Machine learning for quantum data
Bohrdt Group · Research
Machine learning for quantum data
We develop and apply machine learning techniques for the analysis of data from quantum simulation experiments, including quantum gas microscopes and superconducting qubit platforms. Our goal is to provide unbiased and interpretable analysis of quantum snapshots.
Interpretable Snapshot Classification
We train neural networks to classify snapshots from quantum simulation experiments. For example, we distinguish time-evolved states from thermal equilibrium states in the Bose-Hubbard model – when the network performs poorly, this indicates that thermalization has occurred.
We also use networks to distinguish between competing theoretical descriptions of experimental data. For the Fermi-Hubbard model, we compared resonating valence bond theory against geometric string theory predictions using quantum gas microscope snapshots.
To ensure interpretability, we developed a physics-inspired correlator convolutional neural network. This architecture replaces standard nonlinearities with correlation function expansions, providing full interpretability and in particular telling us which (higher-order) correlation functions are the most informative for the classification task. We examine both mean values and full counting statistics of correlations to probe long-range phenomena in 2D systems.
Hamiltonian Reconstruction
We reconstruct effective Hamiltonians from quantum systems to gain physical insight into strongly correlated systems. For mixed-dimensional t-J models, we extract effective spin-sector Hamiltonians and analyze how charge carrier motion affects spin interactions. This approach revealed that charge carrier motion induces J₂-coupling in the spin sector, pushing the system into the frustrated J₁-J₂ regime.