Neural Quantum States for Nuclear Magnetic Resonance Spectroscopy

Abstract

Predicting a nuclear magnetic resonance (NMR) spectrum from first principles requires propagating a quantum state of dimension $2^N$ for $N$ coupled spins, which becomes intractable beyond larger $N$. We benchmark Neural Quantum States (NQS), a class of variational quantum states expressed as an artificial neural network, as an alternative representation for this problem. Using two propagation methods, the Time-Dependent Variational Principle (TDVP) and projected time-dependent Variational Monte Carlo (p-tVMC), we compute the $^1$H spectra of four ($2 \to 5$ spins) experimentally parameterized molecules. TDVP reproduces all line positions and intensities with average spectral mean squared errors of $<10^{-3}$; p-tVMC reproduces the same features, with accuracy determined by its per-step optimization parameters. One dominant obstacle to larger systems is the steep growth of the number of integration steps with spectral bandwidth, which can be removed by propagating in the interaction frame of the chemical-shifted Hamiltonian. Retaining the same accuracy, this reduces the number of integration steps roughly eightfold for the 3-spin system and by at least an order of magnitude for the 4- and 5-spin systems, and it enables a 14-spin molecule (sucrose) to be accurately propagated via Monte Carlo sampling.