omnibias-ferminet

FermiNet bridge for the omnibias closed-form n-th derivative framework.

Install

pip install omnibias-ferminet
# Optional: pull in folx for the bit-stable Laplacian comparison test
pip install omnibias-ferminet[folx]

omnibias-ferminet depends on omnibias-core, omnibias-jax, and jax>=0.4.30. It does not require a FermiNet installation to import or to be unit-tested (mock log|psi| factories are provided for the bridge contracts); a real FermiNet checkout is needed only at production VMC time.

What is in here

• omnibias.ferminet.folx_compat: forward_laplacian, closed_form_forward_laplacian, laplacian_factory -- folx-API drop-ins so any FermiNet / DeepQMC code that imports from folx can switch to omnibias's closed-form Laplacian by changing one import.
• omnibias.ferminet.integration: the production bridge -- envelope value / gradient / Hessian, optional one-body backflow, and the make_omnibias_envelope_local_kinetic_energy and make_omnibias_tier2_local_kinetic_energy factories that the upstream FermiNet laplacian_method switch consumes.
• omnibias.ferminet.restricted: Tier-2 / Tier-2-full restricted FermiNet ansatz with end-to-end closed-form Laplacian (used to bring the bridge up before plugging into the upstream FermiNet checkpoint). tier2_grad_laplacian_log_psi exposes the gradient and Laplacian of log|det M| separately (a bit-identical refactor of the kinetic path) so an additive correlation factor composes correctly.
• omnibias.ferminet.jastrow: closed-form symmetric Padé-Jastrow correlation factor exp(J) -- analytic value / gradient / Laplacian (jastrow_value_grad_laplacian) with physical e-e (1/2, 1/4) and e-n (-Z) cusp slopes, plus jastrow_slater_local_kinetic_energy combining it with a Slater determinant (correct cross term in |grad log|psi||^2).
• omnibias.ferminet.multiblock and omnibias.ferminet.multiblock_integration: multi-block FermiNet primitives for Born-Oppenheimer / nuclear-Hessian work.

Public API (lazy)

from omnibias.ferminet.folx_compat import (
    forward_laplacian, closed_form_forward_laplacian, laplacian_factory,
    OmnibiasFwdLaplResult,
)
from omnibias.ferminet.integration import (
    envelope_value_grad_hessian,
    apply_optional_backflow,
    make_omnibias_envelope_local_kinetic_energy,
    make_omnibias_tier2_local_kinetic_energy,
)
from omnibias.ferminet.restricted import (
    Tier2Params, Tier2SymParams,
    tier2_log_abs_psi, tier2sym_log_abs_psi,
    tier2_local_kinetic_energy, tier2sym_local_kinetic_energy,
)

The top-level omnibias.ferminet namespace deliberately does not re-export these: importing the bridge is opt-in so the package stays cheap to import in scripts that do not need FermiNet.

Contract

omnibias-ferminet produces a Laplacian that is bit-identical to FermiNet's default autograd path on the closed-shell systems we have validated (Be 4-electron 4-walker test in the private release-validation archive row R1: rel_err = 0.0 to ULP, vs folx <= 5.07e-15). This is a bit-stable contract for the CPU deterministic forward path; on stochastic GPU runs we hold to a 1-sigma reproducibility budget.

License

Apache-2.0. See LICENSE and ../../LICENSING.md. You never need a commercial licence for this package.