[GS_C_QI] Learning Measurement-Induced Phase Transition Using Attention
ABSTRACT
The rapid advance of quantum hardware has introduced opportunities to study complex quantum systems while posing challenges in characterizing their behavior under limited and noisy data. In this talk, I will present the Quantum Attention Network (QuAN) [1], a machine-learning framework that leverages attention mechanisms to process quantum measurement snapshots as tokens, accessing high-order moments of bitstring distributions while remaining robust to noise. QuAN has demonstrated its versatility across diverse problems, including witnessing entanglement scaling in driven Bose-Hubbard dynamics, tracking the growth of quantum complexity in deep random circuits, and revealing mixed-state topological order. Building on this foundation, the main focus of the talk is our recent work on measurement-induced phase transitions (MIPTs) in monitored quantum circuits [2], where conventional order parameters are nonlinear in the density matrix and inaccessible without exponential post-selection or classical simulation. By coupling inter-trajectory and temporal attention, QuAN distinguishes weak- and strong-monitoring data from generic Haar-random unitaries under weak measurements, providing a noise-tolerant upper bound on the MIPT without post-selection. The highlight of the talk is the first step toward observing MIPT beyond the reach of classical simulation: using measurement trajectories obtained from Quantinuum's H2-1 quantum computer, QuAN identifies the phase boundary at a system size beyond the limit of exact classical simulation, and its prediction follows the same trend established by smaller noisy-simulation benchmarks. Combined with native-gate circuit compression and transfer learning across system sizes, this opens a concrete path toward the classically inaccessible regime on an experimental device, where non-Clifford monitored dynamics are out of reach for any classical method.
[1] H. Kim et al., Sci. Adv. 11, eadu0059 (2025).
[2] H. Kim et al., arXiv:2508.15895 (2025).
