Interactive inverse design of layered phononic crystals based on reinforcement learning

As supervised learning has been successfully applied in mechanics, reinforcement learning is being attempted to be used to solve mechanical problems more intelligently. In this study, by imagining the mechanical design as a “game” to make clear what is the “score” to maximize, reinforcement learning is successfully applied to the design of layered phononic crystals with anticipated band structures, which can regulate elastic waves by blocking the waves in the range of bandgap. In order to get the desired bandgaps, it is necessary to design unique topological structure of phononic crystals. In this work, the topological structure of layered phononic crystals can evolve itself through interactive reinforcement learning algorithm, and finally reaches the topological structure which meets the given requirements. The reinforcement learning method performs very well both under the goal of maximizing the first-order bandgap width and designing the bandgap of the specified range, respectively. It is worth mentioning that the method is efficient and stable, that is independent of the initial state and target, and can finally learn an evolution route that will keep the objective function increasing. Inspired by the results of exploration, the theoretical analysis is also carried out to explain the design results and gives the feasible bandgap range in layered phononic crystals with given material properties. This reinforcement learning based interactive design scheme can be easily extended to other inverse design problems.