Machine learning-assisted heat transport modelling for full-scale emplacement experiment at Mont Terri underground laboratory

Machine learning (ML)-assisted modelling of deep geological repositories (DGR) is of emerging interest and can help to improve the safe and reliable operation of DGRs as well as the public acceptance. Here, a concept of ML-assisted physical-based 3D heat transport model for the Full-scale Emplacement (FE) experiment performed at Mont Terri Underground Laboratory (URL), is presented. The FE experiment is a 1:1 scale mock-up of a DGR tunnel where heaters simulate emplaced high level waste. ML is applied to sparse sensor data of the water saturation degree in the granulated bentonite material (GBM) yielding a physically and neural network (NN) based surrogate model for the thermal conductivity of the GBM, needed to calculate the temperature evolution in the FE tunnel near-field. In order to investigate the dominant parameters influencing the temperature evolution in the vicinity of the FE experiment, the results of 32 orthogonal test cases have been analysed systematically. For the ML predicted water saturation degree, three NN methods are tested. The Elman NN (with a Pearson’s r coefficient of 0.9911, a mean squared error (MSE) of 4.62, and mean absolute error (MAE) of 1.44) operates better than the back propagation (BP) and cascade BP (CBP) NN methods. Results of ML-assisted heat transport calculations are validated with the large experimental dataset of 137 × 10 temperature sensor data points (errors range within 7%). Parameters uncertainty ranges of ±10% for thermal conductivities of the GBM λGBM and bentonite block λblock are analysed and bands of temperature uncertainties are compared with temperature sensor data, which allows sensor data assessment including identification of the faulty sensor data. The ML-assisted physical modelling framework can be applied to future DGRs.