arXiv:2609.33802
Autonomous phase discovery
Shiyu Zhou, Yuxuan Zhang, Sebastian Wetzel, Roger Melko, Xiu-Zhe Luo
Abstract
Understanding quantum phases of matter has long relied on physicists' intuition and mathematical tools such as symmetry and topology. Remarkably successful as these approaches have been, they provide no universal way to explore a Hamiltonian space whose organizing principle is not known in advance. In this work, we introduce a fully autonomous system combining differentiable programming and unsupervised learning for quantum phase discovery. The search evaluates ground-state data along an adaptive trajectory rather than on a predetermined parameter grid. We demonstrate the system with three different solvers and benchmark it against random sampling at equal ground-state-evaluation budgets. On a generalized cluster chain hosting up to $200$ distinct phases, the search finds up to $25$ more phases at the same budget, and matches random sampling given thirty times its budget. On a $50$-parameter Chern insulator, it reaches sectors not obtained by the simple harmonic constructions considered here, in a family whose inverse problem remains open, while recovering all sectors found by sampling. Our results establish autonomous, gradient-driven exploration of Hamiltonian space as a practical route to discovering quantum phases without phase labels or a prescribed target phase.
AI Summary
Can a search explore Hamiltonian parameter space for unfamiliar quantum phases without phase labels, target invariants, or a predetermined sweep? This paper combines a differentiable ground-state solver with an autoencoder: gradients increase distance from previously sampled states in latent space, while retraining incorporates newly visited regions. Long-horizon benchmarks additionally use parameter-space trail repulsion. Topological invariants are withheld from the search and used afterward to measure coverage. Demonstrations use exact diagonalization, imaginary-time evolution, and DMRG on small Ising, $\mathbb Z_2$ gauge, and transverse-field XXZ systems. The XXZ trajectory visits four gapped phases and registers the Luttinger-liquid line through a gradient-divergence event. Scaling tests use free-fermion representations and matched ground-state-evaluation budgets. A generalized cluster chain with up to $200$ couplings and winding sectors yields $5$–$25$ more sectors than Gaussian random sampling for $d\geq40$; $10^4$ evaluations exceed random sampling’s coverage at $3\times10^5$. In a $50$-parameter Chern-insulator family, the median coverage at $3\times10^5$ evaluations is $207$ sectors versus $139$. Pooling five runs yields $283$ versus $160$, including every randomly found sector, but does not extend known extremal Chern numbers. Stated limitations include repeated evaluations within visited sectors, retraining overhead, and dependence on phase geometry and connectivity. Large-scale evidence is restricted to free fermions; benchmark families are tuned separately. The Chern encoder’s grid misses two couplings, and evaluation labels agree with a finer-grid reference on $98.2\%$ of points. Matched evaluation counts do not establish matched computational cost.
Error found: The Chern-sector coverage totals lack demonstrated sector-by-sector certification, and equal pointwise labeling-error rates do not establish equal bias in distinct-sector counts. Ground-truth phase coverage enters offline hyperparameter selection, so the label-free claim applies to the deployed search rather than the entire configuration-selection pipeline. The 32×32 Chern input grid makes the two highest sine-harmonic coefficients invisible to the learned novelty objective, although the physical family still has 50 parameters.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 10 pages, 5 figures, https://github.com/shiyu-zhou-7/diff_phase