Deep learning-enhanced physical modelling for tape-casting slurry microstructures of solid oxide cell substrates

Solid oxide cells are manufactured by tape casting, in which a ceramic slurry is spread into thin sheets whose microstructure shapes the properties of the finished cell. The discrete element method simulates how that microstructure forms, but its computational cost makes broad exploration of processing options impractical. This practice couples that physics-based simulation with deep learning: a convolutional neural network, adapted from the standard VGG16 image-recognition architecture to a regression task, takes the early frames of a simulation and predicts particle positions at equilibrium, bypassing the long intermediate computation. The predicted structure is then passed back into a short physics-based relaxation step, which removes unphysical artefacts introduced at that transition. The output is a three-dimensional slurry microstructure for a defined NiO/8YSZ/binder formulation, obtained without running the full simulation and assessed against it by density, volume and the radial distribution of particles. AI does not replace the physical model; it accelerates its slowest stage, with physics retained at both ends of the workflow.

AI was used as a tool to simplify and speed up practical industrial processes in process engineering, and to ensure consistent product quality. The work originated in a funding call, and a joint scientific project on solid oxide cells was the nucleus for the cooperative work. The approach was transferred from the group’s earlier work on lithium-ion battery electrodes under the ARTISTIC framework, where the same coupling of particle simulation and neural-network prediction had been developed; the shared simulation configuration allowed the method to move to solid oxide cells with little redesign. A mixed team from computing science, data science, materials science, chemistry and mathematics worked together, and all co-workers are experienced AI developers and/or AI users. Training data was self-generated rather than collected: fifteen simulations were run with LAMMPS on the MatriCS high-performance computing platform at Université de Picardie Jules Verne, eight used for training and validation and the remainder held back for testing, together yielding roughly 130,000 particle-level examples. The physical simulation thus supplies its own ground truth, so no experimental dataset was needed to train the model – though the training data inherits whatever assumptions the underlying physical model carries. The AI models are self-developed in Python, adapting the VGG16 architecture to a regression task, and particle positions were represented in unwrapped coordinates to avoid discontinuities at the simulation boundaries — a problem the group had identified in its earlier battery work. Implementation was in TensorFlow/Keras, using CPU and GPU servers and workstations. The collaboration spans a French materials laboratory and a German research centre, supported by regional, national and European funding across several instruments.

The hybrid surrogate framework enables prediction of tape-cast slurry microstructures for solid oxide cell fuel-electrode substrates without running full-scale simulations, supporting faster validation and early-stage optimisation of manufacturing processes. Calculation time for the whole slurry simulation process was reduced from 1230 to 30 minutes, with most of the remainder spent on the physics-based relaxation rather than the network. Agreement with the full simulation was close, with coefficients of determination between 0.995 and 0.998 for particle coordinates across unseen test cases, and radial distribution functions aligning well at medium and long range. Accuracy is not uniform: predictions degrade as structures grow larger and more complex, the most complex case showing a density deviation of about 24%, and short-range structural detail is reproduced less faithfully than larger-scale arrangement. The work is presented as a proof of concept, restricted to a single formulation and force-field set, and produces single predictions without an accompanying estimate of uncertainty; probabilistic approaches are identified as the next step. Results have been cross-checked between the participating groups and industry partners. The article is open access under CC BY, with force-field settings and simulation parameters published alongside it, while the underlying simulation data is available from the authors on request.

Le-Dinh, Tan, Utkarsh Vijay, Diego E. Galvez-Aranda, Hartmut Schlenz, Norbert H. Menzler, Olivier Guillon, and Alejandro A. Franco. “Deep learning-enhanced physical modelling for tape-casting slurry microstructures of solid oxide cell substrates.” Journal of Power Sources 688 (2026): 240447, https://doi.org/10.1016/j.jpowsour.2026.240447