Natural Quantum Monte Carlo Computation of Excited States

David Pfau, Simon Axelrod, Halvard Sutterud, Ingrid von Glehn, James S. SpencerarXiv 2023 · arXiv · OpenReview

TL;DR

This paper introduces "Natural Excited States for VMC" (NES-VMC), a novel variational Monte Carlo method that accurately computes quantum excited states. By reformulating the problem as a ground-state calculation on an expanded system, NES-VMC overcomes limitations of previous methods and achieves state-of-the-art accuracy when combined with neural network wavefunctions.

What Problem Does This Paper Address?

Calculating the properties of quantum systems in their excited states is notoriously difficult. Unlike ground states, where energy minimization is straightforward, excited states require more complex approaches. Traditional methods like TD-DFT often yield qualitatively incorrect results, while highly accurate methods like MRCI can be computationally prohibitive and struggle with higher excited states. Existing variational Monte Carlo (VMC) methods for excited states often rely on penalty terms or explicit orthogonalization, which can introduce biases, lead to optimization issues, or are incompatible with modern, flexible neural network wavefunctions. This paper aims to bridge this gap by providing a general, parameter-free VMC method applicable to any quantum Hamiltonian, particularly excelling with neural network Ansätze.

Methods

The core innovation is the Natural Excited States for VMC (NES-VMC) method. Instead of directly optimizing individual excited states, NES-VMC constructs a multi-particle wavefunction $\Psi$ as a determinant of single-state Ansätze ($\psi_1, ..., \psi_K$). This determinantal form, inspired by Slater determinants but using many-body wavefunctions as "orbitals," naturally prevents states from collapsing onto each other without explicit orthogonalization. The problem is then transformed into finding the ground state of an "expanded system," where the local energy is a matrix rather than a scalar. Minimizing the trace of this matrix provides an unbiased estimate of the total energy of the lowest $K$ states. Crucially, this formulation allows for standard VMC optimization techniques (like stochastic gradient descent) to be applied directly, ensuring convergence and unbiased gradient estimates. The method is demonstrated with two advanced neural network Ansätze: FermiNet and Psiformer.

Datasets

The paper evaluates NES-VMC on several benchmark datasets:

  • Atomic Spectra: Experimental data from the NIST Handbook of Basic Atomic Spectroscopic Data was used for first-row atoms (Li to Ne).
  • Molecular Systems: The QUEST database was used for vertical excitation energies and oscillator strengths of small molecules.
  • Carbon Dimer (C2): This system was studied to analyze potential energy curves and natural orbital occupancies.
  • Ethylene (C2H4): Used to study excited states across torsional and pyramidalization angles, including conical intersections.
  • Double Excitations: Several systems known for challenging double excitations (Nitrosomethane, Butadiene, Glyoxal, Tetrazine, Cyclopentadienone) were investigated.
  • Benzene: A common benchmark for medium-sized molecules.

No new datasets were introduced; instead, existing, well-established benchmarks were used for rigorous comparison.

Key Insights and Ideas

  • Determinantal Ansatz for Excited States: The key insight is using a determinant of many-body wavefunctions ($\Psi = \det(\Psi)$) as the total Ansatz. This elegantly handles multiple states without explicit orthogonalization, a major hurdle in previous VMC excited-state methods.
  • "Expanded System" Equivalence: The method cleverly reformulates the $K$-excited-state problem into a single ground-state problem on a larger, artificial system. This allows leveraging the well-established VMC machinery for ground states.
  • Unbiased Gradients and Parameter-Free Optimization: NES-VMC provides unbiased estimates of energies and gradients, and crucially, has no free parameters to tune (unlike penalty methods), simplifying its application.
  • Neural Networks as Powerful Ansätze: The paper strongly demonstrates that combining NES-VMC with expressive neural network Ansätze (FermiNet, Psiformer) unlocks unprecedented accuracy for excited states, even for challenging double excitations and systems up to benzene size.

Key Results

  • Accurate Excitation Energies: NES-VMC achieves excellent agreement with experimental and theoretical best estimates (TBEs) for vertical excitation energies across atoms and molecules, often within chemical accuracy (1.6 mHa or 43 meV).
  • Accurate Oscillator Strengths: Oscillator strengths, which are sensitive quantities, are also computed with high accuracy, generally agreeing closely with TBEs.
  • Challenging Double Excitations: The method successfully computes energies for challenging double excitations, outperforming previous neural network approaches.
  • State-of-the-Art for Neural Networks: NES-VMC is presented as the first deep learning approach to achieve accurate vertical excitation energies on benzene-scale molecules, and the only effective method for neural network Ansätze on larger systems.
  • Improved Potential Energy Curves: For C2, NES-VMC accurately reconstructs potential energy curves, showing improved agreement with experiment compared to previous methods like SHCI.
  • Scalability: The method scales quadratically with the number of excited states ($K$) for typical computations, and asymptotically cubically for the determinant calculation.

Limitations and Future Work

  • Numerical Stability: While generally robust, some numerical stability issues were encountered, particularly with larger systems or specific pretraining strategies. Heuristics were employed to mitigate these.
  • Pretraining Sensitivity: The method's performance can be sensitive to pretraining strategies, especially for systems with challenging double excitations. Random pretraining was found to be more robust in some cases.
  • Computational Cost: While more efficient than some traditional methods for excited states, it still involves significant computational cost, especially for larger systems and more excited states.
  • Future Work: The authors suggest applying NES-VMC with classic Ansätze (like Slater-Jastrow) for larger systems, investigating vibronic couplings, and extending its application to nuclear physics, condensed matter physics, and spin systems. They also highlight the potential for combining NES-VMC with deep neural networks for more complex quantum phenomena.

Summary updated

Abstract

We present a variational Monte Carlo algorithm for estimating the lowest excited states of a quantum system which is a natural generalization of the estimation of ground states. The method has no free parameters and requires no explicit orthogonalization of the different states, instead transforming the problem of finding excited states of a given system into that of finding the ground state of an expanded system. Expected values of arbitrary observables can be calculated, including off-diagonal expectations between different states such as the transition dipole moment. Although the method is entirely general, it works particularly well in conjunction with recent work on using neural networks as variational Ans\"atze for many-electron systems, and we show that by combining this method with the FermiNet and Psiformer Ans\"atze we can accurately recover vertical excitation energies and oscillator strengths on a range of molecules. Our method is the first deep learning approach to achieve accurate vertical excitation energies, including challenging double excitations, on benzene-scale molecules. Beyond the chemistry examples here, we expect this technique will be of great interest for applications to atomic, nuclear and condensed matter physics.
BibTeX
@misc{pfau2023natural,
  author = {David Pfau and Simon Axelrod and Halvard Sutterud and Ingrid von Glehn and James S. Spencer},
  title = {{Natural Quantum Monte Carlo Computation of Excited States}},
  year = {2023},
  eprint = {2308.16848},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2308.16848},
}

Loading related content...

Figure