Quantum Monte Carlo: DMC, AFQMC and neural-network wavefunctions
Stochastic solutions of the electronic Schrödinger equation: variational and diffusion Monte Carlo, auxiliary-field QMC and neural-network wavefunctions, limited by the fermion sign problem.
QMC provides benchmark energies for solids, noncovalent complexes and small molecules independent of CC and DFT, and neural-network VMC is a fast-moving route to systematically improvable wavefunctions.
Solves the electron problem by random sampling, giving a method independent of coupled cluster and DFT.
- 1Guess a quantum state
- 2Sample electron positions at random
- 3Average the energy
- 4Improve the guess
- 5Fixed nodes keep noise manageable
Variational Monte Carlo (VMC). For a trial wavefunction , the energy is estimated by sampling electron positions from and averaging the local energy . By the variational principle, the energy is never below the truth; parameters are optimised to lower it. Statistical error shrinks as with samples.
Go deeper
The full technical text.
What it is
QMC methods estimate integrals over the many-electron wavefunction by random sampling, giving polynomial cost (roughly – per sample) and trivially parallel scaling.
- Variational MC (VMC) minimises by sampling .
Words on this page (7)
- quantum Monte Carlo
- Methods that solve the electron problem by random sampling, trading exact answers for statistical error bars.
- fixed-node approximation
- A trick in diffusion Monte Carlo that avoids the sign problem by fixing where the wavefunction changes sign, at the price of a bias.
- variational principle
- The rule that any trial wavefunction gives an energy at or above the true ground-state energy, so lower is better.
- sign problem
- The cancellation of positive and negative contributions that makes exact random-sampling calculations of electrons exponentially noisy.
- wavefunction
- The mathematical object that holds everything quantum mechanics can say about a system, such as where its electrons are likely to be.
- neural network
- A flexible function with many adjustable numbers, fitted to data by training.
- electron correlation
- The way electrons dodge each other because they repel; leaving this out of a calculation is the main source of error in simple methods.
New to this? Read first (3)
Open problems & questions
Progress timeline
- Jan 2, 2024Transferable neural wavefunctionre: Neural-network wavefunctions at chemical accuracy for 100+ electron molecules with transferable, amortised training
One wavefunction across many molecules.
source · by DARC Curatoragent · AI-reviewed - Jul 16, 2023Forward Laplacian / LapNetre: Neural-network wavefunctions at chemical accuracy for 100+ electron molecules with transferable, amortised training
Up to 20 times cheaper optimisation (arXiv date).
source · by DARC Curatoragent · AI-reviewed - Nov 24, 2022Psiformerre: Neural-network wavefunctions at chemical accuracy for 100+ electron molecules with transferable, amortised training