Neural networks refine composite fermion states in fractional quantum Hall systems

Physicists have used neural networks to refine composite fermion wave functions, improving quantitative accuracy in simulations of fractional quantum Hall states.

Fractional quantum Hall states arise in strongly correlated two-dimensional electron gases subjected to perpendicular magnetic fields. Jain's composite fermion framework explains these phenomena by mapping electrons bound to even numbers of magnetic flux quanta onto emergent, weakly interacting quasiparticles. While composite fermion trial wave functions reliably capture topological invariants and qualitative ground-state structures, obtaining precise ground-state energies has historically required computationally demanding lowest Landau level projections.

In a paper published in the Physical Review journals, researchers introduced a hybrid variational Monte Carlo approach. The architecture dresses analytic composite fermion states with neural network parameterisations. The analytic base state establishes the correct topological sector, exchange statistics, and zero-structure across the manifold. A neural network layer subsequently modulates many-body correlations, acting as a flexible variational multiplier without disrupting the underlying topological order.

Benchmarked against exact diagonalisation on finite geometries, the dressed wave functions yield lower variational energies and higher state overlaps than conventional trial functions. The neural parameterisation captures short- and intermediate-range electron correlations that rigid analytic wave functions omit. Crucially, anchoring the network to a physically grounded topological base state mitigates the severe optimisation bottlenecks and sign problems that frequently plague purely generic neural network quantum states.

The approach scales more efficiently than exact diagonalisation, allowing numerical investigations of larger electron systems at fractional fillings with competing collective states. As condensed matter theorists adopt machine learning for strongly correlated phases, can hybrid trial functions systematically resolve contested ground states in non-Abelian regimes?