EAIFR Ph.D. candidate Sunday J. Ogenyi and collaborators have demonstrated a highly efficient strategy for developing accurate machine-learning interatomic potentials (MLIPs) for pure silica zeolites, significantly reducing the computational cost traditionally associated with large-scale atomistic simulations.

In a recent study published in PCCP, the team addressed a key challenge in machine-learning–driven molecular dynamics: the need for extensive and expensive training datasets generated using density functional theory (DFT). While MLIPs can reproduce DFT-level accuracy at a fraction of the computational cost, their development typically requires vast amounts of ab initio data.

To overcome this limitation, the researchers introduced a hierarchical clustering approach tailored to crystal structures with similar chemical compositions but distinct topologies, such as pure silica zeolites. This strategy enabled the identification of a compact set of representative polymorphs, forming the basis of an efficient and targeted training dataset.

The dataset was constructed through an active learning workflow that employed a computationally lightweight invariant neural network potential. Using the ab initio data generated during this phase, the team developed two high-fidelity MLIPs based on invariant and equivariant neural network architectures, respectively.

Both models achieved remarkable accuracy, with energy errors of ≤3 meV per atom compared to DFT results. Beyond energetic accuracy, the neural network potentials successfully reproduced a wide range of material properties—including structural, mechanical, thermal, and phonon dispersion characteristics—showing excellent agreement with first-principles calculations.

This work demonstrates that carefully designed active learning and data-reduction strategies can deliver DFT-level accuracy with optimal computational effort. It opens the door to longer timescale and larger system simulations of complex zeolite materials, offering a promising pathway for accelerating materials discovery across a broad range of crystalline systems.

Source:

Ogenyi, S.J., Shaidu, Y., Seriani, N. and Ojo, O.A., 2026. Efficient Development of a Neural Network Potential for Pure Silica Zeolites. Physical Chemistry Chemical Physics.