Work

Approximating many-electron wave function with physics-aware surrogate models

Year: 2023
Type: dissertation
Abstract: With the ability to obtain sufficiently accurate and computationally accessible solutions to the many-body Schrödinger equation most of modern chemistry and material science could renounce phenomenology and enter the era of ab-initio modeling.Since its publication in 1926, the equation received a lot of attention with regards to finding an accurate, yet efficient approximation which would capture the most relevant quantum behavior while remaining polynomial in computational complexity.To this end, one often resorts to methods based on the mean-field approximation, in particular Hartree-Fock (HF) or the highly popular and flexible Kohn-Sham density-functional theory (DFT).Central to both methods lies the interpretation of the nonlinear eigenvalue problem constituted by the stationary Schrodinger equation for the ground state energy as variational minimization procedure.Both methods, in order to tackle that minimization analytically, restrict their computational basis to a linear combination of orthogonal one-electron wave functions only, which has a severe impact on the overall efficacy of those algorithms.A different family of electronic structure methods known as Quantum Monte Carlo (QMC), overcomes this limitation allowing for an arbitrarily complex wave function ansatz which in recent years has led to a surge in interest from deep learning community which lead to formulation of so called Neural Quantum States (NQS).The price one pays for this freedom, however, is the necessity of solving high-dimensional integrals in order to evaluate the energy expectation objective and a subsequent stochastic optimization procedure of the variational parameters.Given the flexibility of neural networks, that is, their inherent lack of any underlying discretization, other than, perhaps the machine precision, of crucial importance becomes performance of the Monte Carlo simulation, which determines which regions of the wave function get assigned most representational resources.In this work, we address the topic of sampling using physics-aware deep learning surrogates of the wave function Ψ within the Quantum Monte Carlo framework, in particular our aim was to develop a sampling algorithm which could deal with the pathologies of the so called Born probability density |Ψ| 2 .As a secondary endeavour, we have gone to great lengths in providing an extensive, although not exhaustive, review of the state-of-the-art methodology of this young field alongside the traditional, computational approaches of quantum chemistry and physics of fermionic systems published hitherto.Through this work, we wish to improve upon prevailing accuracy-efficiency dilemma in computational quantum mechanics and to further tighten the progressing integration of deep learning and scientific computing which promises significant improvements over the often purely phenomenological elements of the current state-of-the-art methods, closing the gap to the elusive problems of practical significance encountered in engineering, chemistry or medicine.iiiand potential U(r) energy terms which are themselves quantum operators.Hamiltonian mechanics will be dealt with in greater detail in later sections with regards to Hamiltonian Monte Carlo, an efficient sampling technique used as a subroutine in the Quantum Monte Carlo method.For now, it is only important to realize it is a system-dependent, hermitian operator which generates its evolution in time.With no reference to the underlying physics, the wave function is supposed to represent, one can further state that equation 1.1 [4]:
Author: Dawid Pasterny
Language: English
Open Access status: green
OpenAlex ID: W7036178233
Cites: 102
Cited by:
Related to: 10