineering

In holistic terms, the most significant limitations emerged from our analysis is a fundamental lack of an engineering approach to synthetic data beyond "guidelines" (e.g. [210]) and extensive considerations on the relevance of evaluation/validation (e.g. [211]).

A more fine-grained analysis has led to the identification of the interrelated gaps reported in Table 2.

5.3. Current Limitations

The review addresses a broad topic in constant and dynamic evolution as a response to actual needs and technological evolution in the different domains. Such a process is fundamentally led by advances in data science and its applications.

Despite the structured paper retrieval process, the proposed data-driven analysis intrinsically presents some limitations, specifically in terms of exhaustiveness, as the actual application of synthetic data is extremely ramified. Overall, we believe that our approach is effective enough to synthesize the current state of the art and body of knowledge, with a focus on recent technology-driven evolving trends.

The classification schema has been dynamically developed as part of the retrieval process. This pragmatic approach resulted to be very effective in fact, but questionable from a formal perspective. As a possible alternative, an input taxonomy could have facilitated the identification of some peculiar applications. However, by constraining the search strategy,it would have probably also limited the capability to effectively capture the actual trends in the current technological climate. Additionally, a formal taxonomy would have not properly addressed domain overlapping, as well as inter and cross-domain contributions.

The most significant simplification is probably related to the characterization of the selected papers. It relies on a binary approach for the type (conceptual/application) and on a limited number of generic and abstracted keywords. This simplified approach is critical to foster a focused analysis beyond quantitative considerations. However, it inevitably limits the capability to establish fine-grained analysis. In the specific broad context of this work, we believe that the former factor is a determinant to enable an effective re-elaboration and analysis of the content in the literature, while the latter is not a determinant for the quality of the analysis.

Finally, topic modeling has been performed on the title and abstract only. Although this approach does not fully exploit the scalability of AI-powered analysis of large documents, it provides a focused consolidation step to sustain a semi-qualitative analysis in a broad and diverse context across multiple disciplines.

6. Conclusions and Future Work

The application of synthetic data has been recently reviewed in a few specific domains, or it has been indirectly discussed in a more fragmented way. The unprecedented AI capabilities have naturally enabled and quickly consolidated a feedback loop, where synthetic data is, at the same time, a valuable resource to further develop and refine technology, as well as an added, often critical, value within the different application domains.

Addressing such a broad topic required hybrid methods and abstractions. Decisions have been transparently reported in the paper, and results considered in the context of potential limitations.

In this evolving context, this review provides a unified view, where opportunities, trends and current challenges are considered organically. In quantitative terms, the application of synthetic data is definitely in an ascending phase, in domains where the adoption is somehow considered to be usual, as well as in domains where their use was not very common. Depending on the context, there is still some gap between expectations and actual benefits, and, in some cases, largely unexplored opportunities. The analysis has facilitated the identification of research gaps along multiple dimensions, which, overall, seems converging towards a need for a more consistent engineering approach to synthetic data.

Future work might adopt ontology-driven classification with enhanced analysis support [212], and might be aimed at tracking the evolution of the application scenario, including also the establishment of regulatory frameworks and synthetic data engineering.