Identifying optimal cycles in quantum thermal machines with reinforcement-learning

Driven quantum thermal machines, such as heat engines and refrigerators, are quantum devices that allow us to control the conversion between heat and work at the micro-nano scale through time-dependent controls. Their performance is mainly characterized by their power, efficiency, and power fluctuations. However, optimizing such quantities is challenging: in finite-time, the state can be driven far from equilibrium, and the space of all possible time-dependent cycles is exponentially large. While general results have been found in the slow and fast driving regime – general finite-time optimization schemes are currently lacking.

We introduce a general framework based on Reinforcement Learning to discover optimal cycles for quantum thermal machines [1]. Our method makes no assumptions on the shape or speed of the cycle. We employ our method to maximize tradeoffs between high power, high efficiency, and low power fluctuations in simple models of quantum heat engines and refrigerators [2-3], finding cycles that outperform previous proposals made in literature, such as Otto cycles, and showing that such cycles mitigate the detrimental effect of generation of coherence, also known as "quantum friction". The method can be generalized to only observe the heat currents, thus potentially applicable to experimental devices [3].



[1] P.A. Erdman and F. Noé, NPJ Quantum Inf. 8, 1 (2022).

[2] P.A. Erdman et al., arXiv:2207.13104 (2022).

[3] P.A. Erdman and F. Noé, arXiv:2204.04785 (2022).