Inverse design of intermediate band solar cell via a joint drift-diffusion simulator and deep reinforcement learning scheme

Abstract

In this work, we developed an efficient inverse design approach for optimal intermediate band solar cells (IBSC) device design given a target performance by using a joint drift-diffusion simulator and deep reinforcement learning scheme. The drift-diffusion simulator for IBSC simulation was constructed by using the semiconductor module and wave optics module of COMSOL MultiPhysics®. The deep deterministic policy gradient (DDPG) algorithm was chosen as the learning algorithm to optimize the specified device structure. A GaAs quantum well-embedded IBSC was used as the test candidate to verify the performance of the DDPG-based inverse design approach. A maximum efficiency of was reached for the device with optimal structure parameters searched by the DDPG agent, which exceeds the target efficiency of 30%. The subsequent optical analysis revealed that the electric field enhancement due to light absorption at the IB region with a wavelength between 450 nm and 600 nm is mainly contributing to the significantly increased short-circuit current for the optimized device. Meanwhile, a parameters correlation with target conversion efficiency evaluated by topological data analysis successfully identified all the positive and negative parameters with respect to the target parameter, indicating the physical soundness of the optimized structure parameters. Our work presented here demonstrates that a well-trained AI agent can fulfill the target efficiency by searching the optimal parameters for solar cell devices. The AI-based inverse design approach shows promising potential to serve as an efficient device design tool by greatly reducing the number of trial-and-error experiment demonstrations and replacing laborious human-guided device design workload.