To ensure the resistance of materials subjected to severe loadings for protection purposes, materials exhibiting suitable properties under such conditions are identified and then tested in dedicated experiments. Experimental data are used to calibrate material models implemented in fast dynamics simulation codes. A major difficulty arises from the fact that these data are often affected by uncertainties, making the identification of a single parameter set questionable. The objective is therefore to perform a calibration that is robust to these uncertainties. Bayesian calibration provides a rigorous mathematical framework to address this issue: uncertainties can be incorporated into the likelihood, while expert knowledge and physical constraints can be embedded in the prior distribution for a given model. The unknown is no longer a single parameter set that best fits the data, but a distribution over parameters that explains the data and takes into account uncertainties. However, estimating a full probability density introduces significant technical, particularly computational, challenges. Sampling-based methods are required to extract the statistical information contained in the distribution. These sampling algorithms are sequential and typically require a large number of model evaluations. When the model relies on a computationally expensive simulation code, many standard algorithms become impractical, especially in high-dimensional parameter spaces. In our applications, models involve between 10 and 20 parameters, some of which correspond to entire curves, making the problem intrinsically high-dimensional. In this context, Hamiltonian Monte Carlo (HMC) is one of the most efficient algorithms, particularly with respect to scalability in dimension [2], but it requires access to the gradients with respect to the parameters of the model. To address these constraints, we have developed a methodology that: (1) constructs a large, high-dimensional database of model responses in the experimental configuration of interest by sampling the calibration parameters within prior-defined ranges, relying on intensive parallel computations on a supercomputer; (2) trains a neural network with controlled cost and performance on this database to emulate the model, with built-in access to gradients and exploitation of vectorized acceleration; (3) integrates the neural network emulator into an HMC–NUTS [3] framework to efficiently sample the target posterior distribution.