ML molecular dynamics models phase separation in Si-C-N ceramics

The arXiv preprint 2605.20358 by Fabien Mortier et al. presents a machine-learning interatomic potential for silicon carbonitride (Si-C-N-H) systems and applies it to large-scale molecular dynamics. The authors report training the potential on a diversified database of over 9000 configurations, per the preprint, including amorphous models, high-temperature states, surfaces, and crystal predictions. Using this potential, the paper presents simulations of 8000-atom systems that show progressive phase separation: carbon domains nucleate from an amorphous SiCN matrix and form graphene-like sheets while the ceramic network remains intact, according to the preprint. The models are reported to reproduce experimental atomic pair distribution functions with high fidelity, and the authors describe defect-mediated ring transformations that convert 5- and 7-member carbon rings into stable 6-member aromatic structures. Editorial analysis: this methodology extends first-principles accuracy to experimentally relevant length scales for polymer-derived ceramics.
What happened
The arXiv preprint 2605.20358, submitted 19 May 2026 by Fabien Mortier and coauthors, introduces a machine-learning interatomic potential targeted at silicon carbonitride (Si-C-N-H) systems. Per the preprint, the authors trained the potential on a diversified database of over 9000 configurations, which includes amorphous structures, high-temperature snapshots, surfaces, and crystal-structure predictions. The paper reports large-scale molecular dynamics simulations of 8000-atom systems that reveal atomic-scale thermal evolution during polymer-derived ceramic processing. The simulations show progressive phase separation in which carbon domains nucleate from an amorphous SiCN matrix and assemble into graphene-like sheets while the ceramic network remains, per the preprint. The authors report that their models reproduce experimental atomic pair distribution functions with exceptional fidelity and describe defect-mediated transformations where 5- and 7-member carbon rings convert into stable 6-member aromatic rings.
Technical details
Editorial analysis - technical context: The work centers on a data-driven interatomic potential trained to emulate first-principles energetics across a wide configuration space. For practitioners, this pattern follows recent trends where high-quality training sets and flexible ML potentials enable atomistic simulations at scales (thousands of atoms) and temperatures that were previously costly for direct density functional theory (DFT) sampling. The paper's emphasis on amorphous-to-graphitic transitions mirrors other ML-MD studies that combine structural diversity in training data with long-timescale MD to capture nucleation and phase separation phenomena.
Context and significance
Editorial analysis: Extending accurate interatomic potentials to complex, multicomponent amorphous materials lowers the barrier for mechanistic interpretation of experimental observables such as pair distribution functions and Raman signatures. For the materials-simulation community, validated ML potentials for polymer-derived ceramics provide a route to test processing-structure-property hypotheses at near-experimental scales without the full cost of DFT for every timestep.
What to watch
For practitioners: look for code or potential release and benchmarking details that enable reproduction and reuse. Also watch for follow-up work validating kinetic rates, larger system sizes, or coupling to mesoscale models for property prediction. Reproducibility hinges on published training datasets and hyperparameters, so availability of those artifacts will determine practical adoption.
Scoring Rationale
The paper advances ML interatomic potentials for complex multicomponent amorphous materials and demonstrates MD at **8000-atom** scale with experimental validation, which is notable for computational materials practitioners. It is a domain-specific advance rather than a cross-field paradigm shift.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems

