Neuro -symbolic AI Approach for the Classification of Gunshot Residue

Keywords

Artificial intelligence (AI), symbolic AI, neuro-symbolic AI, ontologies, rule-based reasoning, machine learning, classification, interpretability, explainable AI, forensics, gunshot residue, decision support.

Description

General context

The topic is part of the IGNIS project (Intelligent System for Gunshot Residue Classification), jointly led by the Laboratoire d’Informatique de Bourgogne (LIB), the Institut de Mathématiques de Bourgogne (IMB), and the French National Forensic Science Service (SNPS). The IGNIS project aims to design an automatic and explainable classification system for gunshot residues observed by scanning electron microscopy coupled with energy-dispersive spectroscopy (SEM-EDX). The research environment is interdisciplinary, at the interface of artificial intelligence, knowledge representation, machine learning, statistics, optimisation and forensic science. It benefits from access to real data from SNPS laboratories.

Scientific context

Gunshot residues are particles produced when a firearm is discharged, consisting of inorganic and organic components originating from ammunition. Their SEM-EDX analysis is now considered the reference technique, as it enables observation of particle morphology and estimation of elemental composition, in accordance with the ASTM E1588-20 standard framework. In practice, the classification of gunshot residue particles relies on an initial automated analysis phase, followed by manual validation by an SNPS expert.
This situation raises several scientific challenges: (1) formalising the expert knowledge used during particle validation; (2) the robustness of learning models, which is strongly dependent on data quality, representativeness and evolution, particularly with the emergence of new ammunition types and lead-free compositions; and (3) explainability, which is central in a judicial context.
The thesis aims to go beyond these limitations by proposing a neuro-symbolic AI approach capable not only of classifying particles, but also of explaining its decisions in language compatible with professional expertise. The scientific originality of the project lies in integrating a formal representation of forensic knowledge into a neuro-symbolic architecture that combines data-driven learning with explicit reasoning grounded in expert knowledge.

Research questions

The thesis will seek to answer the following question: how can a neuro-symbolic artificial intelligence system be designed to automatically classify gunshot residue particles in a reliable, adaptable and explainable way, while complying with the scientific, operational and judicial constraints specific to forensic science?
This general issue breaks down into several research questions:

• How can expert knowledge relating to gunshot residues be formally represented, integrating ASTM standards, morphological criteria, elemental compositions and business rules, within a machine-readable knowledge base?
• How can this symbolic representation be combined with learning models to refine existing classification rules, learn new classes or sub-classes of particles, and maintain system robustness?
• How can understandable and acceptable explanations be produced for SNPS experts, by combining symbolic reasoning, post-hoc explanation methods and sensitivity analysis?
• How can such a system be evaluated not only in terms of algorithmic accuracy, but also in terms of operational usefulness, integration safety, reproducibility and potential transferability to other laboratories?

References

Maitre, M., Kirkbride, K. P., Horder, M., Roux, C., & Beavis, A. Current perspectives in the interpretation of gunshot residues in forensic science: A review. Forensic Science International, 270, 1-11, 2017.

ASTM International. ASTM E1588-20 – Standard Practice for Gunshot Residue Analysis by Scanning Electron Microscopy/Energy Dispersive X-Ray Spectrometry, 2022.

Mandel, M., Israelsohn Azulay, O., Zidon, Y., Tsach, T., & Cohen, Y. Classification Improvements in Automated Gunshot Residue (GSR) Scans. Journal of Forensic Sciences, 2018.

Matzen, T., Kukurin, C., van de Wetering, J., et al. Objectifying evidence evaluation for gunshot residue comparisons using machine learning on criminal case data. Forensic Science International, 335, 111293, 2022.

de Bie, K., Vinkenoog, M., Arins, S., et al. Proposed method to objectively evaluate gunshot residue comparisons does not generalise to different-location settings. Forensic Science International, 369, 112414, 2025.

Lundberg, S. M., & Lee, S.-I. A unified approach to interpreting model predictions. NeurIPS, 2017.

Ribeiro, M. T., Singh, S., & Guestrin, C. Why should I trust you? Explaining the predictions of any classifier. KDD, 2016.

Sobol, I. M. Global sensitivity indices for non-linear mathematical models and their Monte Carlo estimates. Mathematics and Computers in Simulation, 55, 271-280, 2001.

Candidate profile

Candidates must hold a Master 2 or an engineering degree in computer science, artificial intelligence, data science, or applied mathematics. Candidates must have a good level of French and English (at least C1). Candidates must have an interest in research, a strong scientific background, and programming skills. Skills in machine learning, knowledge representation and big data analysis are expected. Skills in statistical modelling, optimisation or symbolic reasoning will be an asset.

The desired profile also requires strong writing skills, the ability to communicate with subject-matter experts, a taste for interdisciplinary work, and the ability to operate in a partnership-based environment under confidentiality and security constraints. An interest in forensic science or in fields with strong traceability requirements would be particularly appreciated. Due to the constraints linked to the sensitive application context, the selected candidate will be approved following a background check.

The profile will be particularly suitable for someone able to work at the interface between fundamental research and technology transfer, with a taste for building prototypes, analysing real-world data, and co-designing solutions with institutional partners.

Administrative information

Application deadline: 13th of July 2026

Starting date: October 1st 2026

Supervisor(s): Ana ROXIN PU (lead) / Laurence DUJOURDY IR-HDR (co-lead) / Ludovic JOURNAUX MCF (co-lead)

Contact

Applicants are invited to submit their applications to the PhD supervisors.
The application must include the following documents:

• CV
• A cover letter
• M1/M2 transcript of records (or equivalent degree)
• At least one reference letter