Author(s) :

Mia Moser-Gitter, David Yaron.

Volume/Issue :

Abstract :

Neural network potentials (NNPs) have achieved impressive accuracy on local, short-range chemical environments, yet their ability to capture long-range electronic effects remains unclear. We investigate this question using a controlled testbed for electron delocalization: 1D tight-binding chains with 2–10 atoms. We generated synthetic datasets by drawing nearest-neighbor coupling values from a narrow normal distribution and determining the resulting ground-state energies using Hamiltonian diagonalization. We benchmark two architectures: (1) a “full chain” feed-forward MLP that ingests the entire coupling vector and directly predicts the total energy, and (2) a Behler–Parrinello–style “atom environment” model that sums per-atom energies from local windows (sizes 1–3). Across chain lengths, the full-chain model consistently attains lower error and cleaner predicted-vs-true alignments, though accuracy degrades with system size for all models. For 10-atom chains, mean errors are 0.0279 (full-chain) versus 0.0613, 0.0524, and 0.0417 for window sizes 1, 2, and 3, respectively. Loss curves show stable optimization with limited overfitting on small and mid-sized systems. These results indicate that even simple global-feature MLPs can learn delocalization effects that challenge purely local descriptors. Key limitations include the idealized 1D tight-binding ground truth, nearest-neighbor couplings only, and the absence of symmetry/equivariance constraints, which limit transferability to realistic 3D chemistry. Nonetheless, the findings motivate combining global receptive fields (or longer range descriptors) with physics-aware architectures to improve generalization in delocalized electronic systems.