arXiv:2610.05647
Transformer-Based Time-Series Inference of Lindblad Dynamics in Open Quantum Systems
Julian Guam, Jianqing Liu
Abstract
The Lindblad master equation is the standard framework for describing the non-unitary evolution of open quantum systems, where environmental interactions induce dissipation and decoherence. When both the system Hamiltonian and the dissipation rates are partially unknown or explicitly time-dependent, traditional analytical inversion and system-identification techniques become intractable. Recent works have demonstrated that Transformer-based models can infer unknown dissipation rates from observable time series, yet these approaches typically rely on hand-crafted statistical features under idealized and highly restricted conditions. Here we advance the paradigm by introducing a raw time-series Transformer that directly ingests the full trajectories of Pauli expectation values $\langleσ_x(t)\rangle$, $\langleσ_y(t)\rangle$, and $\langleσ_z(t)\rangle$, thereby fully exploiting the self-attention mechanism for temporal modeling. The architecture is further extended to jointly learn unknown Hamiltonian parameters, handle multiple dissipation channels, and operate robustly under realistic measurement noise. Across all tested scenarios the model achieves consistently high reconstruction accuracy while eliminating manual feature engineering. This provides a scalable, robust, and versatile framework for quantum environment sensing in realistic open quantum systems.
AI Summary
The paper studies supervised inference of time-dependent Lindblad dissipation rates from single-qubit trajectories of $\langle\sigma_x\rangle$, $\langle\sigma_y\rangle$, and $\langle\sigma_z\rangle$, including a setting with an unknown Hamiltonian parameter. An encoder-only Transformer processes raw trajectories and predicts parameters specifying the dynamics, replacing hand-crafted statistical and spectral features. Training and testing use synthetic trajectories generated with QuTiP from Haar-random initial pure states. For amplitude damping with $H=\sigma_z$ and a nonnegative quadratic Bernstein-polynomial rate, training on 5,000 trajectories yields held-out coefficient $R^2$ values of 0.9956, 0.9920, and 0.9880. With $H=\sigma_z+h\sigma_x$, joint inference gives $R^2=0.9982$ for $h$ and 0.9859–0.9932 for the dissipation coefficients. For sinusoidal rates, direct amplitude–frequency–phase regression is reported to fail; replacing phase with the initial rate gives $R^2$ values of 0.9956, 0.9750, and 0.9893 for amplitude, frequency, and initial rate, respectively. With additive Gaussian measurement noise having standard deviation 2% of the local signal magnitude, coefficient $R^2$ values range from 0.9696 to 0.9935. Combining this noise with random initial-state rotations of standard deviation 0.015 radians and increasing training to 8,000 trajectories gives 0.9827–0.9947. The demonstrations are restricted to simulated single-qubit amplitude-damping dynamics and prescribed low-dimensional rate families. Multi-qubit systems and additional dissipation channels are presented as extensions rather than evaluated cases. The paper identifies parameterization sensitivity in the sinusoidal task and uses additional training data for the combined-noise setting.
Central claim may not hold: The abstract claims multiple-channel inference, but every reported experiment contains only one amplitude-damping channel. The claimed scalability is not established by the exclusively single-qubit demonstrations or any scaling analysis. The measurement-noise model has variance proportional to the squared signal, unlike finite-shot Pauli variance (1-m^2)/N, so the tests do not establish the claimed finite-shot robustness. Random unitary perturbations preserve the Haar-random initial-state distribution used throughout, making the preparation-noise test distributionally redundant rather than evidence of additional preparation-error robustness. The sinusoidal outputs amplitude, frequency, and initial rate omit the phase branch and therefore cannot uniquely reconstruct the stated unrestricted sinusoidal rate.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.