Abstract
$n$-particle reduced density matrices ($n$-RDMs) play a central role inunderstanding correlated phases of matter. Yet the calculation of $n$-RDMs isoften computationally inefficient for strongly-correlated states, particularlywhen the system sizes are large. In this work, we propose to use neural network(NN) architectures to accelerate the calculation of, and even predict, the$n$-RDMs for large-size systems. The underlying intuition is that $n$-RDMs areoften smooth functions over the Brillouin zone (BZ) (certainly true for gappedstates) and are thus interpolable, allowing NNs trained on small-size $n$-RDMsto predict large-size ones. Building on this intuition, we devise two NNs: (i)a self-attention NN that maps random RDMs to physical ones, and (ii) aSinusoidal Representation Network (SIREN) that directly maps momentum-spacecoordinates to RDM values. We test the NNs in three 2D models: the pair-paircorrelation functions of the Richardson model of superconductivity, thetranslationally-invariant 1-RDM in a four-band model with short-rangerepulsion, and the translation-breaking 1-RDM in the half-filled Hubbard model.We find that a SIREN trained on a $6\times 6$ momentum mesh can predict the$18\times 18$ pair-pair correlation function with a relative accuracy of$0.839$. The NNs trained on $6\times 6 \sim 8\times 8$ meshes can providehigh-quality initial guesses for $50\times 50$ translation-invariantHartree-Fock (HF) and $30\times 30$ fully translation-breaking-allowed HF,reducing the number of iterations required for convergence by up to $91.63\%$and $92.78\%$, respectively, compared to random initializations. Our resultsillustrate the potential of using NN-based methods for interpolable $n$-RDMs,which might open a new avenue for future research on strongly correlatedphases.