Unifying neural-network architecture for interacting electron systems

Determine whether there exists a single neural-network wavefunction architecture that can be applied, without system-specific tailoring, to accurately and efficiently approximate ground states across a wide range of interacting electron systems, thereby serving as a unified variational ansatz for correlated electrons.

Background

The paper surveys recent advances in neural-network variational Monte Carlo for many-electron systems and introduces a self-attention-based wavefunction ansatz that performs well on moiré semiconductors. The authors note that various neural architectures have been proposed across different problems (atoms, molecules, lattice models, electron gas, moiré materials, fractional quantum Hall liquids).

Despite encouraging evidence that self-attention may be broadly effective, the authors explicitly raise the question of whether a single, general-purpose neural architecture can unify these domains. This problem seeks a definitive determination of the existence and scope of such a unifying architecture.