Purpose A deep geological repository for radioactive waste, such as Andra’s Cigéo project, necessitates long-term monitoring. This monitoring is achieved by collecting data from various sensors. However, due to environmental conditions (radioactivity and mechanical constraints), this set of sensors is prone to deterioration over time. Therefore, it is essential to replace the responses of faulty sensors with comprehensive predictions. Graph neural networks (GNNs) are appropriate models for these predictions, as they efficiently characterize the physical phenomena present in the system, leverage the underlying topology of the data and can be used to infer general dependencies. The purpose of this paper is to study the effectiveness of GNNs for this temperature interpolation task. Design/methodology/approach In this paper, the authors trained several types of GNNs for temperature forecasting using experimental data from Andra’s Underground Research Laboratory. The specific experiment used to train the machine learning algorithms simulates the heating of a high-level waste (HLW) demonstrator cell by radioactive waste within a deep geological layer. The model the authors used is a forward-integrating GNN that takes the initial temperature and boundary conditions as input and outputs the temperature field at all future time steps. Findings By comparing GNNs to other machine learning algorithms (Gaussian processes, artificial neural networks and kriging), the authors proved their effectiveness for data completion. Originality/value This work is original as it proposes the use of time-integrating GNNs for data completion through transfer learning, using data collected from an industrial demonstrator that simulates the heating of a HLW cell by radioactive waste.
Hembert et al. (Tue,) studied this question.