Graph-Theoretic Neural Network Fragmentation for Coupled-Cluster Accuracy AIMD

A new machine learning framework integrates graph-theoretic molecular fragmentation with direct force prediction to enable coupled-cluster accuracy ab initio molecular dynamics (AIMD) for fluxional systems. The approach bypasses automatic differentiation on learned energy surfaces by directly modeling nuclear force vectors, using covariant descriptors projected onto fragment-fixed principal axes of inertia to ensure rotational, translational, and permutational invariance. This vector-valued training protocol reduces trainable parameters by over an order of magnitude, achieving high parameter efficiency. The method addresses the computational bottleneck of correlated electronic structure methods, allowing accurate AIMD simulations of complex chemical systems.