(1) Context and motivation
Large-scale digitization initiatives in galleries, libraries, archives, and museums (GLAMs) have made vast numbers of historical documents available and accessible for scholars (Sotirova et al., 2012; Terras, 2015; Chowdhury & Ruthven, 2015). Yet access to page images does not automatically yield data that are searchable, interoperable, or reusable for research (Nockels et al., 2024). Many collections remain difficult to analyze because their content is not available in structured, machine-readable formats. This bottleneck is particularly evident for historical exhibition catalogues, which contain rich information on artists, artworks, prices, locations, and institutional practices, but are typographically heterogeneous and structurally complex (Joyeux-Prunel & Marcel, 2015; Joyeux-Prunel, 2019).
Within digital art history, much attention has been given to databases, computational analysis, and digital forms of publication, as well as to the methodological implications of working at scale (Rodríguez-Ortega, 2013; Baca et al., 2019). By comparison, less attention has been paid to the preparatory stages that make such work possible: the extraction, structuring, and normalization of historical source material. This gap is especially visible in the case of exhibition catalogues, which are central sources for the study of artistic circulation and participation but remain difficult to formalize as reusable data (Joyeux-Prunel & Marcel, 2015).
This discussion paper argues that the difficulty of transforming such catalogues into reusable data cannot be understood solely as a problem of open character recognition (OCR) accuracy or extraction performance. In the publicly available set of Exhibitions of Living Masters catalogues held at the RKD – Netherlands Institute for Art History, the obstacles to structuring data arise at multiple levels: from page-level issues of legibility and segmentation, from inconsistencies across catalogues, and from contextual conventions that remain only partially explicit in the source itself (Philips & Tabrizi, 2020). What is often described as “messy data” is therefore better understood as a form of structural complexity embedded in the documentary, historical, and institutional conditions of the corpus. A broader body of research in historical document processing has shown that the transformation of digitized documents into structured data depends not only on OCR quality, but also on layout analysis, preprocessing, and domain-specific representations (Philips & Tabrizi, 2020; Goes et al., 2019; Khemakhem et al., 2020).
To address this problem in the context of exhibition catalogues, the paper develops a taxonomy of structural data complexity in exhibition catalogues and uses selected artificial intelligence (AI) assisted extraction experiments to clarify how different forms of complexity generate different kinds of failure and intervention. The aim is not to propose a universal extraction method or a benchmark evaluation. Rather, it is to show why apparently similar difficulties in catalogue processing do not arise from the same source, and how different approaches perform in relation to these structural data challenges.
In this paper, we first propose a taxonomy that distinguishes among document-level, cross-catalogue, and contextual forms of structural complexity in historical exhibition catalogues. Second, we use exploratory encounters with OCR-based and AI workflows to show how these different forms of complexity affect extraction unevenly. Finally, we discuss the implications of these findings for the preparation, validation, and reuse of humanities data in digital art history and cultural heritage research.
(2) Corpus description
The discussion in this paper is based on a corpus of 19th-century exhibition catalogues from the Exhibitions of Living Masters held in the Netherlands between 1808 and 1917 (Koot & Kapelle, 2015). The catalogues are part of the RKD – Netherlands Institute for Art History.
The Exhibitions of Living Masters were a pivotal feature of the 19th-century Dutch art scene, taking place in various cities across the Netherlands from 1808 to 1917. Initially organized in major cultural centers like Amsterdam, Rotterdam, and The Hague, the exhibitions gradually expanded to smaller cities including Haarlem, Dordrecht, Den Bosch, Utrecht, Arnhem, Leeuwarden, and Groningen. These exhibitions provided a unique, commission-free platform where artists, ranging from emerging talents to established figures, could showcase and sell their works across diverse disciplines. They also fostered artistic exchange by bringing together art collectors, critics, dealers, and artists from different regions and countries, stimulating the growth of the Dutch art market. The accompanying exhibition catalogues serve as vital primary sources for understanding this dynamic period in Dutch art history. Exhibition catalogues from this period have traditionally functioned as essential documents that provide insight into participating artists and artworks. They vary widely in scope, ranging from simple listings of works to extensive research tools that may include detailed artist biographies, illustrations, prices, and additional contextual information such as venue details. The Living Masters catalogues follow this diverse tradition, offering rich but sometimes inconsistent snapshots of the exhibitions. Entries typically include the artist’s name and city along with the titles of artworks. Some catalogues offer further details such as brief descriptions, poems, or notes on the medium and genre of the works. A system of abbreviations is commonly used to denote important information: participation in competitions, previous awards, or membership in art associations. Additional symbolic notations include crosses to indicate recently deceased artists or asterisks to mark works not for sale. Moreover, many catalogues contain front- and back-page price lists with corresponding currency values, offering valuable data on the commercial aspect of the exhibitions.
The corpus comprises 877 volumes representing 234 unique catalogue editions and spans a period in which catalogue structure, descriptive conventions, and exhibition practices changed substantially over time. The source material is available as scanned PDF files derived from printed catalogues in the RKD Library.1 It is therefore publicly accessible as digitized documentary material, but not as natively structured, machine-readable data. Its transformation into reusable data requires the extraction and formalization of information on artists, artworks, media, prices, dimensions, owners, and exhibition locations. These entities and attributes do not appear in stable or uniform form across the corpus. Instead, the catalogues contain substantial variation in page layout, section structure, typographic emphasis, abbreviations, multilingual expressions, and implicit cross-references, all of which complicate computational processing.
(3) Sources of Structural Data Complexity in Exhibition Catalogues
While technical obstacles to digitization and data extraction are often framed in terms of OCR errors or layout recognition, our experience working with the Exhibitions of Living Masters catalogues revealed a broader and more systemic set of challenges. These go beyond technical parsing and speak to the interpretive, linguistic, and infrastructural messiness inherent to historical archival material. In this section, we synthesize these challenges into a taxonomy that categorizes the structural complexity of digitized exhibition catalogues across three interrelated levels: (1) document legibility and semantic noise, (2) cross-catalogue inconsistency, and (3) contextual and institutional frameworks. This taxonomy is intended as both a diagnostic framework for similar projects and a methodological contribution to digital art historical scholarship. Figure 1 shows a synthesis of the taxonomy, while Figure 2 illustrates the sources of structural data complexity in two selected catalogues.
Figure 1
Sources of structural data complexity in exhibition catalogues.
Figure 2
Annotated catalogue pages illustrating data complexity in exhibition catalogues: (a) legibility issues such as faint print or overwritten text, (b) implicit reference to the artist in the previous line (‘dezelfde’), (c) ambiguous metadata formatting and structure, (d) extensive annotations including additional metadata such as lender and material information, (e) multilingual text, and (f) layout inconsistencies across catalogues. Photo RKD — Nederlands Instituut voor Kunstgeschiedenis.
(3.1) Document legibility and semantic noise
The first category involves challenges at the level of the individual exhibition catalogue: how the visual, material and semantic qualities of the text frustrate machine readability.
Text legibility: Textual information in scanned documents can represent a challenge for legibility and therefore extraction. While it is true that OCR and handwritten text recognition (HTR) methods have improved greatly in recent years, and show high performance on text extraction, characteristics of historical documents increase the difficulty of extraction. These range from original source degradation over time, to difficulty handling handwritten text. Text legibility is a major concern for historical data extraction, as every subsequent step of structuring and analysis relies on quality and accuracy of the extracted text.
Implicit content: A recurring obstacle is the presence of implicit or abbreviated entries. For instance, artists listed in multiple consecutive entries often appear only once, with subsequent rows marked by shorthand such as “idem,” “ditto,” or even quotation marks. While readers readily infer continuity, models often fail to replicate this logic, leaving fields blank or misattributing values.
Semantic ambiguity and interpretation gaps: catalogues often rely on semi-formal or assumed conventions (e.g. parentheses, italics, or indentation) to distinguish between titles, media, ownership notes, or quotes. Without standardized punctuation or schema, it becomes difficult to disambiguate whether a term like “(aquarel)” refers to the medium, the technique, or a parenthetical aside. Large language models (LLMs) can infer context in many cases, but their interpretations were not always consistent or transparent, especially in ambiguous or compound entries.
Post-production annotations and marginalia: Historical catalogues commonly contain handwritten notes – such as comments, corrections, or additional notes – added by viewers, collectors, or institutional staff. While often of historical interest, these annotations introduce additional noise during extraction, especially in cases where handwritten and printed text appeared in close proximity. LLMs fared better than OCR engines at excluding these elements, but their presence still complicates the extraction pipeline, raising broader questions about what constitutes “authoritative” content in a digitized archival object.
(3.2) Cross-catalogue inconsistency
The second category pertains to inconsistencies visible across catalogues: how differences in form, layout and language hinder clean extraction.
Layout and font variation: Catalogues in our corpus displayed substantial variation in layout between years and organizing institutions. Artist names might appear in bold at the top of a section in one volume, while in others, they are embedded mid-paragraph or preceded by a numbering scheme. Similarly, titles may appear on the same line as the artist or separated by a line break. These inconsistencies – while semantically clear for readers – function as model traps for extraction models, as they undermine expectations of spatial hierarchy or alignment. Even with preprocessing steps such as binarization and segmentation, layout variation remains a persistent issue.
Inconsistent metadata structures: While all catalogues contain core categories such as artist name and artwork title, the presence and order of additional metadata (e.g. artist address, awards, or pricing) varied greatly. For example, price lists might appear on a separate page, use different currencies, or include only partial coverage of artworks. Even where fields were present, formatting was inconsistent: currencies might appear as “f,” “gulden,” “ƒ,” or spelled out entirely, complicating parsing.
Multiplicity of languages: Many of the catalogues use 19th-century Dutch, sometimes mixed with French, German, or Latin terms depending on city or institutional affiliation. These language shifts challenge not only recognition, but also the downstream task of categorizing metadata, such as standardizing city names or detecting medium types.
(3.3) Contextual and institutional frameworks
The final category involves structural issues that emerge particularly in relation to external frameworks such as institutional conventions and larger databases.
Historical shifts: As the exhibition series spanned over a century, artists moved cities, changed signatures, or were listed under various aliases. Institutional naming conventions also evolved. These changes introduce inconsistencies that are difficult to resolve algorithmically but are crucial for disambiguating entries and enabling linked data efforts. Moreover, the same artist may appear with slightly different metadata across years, posing a challenge for record matching and normalization.
Legacy structuring bias: Previous efforts at manual or semi-automated catalogue structuring introduced an existing structure of metadata and information selection that needs to be reconciled with the data at hand. In the absence of a validated unified data structure, manual curators keep as close as possible to the catalogues, retaining metadata types inconsistently: The field ‘location’, for instance, might be used to refer to artwork location or artist origin, and can further differ in level of field granularity depending on the catalogue needs. In addition, the legacy structure also affected the focus of the extraction. In this case, since it was centered on entries, rather than catalogues, we ignored catalogue-level information such as opening days, pricing or membership options which might have been valuable in a different structuring scheme. In this context, legacy structuring sets the tone for the extraction process and exacerbates the need for reconciliation.
This taxonomy synthesizes the multi-layered messiness of exhibition catalogue data and reflects the way that cultural records are produced, annotated, and preserved. Importantly, these categories are not isolated; layout irregularities exacerbate semantic ambiguity, while missing data complicates structural alignment. Identifying these challenges is an essential step for developing effective, scalable, and epistemologically sound methods of historical data extraction and structuring.
(3.4) From catalogue entry to structured record
Before extracting the data from the catalogues, the first step was to define a unified data structure that captured the necessary categories while allowing variability in others. Because the primary aim of the present workflow is integration into an existing database, we prioritize the extraction and normalization of entry-level information over encoding approaches such as TEI, which are better suited to preserving the full textual layering of the catalogue. We collaborated closely with a team of experts (curators, data specialists) at the RKD and examined their subset of manually structured catalogues in order to develop a unified data dictionary. The dictionary consists of three meta-categories:
(≡) Literal types: standard categories such as artist name, entry number, and artwork title
(△) Standardized types: categories which we standardize to ease subsequent manipulation, such as a newly added currency column, or artist address.
(∪) Catch-all types: meant to keep the original wording and formatting in order to retain the specificity found in the original catalogues.
Table 1 summarizes these fields, their function in the schema, and shows how a sample catalogue entry (Figure 3) is structured based on them.
Table 1
Data dictionary synthesizing the category titles, their meanings, and the reasoning behind their creation.
| CATEGORY | MEANING AND REASON | EXAMPLE |
|---|---|---|
| Catalogue ID ≡ | Document title. Catalogue ID and database linker. | 201503502 |
| Keyword Person ≡ | Full name of the artist. Matches category title of existing database. | A. H. BAKHUIJZEN, Jr. |
| Artist City ≡ | City of artist. Location marker. | te ’s Hage |
| Artist Address △ | Extended address line if applicable. Additional location marker, less often specified in catalogue. Valuable for insight on artist mobility, especially for international artists who were more likely to provide full addresses. | — |
| Artist Abbreviation ∪ | Catch-all category for exhibition-relevant information, such as membership and competition status. Not all exhibitions used the same abbreviations, and artists sometimes had multiple in the same entry. Created as a catch-all due to non-standardized abbreviations across exhibitions. | BL |
| Entry Number ≡ | Entry ID. Listing number of the artwork in the catalogue. | 1 |
| Free Title ≡ | Title of the artwork. Artwork distinction and matches existing database category. | Een Panorama in de omstreken van ’s Gravenhage |
| Additional artwork info ∪ | Anything listed about the artwork that is not part of the title, such as material, quote, ownership, or location. Created as a catch-all due to the variety of entries and inconsistencies across catalogues. | unknown |
| Asterisk △ | True/False depending on the presence of the * symbol. This symbol usually indicates whether an artwork is for sale, but can have other meanings throughout the catalogues. The meaning is only found on introductory pages. | False |
| Amount Type ≡ | Asking price. Matches the existing database. | — |
| Currency △ | Currency in alphabetical abbreviated label, for example HFL. This format matches the existing database instead of a currency symbol or other abbreviations. | — |
| Price ≡ | Price of the artwork. Matches existing database and adds price data to artworks. | — |
| Full Entry Quote ∪ | Complete quote of the entry as formatted in the catalogue. Matches existing database and stores data without adjustments. | A. H. BAKHUIJZEN, Jr., te ’s Hage. B.L. 1 Een Panorama in de omstreken van ’s Gravenhage. |
Figure 3
Entry from the Tentoonstelling van schilder- en andere kunstwerken van levende meesters, held in Amsterdam by the Maatschappij Arti et Amicitiae, 1857. Photo RKD — Nederlands Instituut voor Kunstgeschiedenis.
Building on the data dictionary, we iteratively refined the prompt used in the multimodal workflow to improve the consistency of entry-level extraction. In the OpenAI-based pipeline, catalogue PDFs were converted into 300 dpi JPEG page images and submitted to the gpt-4o model through the Chat Completions API, using a low-temperature setting (0.2) to favor stable structured output. The model output was then supplemented by rule-based post-processing, including the propagation of repeated values, normalization of asterisk markers, filtering of incomplete rows, and the reconciliation of prices with entries through entry numbers. Figure 4 shows a representative prompt from this exploratory phase.
Figure 4
Representative prompt used in the VLM workflow to extract and structure catalogue entries into predefined fields. As the model’s behavior evolved over time, the prompt required periodic refinement during the exploratory phase.
(4) Initial findings on AI-assisted extraction and structuring
(4.1) Process
Transforming catalogue pages into database-ready data requires two steps: (1) extraction, obtaining raw text from scanned PDFs, and (2) structuring, mapping that text onto the data dictionary defined. In this study, we conduct an initial qualitative comparison of two AI-assisted workflows: (a) a sequential OCR+LLM workflow, in which OCR output is processed by large language models; and (b) a vision-language model (VLM) approach, in which a vision-enabled model operates directly on page images and produces structured output without a separate OCR step.
Given the absence of a validated ground truth dataset, and the experimental nature in this phase, we chose to evaluate the two pipelines against the structural data challenges introduced in Section 3, using qualitative inspection to interpret where each approach succeeds consistently, partially, or minimally. Here, consistent, partial, and minimal are used as qualitative assessment categories based on the manner in which the pipeline handled the challenge. We noted ‘consistent’ when the workflow handled a given challenge reliably with limited correction; ‘partial’ when it captured some aspects of the challenge but required repeated manual intervention; and ‘minimal’ when it did not address the challenge and left substantial corrective work to the researcher. In both workflows, extraction is guided by domain-informed prompts that describe the catalogue structure, the target fields, and how ambiguous or missing information should be handled.
In the OCR + LLM pipeline, prompts instruct the model to segment OCRed text into individual entries and to map tokens to fields in the data dictionary. Additional rules specify how to deal with abbreviations and repeated information.
In the VLM approach, prompts explain the visual structure of the pages (columns, headings, numbering) and ask the model to identify entry boundaries and extract all visible fields, including those that are only implied by layout or repetition markers.
Prompt design is iterative: outputs are reviewed to identify recurring errors, leading to refinements in the prompt wording, the data dictionary, and post-processing rules.
(4.2) Comparative performance of the pipelines
The VLM pipeline generally outperforms the OCR + LLM approach across several categories of structural complexity, particularly where layout, visual cues, and implicit repetition are crucial:
Document-level legibility and annotations. The VLM approach more robustly distinguishes main entries from marginalia, resolves multi-line entries, and maintains field boundaries despite irregular line breaks or column shifts.
Implicit content and multilinguality. The VLM approach is better able to propagate values indicated through “idem” or similar markers, and to interpret entries that blend Dutch, French, and other languages within a single line.
Cross-catalogue layout variation. When catalogue formats change across years and institutions, the VLM adapts more effectively to new column structures and typographical emphasis.
However, both pipelines remain fragile in relation to semantic ambiguity and contextual knowledge. Particularly, they show limitations in the following fields:
Abbreviations and addresses. Initials are sometimes misinterpreted as catalogue-specific abbreviations, and street addresses are merged with city names, complicating geocoding and authority file alignment.
Pricing fields. Attributes such as price type and currency are inconsistently extracted, leading to simplification in the final dataset.
Over-correction and paraphrasing. Models occasionally standardize spelling or paraphrase artwork titles, which may not be desirable for some philological or attribution-focused research questions.
These patterns highlight that the structured data should be treated as a proposed representation of catalogue content needing supervision and validation from domain experts. Maintaining links to page images and providing access to full entry text remain crucial for responsible reuse. Table 2 outlines the data challenges along with the degree of observed resolution per method, the typical failure mode and intervention required for repair.
Table 2
Comparative performance of OCR+LLM and VLM pipelines across structural data challenges.
| LEVEL | DATA CHALLENGE | DEGREE OF RESOLUTION | FAILURE MODE | INTERVENTION NEEDED | |
|---|---|---|---|---|---|
| OCR+LLM | VLM | ||||
| Document | Text legibility | Partial | Consistent | Character-level noise propagates into fields | Spot-check output and correct errors |
| Document | Implicit content (e.g. “idem”, “dezelfde”) | Minimal | Consistent | Field left blank or misassigned | Adjust prompts or manually propagate values |
| Document | Semantic ambiguity and interpretation gaps | Partial | Partial | Classification of values in incorrect fields | Review and correct field assignments |
| Document | Post-production annotations and marginalia | Minimal | Consistent | Incorrect extraction or annotation merged with entry text | Filter out or verify non-original content |
| Document | Multilinguality | Partial | Consistent | Language switching confuses field assignment | Standardize or translate extracted values |
| Cross-cat. | Font variation | Partial | Consistent | Character forms or typographic emphasis are misread or inconsistently interpreted | Correct misread values and normalize formatting cues |
| Cross-cat. | Layout variation | Partial | Consistent | Entry splitting or merging | Correct segmentation or merge entries |
| Cross-cat. | Metadata variability | Minimal | Partial | Fields missing or inconsistently populated | Manually align outputs to schema |
| Cross-cat. | Language drift | Minimal | Partial | Inconsistent field assignment and lack of normalization | Standardize terms across corpus and reconcile variants |
| Contextual | Historical shifts | Minimal | Minimal | Entity variation across time prevents matching | Disambiguate entities using external sources |
| Contextual | Legacy structuring bias | Minimal | Minimal | Conflicting field definitions and granularity | Reconcile fields with existing database conventions |
(5) Implications/Applications
This paper contributes a formalization of the structural sources of data complexity in exhibition catalogues, along with initial findings from a comparative workflow using OCR and vision-language models. The taxonomy outlines different types of challenges in exhibition catalogues, supporting researchers in designing appropriate workflows for data extraction and structuring. Our initial comparison of OCR+LLM and VLM approaches shows that layout heterogeneity and semantic ambiguity often stump traditional extraction approaches, and that OCR-based extraction ends up propagating legibility errors to subsequent stages of classification. These findings expand on recent work identifying OCR and LLM as a methodological workflow toward validated historical datasets (Motilla-Chávez et al., 2026), as the combined extraction and structuring allowed by multimodal models shows improved results on heterogeneous data.
This is especially relevant for digital art history, where exhibition catalogues are increasingly used to reconstruct participation, circulation, institutional networks, and other large-scale historical patterns (Joyeux-Prunel & Marcel, 2015; Joyeux-Prunel, 2019). The significance of the present approach is that it supports the creation of structured exhibition data from sources that are rich but formally unstable. At the same time, the method and resulting data may also be valuable to print culture, social and cultural historians, or GLAM professionals interested in how exhibition catalogues function as documentary artefacts and as structured historical sources. In the case of the Exhibitions of Living Masters, our approach has already led to such exploration. One example is subsequent research on exhibition access, which uses the structured data derived from these catalogues to examine how access was distributed across the Netherlands between 1808 and 1917, including the geography of exhibitions, their duration, and opening conditions (Zandieh Doulabi et al., 2026). There, the formalization of catalogue information into fields such as location, duration, opening hours, entry prices, and membership arrangements made it possible to analyze exhibition access at scale across 234 catalogues. The resulting analysis showed how access was structured by intersecting geographic, temporal, economic, and gendered constraints, illustrating how the workflow supports historical questions that would be difficult to pursue systematically from page images alone.
Yet the analytical value of such structured outputs should not obscure the conditions of labor under which they are produced. The current value of AI-assisted extraction lies less in full automation than in large-scale data preparation and pre-structuring prior to expert validation. In labor terms, this creates a tension. Labor shifts away from transcription and manual data entry toward data design, prompting, exception handling, manual verification, and validation. That shift may be productive, but it should not be taken as neutral. It recalls longer-standing critiques in digital humanities and digital art history concerning neoliberalization, platform logics, and the tendency of data-driven scholarship to obscure the forms of labor on which it depends (Jarlbrink, 2020).
Conclusion
This discussion paper has examined how structural data complexity in exhibition catalogues shapes the possibilities and limits of AI-assisted extraction and structuring.
Our first contribution is a taxonomy that distinguishes document-level, cross-catalogue, and contextual forms of structural complexity in exhibition catalogues, clarifying how legibility issues, layout variation, metadata inconsistency, and historical shifts affect data formalization in different ways. Second, through a qualitative comparison of OCR+LLM and VLM workflows, we show that VLMs generally better handle layout cues, implicit repetition, multilinguality, and marginalia, while both approaches remain fragile when confronted with semantic ambiguity, abbreviations, and context-dependent fields such as addresses and pricing. Third, we use these findings to articulate implications for humanities data preparation and reuse, arguing that AI-assisted extraction should be understood as a provisional, prompt-dependent representation that requires domain-informed modelling, iterative validation, and continued alignment with authority files and institutional frameworks.
These outcomes have practical consequences for digital art history and GLAM institutions. The structured data derived from catalogues like the Exhibitions of Living Masters already support new analyses of exhibition histories, artistic mobility, markets, and institutional practices that would be difficult to pursue with page images alone, and offer workflows through which digitization projects can extend beyond access to images toward interoperable, machine-readable surrogates. At the same time, the lack of a fully validated ground truth dataset, the persistence of contextual and historical ambiguities, and the risk of error propagation across pipelines underline that multimodal AI does not replace scholarly interpretation or curatorial labor. Rather, its value lies in making structural difficulties more legible, in helping to prioritize where human intervention is most needed, and in shifting labor from transcription toward data design, exception handling, and critical validation. Recognizing these dynamics is essential if AI-assisted workflows are to support robust, reusable humanities data without obscuring the interpretive and institutional work on which such data ultimately depend.
Note
[1] https://www.rkd.nl/en/collection/digital-collection/rkdexcerpts, last accessed June 12th, 2026.
AI Declaration
AI was primarily used in the extraction and structuring workflow as described in the text. It was additionally used for minor text editing and formatting prior to submission.
Acknowledgements
We would like to thank Sabine Craft-Giepmans, Djairo Terpstra, Jeroen Kapelle, Reinier van ‘t Zelfde, Rianne Piening and Evelien de Visser from the RKD – Netherlands Institute for Art History for their support during this collaboration.
We also extend our gratitude to Ivan Kisjes, whose help was meaningful in putting together the technical parts of the project, and whose generosity and heart are sorely missed.
Author Contributions
Yağmur Sarigül, Pascale Stomp, Maria Elpiniki Zafeiraki, Ava Zandieh Doulabi (equal contribution, alphabetical): Conceptualization, Methodology, Software, Formal analysis, Investigation, Data curation, Writing – original draft.
Rudy Jos Beerens: Conceptualization, Supervision, Writing – review & editing.
Houda Lamqaddam: Conceptualization, Supervision, Writing – review & editing.
Yağmur Sarigül, Pascale Stomp, Maria Elpiniki Zafeiraki and Ava Zandieh Doulabi are contributed equally and are listed alphabetically.