Speaker
Description
The direction of time's arrow is easily discernible by us in the real, macroscopic world. In the microscopic regime, on the other hand, the dynamics of physical systems are time reversible, and fluctuations prevent us from determining the arrow of time with anything more than probabilities. It has been shown that a machine learning algorithm can discern the arrow of time even in the fluctuating classical microscopic regime[1]. When trained on forward and time-reversed backward trajectories of some dynamical stochastic system, such as a particle in a harmonic potential undergoing Brownian motion, it returns a probability that closely resembles the theoretical one.
Here we find that a machine learning algorithm can also learn the arrow of time when given data produced by a quantum system. We take as our system an optical cavity with an AC driving field, coupled weakly to a reservoir of bosons. Forward trajectories are created by driving the system from an equilibrium state to a non-equilibrium steady state, while our backward trajectories are defined by the opposite process. We then take the labelled forward trajectories and time-reversed backward trajectories as input to a multilayer perceptron classifier algorithm. After hyperparameter tuning we find the neural network learns and is able to classify new trajectories at 76% accuracy, indicating that it can indeed learn to distinguish forward from backward. I will discuss these methods, and how the quantum case relates to Crooks' fluctuation theorem, and compares to the classical stochastic results.
[1] A. Seif, M. Hafezi, and C. Jarzynski, "Machine learning the thermodynamic arrow of time," Nature Physics 17, 105–113 (2021).
| I am the presenting author | Yes |
|---|