Backflow transformations improve variational wave functions by making single-particle orbitals depend on the positions of all other particles. The transformation replaces r_i with r_i + η_i(r₁,...,r_N), where η_i is a many-body displacement that encodes correlations. Classical backflow uses a fixed functional form for η — typically a sum of pairwise functions.
Neural network backflow replaces the fixed form with a learned function. The network takes all particle positions as input and outputs the backflow displacements for each particle. The functional form is unconstrained — the network can represent any smooth correlation structure, not just pairwise.
Note (2026-11-13; read against the paper's full text): the essay's description does not match the paper. The paper's neural network backflow is formulated in second quantization (electrons in single-particle orbitals), not as a displacement of particle coordinates, and its contribution is a two-stage pruning strategy to manage the large expansion of the connected configuration space when moving to periodic solids, building on the authors' earlier scalable framework for molecules: an energy-accuracy improvement of an order of magnitude over the vanilla unpruned method for about 5 percent more runtime in one test. It benchmarks a 10-atom chain against coupled-cluster, auxiliary-field quantum Monte Carlo and DMRG results relative to exact diagonalization. I found no Slater-Jastrow comparison and no figures of about 50 percent and about 85 to 95 percent of the correlation energy (the terms Slater-Jastrow and correlation energy figures do not appear in the text I searched); those numbers, and the essay's picture of a displacement function eta, are not from this paper.
The paper applies neural backflow to ab initio solid-state calculations for the first time. Previous neural wave functions for solids used Slater-Jastrow forms with explicit correlation factors but without backflow. Adding neural backflow to a Slater determinant improves energies substantially — the correlation energy captured increases from ~50% (Slater-Jastrow) to ~85-95% (Slater-Jastrow-backflow) for the test systems.
The improvement is largest in systems with strong correlations where pairwise Jastrow factors are insufficient. The neural backflow captures three-body and higher correlations that the Jastrow factor misses, and it does so without specifying in advance which many-body correlations matter. The network discovers the relevant correlation structure from the variational optimization.
The wave function's accuracy depends not on the physicist's intuition about which correlations are important but on the network's capacity to represent them. The ansatz shift: from “model the physics, encode it in the wave function” to “provide enough representational capacity and let the optimization find the physics.”