Towards digital twin of an in-situ experiment: a physics-enhanced machine-learning framework for inverse modelling of mass transport processes

• A digital twin integrates in-situ data, physics models, ML, and inverse optimization. • Fast fitting enabled by high-fidelity 3D simulations in realistic geometries. • ML surrogates accelerate inverse modelling, with validated speed and accuracy gains. As a prove of concept for experimental geochemistry, an advanced 3D numerical framework, here and after called Digital Twin (DT), of a diffusion experiment conducted at a synchrotron beamline, has been implemented using in-situ measurements data, physics-based modelling, a machine learning (ML) model, and parameter optimization module. The physics-based model enables finely discretized high-resolution 3D mass transport simulations, which provide the training set for the ML model. The resulting ML model greatly accelerates the computationally intensive calculations needed for the interpretation of the experimental observations during inverse modelling. The framework is applied to interpret the in-situ non-destructive micro-X-ray fluorescence (μ-XRF) imaging data from a bromide diffusion experiment through a silica-gel-filled capillary system. The computational framework is refined, and several optimization algorithms are implemented to fit the experimental data. The gain in computational efficiency allows modelling the experiment practically in real-time.