PhD Position M/F: Digital Heritage Twins and Generative AI: New Architectures for the Cultural and Creative Industries

Job Description

Thesis Subject

Voici la traduction en anglais, dans un style adapté à une publication d'offre de doctorat.

1. Research Context

The cultural and creative industries (CCIs) are currently undergoing profound transformation through the rise of generative artificial intelligence, 3D digitization, and immersive environments (XR). These technologies enable new forms of cultural production, including heritage reconstruction, interactive storytelling, digital museography, and 3D content creation. However, they also raise major scientific and societal challenges relating to the reliability of reconstructions, traceability of sources, transparency of algorithmic processing, and bias management.

In the field of cultural heritage, the rapid democratization of image and 3D content generators has led to the widespread circulation of visually appealing reconstructions that are not scientifically justified, or are difficult to justify, and that rely on opaque and uncontrolled training datasets. This situation is particularly problematic for the CCIs: museums, heritage institutions, companies producing virtual and/or immersive content, and even local authorities must now arbitrate between visually attractive outputs and the need to ensure scientific rigor and ethical dissemination.

With the massive democratization of generative AI tools, including chatbots and image and video generators, this PhD project also contributes to addressing issues of professional ethics, scientific responsibility, and the fight against visual misinformation. It proposes methods for producing cultural content based on source traceability, explicit hypothesis formulation, and the representation of uncertainty.

In this context, the Nantes Digital Sciences Laboratory (LS2N, UMR CNRS 6004), and more specifically the CPS3 team, has for several years been developing advanced research in digital heritage and immersive museography, combining 3D digitization, knowledge management, CAD, virtual reality, archiving, and mediation tools for the web, museums, and other contexts.

These areas of expertise have recently been strengthened through the interdisciplinary PIONNIER project, which aims to study how AI and XR can enrich research in the humanities and social sciences and renew heritage reconstruction methods through a case study conducted in collaboration with a humanities and social sciences laboratory specializing in archaeology, CReAAH-LARA.

An exploratory phase funded in 2025 validated the feasibility of a complete pipeline combining an ontology-structured heritage database based on CIDOC-CRM, human-in-the-loop annotation, fine-tuning of generative models, and an expert evaluation protocol incorporating initial experiments with RLHF applied to images. This work has already led to national and international presentations, notably at the international Digital Heritage conference in Italy in September 2025. It was also presented during the ICCARE-LAB acceleration day, in collaboration with the GdR IASIS, on the theme “AI and Heritage” in November 2025.

The proposed PhD project follows directly from these results, with the aim of increasing scientific maturity: moving beyond the experimental stage in order to produce a reproducible, generalizable, and interoperable methodology serving the needs of the CCIs.

2. Scientific Objective and Research Question

The CCIs need tools capable of producing digital reconstructions and immersive content that are at once plausible, explainable, and scientifically assessable. These tools must also guarantee the traceability of sources and hypotheses, as well as the interoperability of results with national, European, and international infrastructures.

However, current generative models mostly operate as black boxes: they generate images or environments without being able to justify their choices or integrate the epistemological constraints specific to the humanities. This limitation hinders their institutional adoption by experts and their long-term integration into CCI production chains, for example in museums, publishing, audiovisual production, XR, or any other form of heritage mediation.

The objective of the proposed PhD project is therefore to design a hybrid scientific architecture combining generative AI, heritage ontologies, reinforcement learning, and probabilistic modeling, in order to formalize the historical and archaeological reasoning required to produce heritage digital twins that can be used within the CCIs.

The resulting scientific question is as follows: can historical and archaeological expertise be formalized as computational signals, including formal ontologies, probabilistic rules, reward models, and human learning signals, in order to train a generative AI system capable of producing heritage reconstructions that are assessable, explainable, and directly integrable into CCI value chains?

The scientific objectives are structured around four key challenges:

1. Formalizing heritage reconstruction reasoning, including comparative analysis, chains of evidence, and source hierarchies, as computational models.
2. Developing controlled generative models capable of integrating spatial, iconographic, and material constraints derived from standard heritage ontologies and 3D data.
3. Defining an evaluation and uncertainty quantification protocol in order to produce probabilistic results rather than “visual truths”.
4. Ensuring CCI/Huma-Num interoperability by guaranteeing FAIR, traceable, and reusable data within museographic and creative pipelines.

3. State of the Art and Theoretical Foundations

Recent work on generative AI, including diffusion models, LoRA, and ControlNet, has demonstrated the ability to produce high-quality visual content. However, in heritage applications, these approaches suffer from a lack of scientific control, leading to hallucinations, anachronisms, and medievalist biases.

At the same time, internationally recognized heritage ontologies such as CIDOC-CRM and Dublin Core provide an established foundation for structuring sources, but they remain underused as active tools for learning and guiding generative models. Finally, RLHF methods, or Reinforcement Learning from Human Feedback, have developed considerably in the field of language models, but remain exploratory for diffusion models, particularly in heritage contexts where evaluation criteria are strongly dependent on historical expertise.

The main gap identified is therefore the absence of an integrated approach combining:

* controlled generative AI;
* ontological reasoning;
* expert-based scoring;
* probabilistic uncertainty;
* use and verification of results in immersive environments.

Beyond this observation, the PhD project is grounded in a structuring theoretical framework articulated around two complementary components, respectively related to digital sciences and to the humanities and social sciences. This dual conceptual foundation aims precisely to move beyond a simple juxtaposition of disciplines. The objective is to produce an integrated approach in which computational methods become capable of formalizing forms of reasoning specific to the practices of historians and archaeologists.

The first component, rooted in digital sciences, mobilizes knowledge engineering, including ontologies and semantic graphs, in order to structure heritage data and model chains of evidence, drawing on standards such as CIDOC-CRM. It also builds on diffusion-based generative models and lightweight fine-tuning techniques such as LoRA and DoRA to adapt foundation architectures to specialized corpora. Within this framework, reinforcement learning through human feedback and multimodal reward models form a central contribution: they make it possible to explicitly integrate expert feedback into the optimization loop and to transform human evaluation into a learning signal.

Finally, probabilistic modeling and uncertainty quantification provide a mathematical framework for qualifying the degree of confidence associated with generated hypotheses, moving beyond the production of merely “plausible” images toward assessable and interpretable results.

The second component, rooted in the humanities and social sciences, lies within the fields of history, art history, and building archaeology. It considers heritage reconstruction above all as a critical process of hypothesis construction based on the analysis of primary and secondary sources. The PhD project therefore aims to advance research in history and archaeology by proposing new instruments for exploring and validating hypotheses, notably through the formalization of comparative reasoning, the explicit identification of source biases, and the visualization of uncertainties.

Immersive environments and the concept of digital twins developed by the LS2N research team thus become research tools in their own right, enabling direct confrontation between physical traces, archives, successive hypotheses, and scientific interpretations.

The articulation between these two components underpins the deeply interdisciplinary nature of the PhD project. The challenge is not merely to apply AI tools to heritage, but to produce a framework in which the scientific requirements of the humanities and social sciences—evidence, interpretation, and source criticism—become formalized and computable constraints capable of guiding digital models and renewing research practices in support of the CCIs.

4. Work Plan and Methodology

The proposed work plan follows directly from the initial experiments carried out in 2025. While these experiments validated the feasibility of the pipeline, the PhD project now aims to transform this experimental prototype into a formalized, robust, and generalizable scientific architecture.

First, the PhD candidate will consolidate and extend the structuring of heritage data by developing an operational ontology oriented toward reconstruction reasoning. Particular attention will be paid to embedding this knowledge base within the ecosystem of the French national IR* Huma-Num infrastructure and its consortia, notably 3D HN. This will involve not only enriching existing multimodal corpora, including textual archives, iconographic sources, surveys, and 3D scans, but above all formalizing the chains of evidence, dependency relations, and hierarchies of clues mobilized by historians and archaeologists. This stage aims to transform a descriptive documentary database into a genuine knowledge model that can be used computationally.

Second, the research will focus on developing a controlled generative AI system capable of integrating these ontological and geometric constraints. Diffusion models will be adapted through lightweight fine-tuning, including LoRA and DoRA, on specialized corpora, while spatial constraints derived from plans, surveys, or depth maps will be integrated to ensure architectural coherence. The challenge will be to connect 2D generation with the production of usable 3D models, while maintaining an explicit link between visual hypotheses and documentary sources.

A third phase will consist in formalizing a scientific scoring protocol that explicitly integrates expert feedback into the learning loop. Based on historiographical criteria such as iconography, stylistic coherence, materiality, and architectural plausibility, a multimodal reward model will be developed in order to implement an RLHF-type strategy applied to diffusion models. The objective is to transform human evaluation into a measurable and reproducible optimization signal.

In parallel, particular attention will be paid to uncertainty quantification and the explainability of results. Probabilistic approaches will make it possible to associate each generated hypothesis with a degree of confidence, making visible the distinction between a strongly supported reconstruction and a speculative proposal. This dimension is essential in order to avoid any deterministic interpretation of the images produced and to ensure their responsible use by the CCIs.

Finally, the results produced will be integrated into an immersive digital twin enabling hypotheses to be situated and validated in an XR environment. This system will serve both as a research tool for historians and archaeologists and as a demonstrator for the CCIs, illustrating the ability of the developed framework to be integrated into museographic and immersive production chains.

Your Work Environment

Voici la traduction en anglais :

The proposed doctoral project is fully aligned with the strategy of the PEPR ICCARE programme, which aims to structure ambitious research in support of the digital transformation of the cultural and creative industries (CCIs). It contributes to the objectives of technological sovereignty, methodological innovation, and the structuring of CCI ecosystems by proposing a scientific framework for mastering the use of generative AI in cultural and heritage production.

By developing hybrid architectures combining ontologies, generative models, reinforcement learning, and uncertainty quantification, the PhD project addresses the challenges identified by ICCARE in terms of the reliability, traceability, and interoperability of digital content. It thus contributes to the programme's structuring research questions, particularly those addressed by the Harmonie project, by proposing methodological tools that make it possible to articulate cultural content production, structured data, and national infrastructures.

As such, the project contributes to the long-term structuring of the immersive sector by providing reproducible methods that may be adopted by economic stakeholders, while strengthening the territorial and national anchoring of the research through the partnerships already established.

This research programme is also aligned with the strategic projects led by the Maison des Sciences de l'Homme Ange Guépin in the Pays de la Loire region, within the “Society and Digital Technology” research axis, confirming its structuring role at the interface between digital sciences, the humanities and social sciences, and the cultural and creative industries.

Constraints and risks

-

Compensation and benefits

Compensation

2300 € gross monthly

Annual leave and RTT

44 jours

Remote Working practice and compensation

Pratique et indemnisation du TT

Transport

Prise en charge à 75% du coût et forfait mobilité durable jusqu’à 300€

About the offer

About the CNRS

The CNRS is a major player in fundamental research on a global scale. The CNRS is the only French organization active in all scientific fields. Its unique position as a multi-specialist allows it to bring together different disciplines to address the most important challenges of the contemporary world, in connection with the actors of change.

CNRS

The research professions