DeepH-Polar brings distant electric fields into an AI materials model

DeepH-Polar, a 27 October preprint, addresses a blind spot in AI materials calculations: similar atomic neighbourhoods can experience different electric fields. It combines learned local electronic terms, atomic dipoles and polarizabilities with self-consistent Poisson electrostatics. Dipoles describe separated charge; polarizability determines how those dipoles respond to an electric field.[1]

The object being learned is a Hamiltonian: here, a matrix used to calculate electronic energies and states. The earlier DeepH-E3 model, published in 2023, learns this representation from atomic structures. Its architecture preserves the required transformations under rotation, so changing the coordinate system respects the same underlying physics. Predicting the Hamiltonian also provides a route to several electronic properties, rather than requiring a separate prediction of every quantity of interest.[2]

The electrical problem is easy to see in a silicon device. A gate can sit on the other side of an oxide layer. Changing its voltage redistributes charge in the semiconductor, shifting the balance between depleted regions and layers rich in mobile carriers. The atoms need not be rearranged for those electrical conditions to change. This is familiar device physics: applied voltage is an essential part of the description.[3]

Poisson’s equation links the charge distribution to the electrical potential. The charge distribution, in turn, depends on that potential. A self-consistent calculation seeks values that satisfy both relationships together. Treating only one side as fixed can miss the response of the other. This explains why knowing the local geometry is a different task from determining the electrical state of the whole device.[3]

The new model was tested against density-functional-theory calculations, including one held-out silicon/oxide structure under a strong field. A separate 8,194-atom device simulation used zero source–drain voltage; measured transistor performance was not tested.[1]

A bandgap is an energy interval without allowed electronic states. A metal has available states immediately above its highest occupied levels.[4]

One model-design lesson follows. If two physical situations are encoded as identical inputs, a deterministic predictor cannot distinguish them. Training can refine its answer, but cannot tell it which of those situations it has been given. Choosing the inputs is therefore part of choosing which physics a calculation can represent.[2]

Report an error or suggest a correction