omnibias.ferminet
The FermiNet bridge: wire omnibias's closed-form Laplacian into neural
variational Monte Carlo as a drop-in for the default local kinetic energy.
Install with pip install omnibias-ferminet.
import omnibias.ferminet
Importing omnibias.ferminet is cheap — it does not import JAX or a
FermiNet checkout. The submodules below import JAX on demand, so the top-level
__all__ is intentionally minimal (__version__). Import the submodule you
need explicitly.
Public surfaces
omnibias.ferminet.folx_compat
folx-compatible adapters — a direct replacement for the folx Laplacian in
FermiNet's laplacian_method switch.
| Symbol | Role |
|---|---|
forward_laplacian | folx-signature forward Laplacian |
closed_form_forward_laplacian | the closed-form implementation |
laplacian_factory | build a Laplacian callable |
omnibias.ferminet.integration
The production bridge: envelope value / gradient / Hessian kernels, optional one-body backflow, and the factories the upstream FermiNet branches consume:
make_omnibias_envelope_local_kinetic_energy→laplacian_method == "omnibias_envelope"make_omnibias_tier2_local_kinetic_energy→laplacian_method == "omnibias_tier2"
omnibias.ferminet.restricted
The Tier-2 / Tier-2-full restricted FermiNet ansatz with a closed-form Laplacian — an end-to-end omnibias-only path.
omnibias.ferminet.multiblock / multiblock_integration
Multi-block FermiNet primitives (per-geometry blocks for nuclear-Hessian work)
and their composition into a FermiNet-shaped log|ψ| / local kinetic energy.
Example
from omnibias.ferminet.integration import (
make_omnibias_envelope_local_kinetic_energy,
)
local_kinetic = make_omnibias_envelope_local_kinetic_energy(network=ansatz)
E_kin = local_kinetic(params, walkers) # bit-identical to the default method