AI.iO Artificial Intelligence in Orléans: Learning from heterogeneous data and expert knowledge. Applications in geological and environmental sciences – AI.i0 PhD Fellowship

The University of Orléans benefits from an important research environment in Artificial Intelligence (AI) with both fundamental and applied research activities. It is supported by three laboratories: LIFO (Laboratoire d’Informatique Fondamentale de l’Université d’Orléans), IDP (Institut Denis Poisson), and PRISME (automatics, signal and image processing). The University is also one of the partners of the “Orleans Grand Campus” that regroups several research institutes and laboratories with a high-level expertise in Environment (ranging from under soil know-how to plants) such as OSUC, BRGM, INRA and CNRS. More specifically, Orléans (BRGM, INRA, …) is one of the biggest clusters in Europe, hosting national and international geo-environmental databases.

The strategic plan of the University is to develop a strong and recognized research center in Environment and Digital Sciences. In this context, several initiatives have been taken: (1) the proposal (although not accepted) of a EUR project GEODE integrating the University of Orléans, BRGM, INRA, CNRS, Atos and ANTEA Group in the field of Digital Sciences and Environment, (2) the proposal of a chair of research and teaching in AI, Ch.A.I.R.E.-O (Chair Artificial Intelligence Research for Environment in Orléans, leader F. Ros).

This project aims at strengthening the research activities in fundamental AI with a focus on applications in the field of Environment and Cultural Heritage. In these domains we have often to deal with a great amount of data coming from different sources, leading to different types of data that could come at regular time or at any moment and could also come from different places. All these lead to the generation of heterogeneous data (text, image, signal, captors) described at different granularity degrees. It is important to point out that in many applications in these two domains the temporality is an important factor. Temporal data have already been a lot studied in the case in which the data comes at predefined time stamps. In this project, we consider the case in which events could come at any given moment of time and are heterogeneous. Also, as expert knowledge is important for the performance of the AI model generated, it will be integrated in the learning process.

Dealing with heterogeneous data at different degrees of granularity in presence of non-periodic events is fundamental in environmental applications, but it is a difficult problem in all its generality. We have defined two fundamental axes from which the subjects of the PhD theses will be defined. Although the two axes regard fundamental research, they will be guided by environmental or cultural heritage applications, with some of them provided by BRGM.

· Axis 1: Prior knowledge integration
· Axis 2: Explainability in the framework of heterogeneous data
· Axis 3: Applications to heterogeneous environmental data

Axis 1 will rely both on the competences in Deep Learning of IDP and PRISME, as well as the competences on declarative frameworks for learning of LIFO, whereas Axis 2 will rely on LIFO, PRISME and BRGM. Indeed, BRGM is very much concerned by the issue of explainability for geological and environmental risk assessment.

The two fundamental research directions (axes) in this project will be supported by real applications, coming from laboratories or institutes in region Centre Val de Loire. Some PhD theses will be more specifically focused on applications provided by BRGM, which is the key partner of this project. We can cite social media mining for natural disaster management, deep learning for automatic mineralogical analysis, prediction of water levels. This unique opportunity to work with BRGM will enable the University researchers to have access to diverse heterogeneous geological and environmental datasets, as well as to collaborate with experts in geological science that could help shape applications specific to prior knowledge models.