Learning the best nanoscale heat engines through evolving network topology

The quest to identify the best heat engine has been at the center of science and technology. Considerable studies have so far revealed the potentials of nanoscale thermal machines to yield an enhanced thermodynamic efficiency in noninteracting regimes. However, the full benefit of many-body interactions 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. We tackle this problem with developing a framework for reinforcement learning of network topology in interacting thermal systems. We find that the maximum possible values of the figure of merit and the power factor can be significantly enhanced by electron-electron interactions under nondegenerate single-electron levels with which, in the absence of interactions, the thermoelectric performance is quite low in general. This allows for an alternative strategy to design the best heat engines by optimizing interactions instead of single-electron levels. The versatility of the developed framework allows one to identify full potential of a broad range of nanoscale systems in terms of multiple objectives. While the thermodynamic power and efficiency of nanoscale heat engines in noninteracting regimes has been well-explored, revealing effect of many-body interactions remains a challenge. Here, the authors develop a reinforcement learning framework to achieve optimal power and efficiency in nanoengines where two-body interactions among elementary components are nonnegligible.