# Learning the best thermoelectric nanoscale heat engines through evolving network topology

## TL;DR

This paper uses reinforcement learning to optimize network topology in thermoelectric nanoscale heat engines, revealing how many-body interactions can significantly enhance efficiency and power output, and proposing practical quantum-dot array setups.

## Contribution

It introduces a reinforcement learning framework for optimizing many-body interactions in thermoelectric systems, achieving near-Carnot efficiency with stable power.

## Key findings

- Maximum thermoelectric figure of merit can be increased by orders of magnitude.
- Reinforcement learning identifies optimal network topologies for nanoscale engines.
- Proposed quantum-dot array setups realize the theoretically optimal engines.

## Abstract

The quest to identify the best heat engine has been at the center of science and technology. Thermoelectric nanoscale heat engines convert heat flows into useful work in the form of electrical power and promise the realization of on-chip power production. Considerable studies have so far revealed the potentials to yield an enhanced efficiency originating from quantum confinement effects and energy-dependent transport properties. However, the full benefit of many-body interactions in thermoelectric is yet to be investigated; identifying the optimal interaction is a hard problem due to combinatorial explosion of the search space, which makes brute-force searches infeasible. Here we tackle this problem with reinforcement learning of network topology in interacting electronic systems, and identify a set of the best thermoelectric nanoscale engines. Harnessing many-body interactions, we show that the maximum possible values of the thermoelectric figure of merit and the power factor can be enhanced by orders of magnitudes for generic single-electron levels. This allows for simple and flexible design of realizing the asymptotic Carnot efficiency with subextensive, but still nonzero and stable power. To realize the optimal nanoscale engines, we propose concrete physical setups based on quantum-dot arrays. The developed framework of reinforcement learning through evolving network topology thus enables one to identify full potential of nanoscale systems.