Variational Monte Carlo with neural-network wavefunctions (FermiNet, DeepQMC and friends) spends a large fraction of its compute in one place: the local kinetic energy, which is a Laplacian of the log-wavefunction evaluated at every walker, every step.
The bottleneck
The textbook way to get that Laplacian is autodiff - either a dense Hessian trace or a forward-Laplacian. Both are correct. The dense path is quadratic in the electron-coordinate dimension and dominates memory; the forward-Laplacian is far better but still re-pays a structured pass per evaluation.
A drop-in adapter
Derivon ships a folx-compatible forward_laplacian adapter. On closed-shell systems it is bit-identical to FermiNet's default autograd Laplacian - we verified rel-err = 0.0 to ULP on a Beryllium 8-determinant body, and ≤ 5.07e-15 versus folx.
from omnibias.ferminet.folx_compat import forward_laplacian
from omnibias.ferminet.integration import (
make_omnibias_envelope_local_kinetic_energy,
)
local_kinetic = make_omnibias_envelope_local_kinetic_energy(...)
Because the envelope's derivative tower is closed form, the kinetic-energy primitive is one forward evaluation regardless of order. The relativistic mass-velocity corrections Δᵏψ - which need iterated Laplacians - are where this matters most: they stay flat in k instead of exploding.
Why "bit-identical" is the headline
Swapping a numerical kernel in the inner loop of a physics calculation is scary. If the new kernel gives subtly different numbers, you cannot tell a speedup from a regression. "Bit-identical to the method you trust" removes that fear: the energy curve is the same, the optimizer sees the same landscape, and the only thing that changes is the wall-clock and the memory ceiling.
What we are working on next
Threading the closed-form path through deeper ansätze via the multivariate jet machinery, and publishing a reproducible GPU benchmark harness for the full (value, grad, Laplacian) envelope across electron counts.