Graph-Theoretic Neural Network Fragmentation with Covariant Direct Molecular Force Learning: Enabling Coupled-Cluster Accuracy AIMD for Fluxional Systems

The paper proposes a graph-theoretic molecular fragmentation framework that learns post-Hartree–Fock nuclear forces directly, targeting coupled-cluster accuracy without differentiating a learned energy surface. Force vectors are projected onto fragment-fixed principal inertia axes to construct covariant representations while respecting rotational, translational, and permutational symmetries. The authors report over an order-of-magnitude reduction in trainable parameters and use mini-batch k-means selection to retain only 10–20% of reference configurations. Validation is reported for the fluxional solvated Zundel cation, H13O6+, using structural distributions and velocity-autocorrelation spectra.