Résumé :

This seminar will present an original approach to artificial intelligence application in geosciences, as proposed in the MaLISSiA project. Artificial intelligence and digital twins offer tremendous opportunities for modelling and understanding Earth systems. However, these new developments have mainly been applied to geophysical processes occurring within relatively well-constrained geometries (e.g., on earth surface) and they have yet to make a breakthrough in the characterization of sub-surface architectures. Yet these geometrical considerations are determinant in the localization of sub-surface Earth processes. The difficulty to access sub-surface causes paramount epistemic uncertainties, which result in peculiar scientific challenges, e.g., the dependencies to scale, to time, and a general lack of observation with respect to the complexity of structures. These limitations are generally counterbalanced by expert interpretations that rely on human learning from analogues (outcrops and simulations). The search for more formal and automated solutions is challenging both the formalization of geological knowledge and the development of innovative artificial intelligence methods. The MaLISSiA project draws its inspiration from the geocognitive process developed by human learning and interpretation. It proposes a new paradigm for automatically interpreting and explaining sub-surface architectures. This paradigm is based on two founding components: (1) a formalisation of geological and interpretation concepts in dedicated ontologies; (2) a Geocognitive Interpretation Process that breaks a complex spatial data assimilation problem into a series of explanation proposal and testing. Such formalism also enables the training of machine learning algorithms to answer key questions in the process. This training calls for the development of a corpus of interpreted geological references, based on both natural objects and process-based simulations that reproduce geological history.