Optimization of Single Photon Avalanche Diode Design using Surrogate-Reinforcement Learning
Article
Mamun, KM, Pala, N, Shawkat, MSA. (2026). Optimization of Single Photon Avalanche Diode Design using Surrogate-Reinforcement Learning
. IEEE Photonics Journal, 10.1109/JPHOT.2026.3710044
Mamun, KM, Pala, N, Shawkat, MSA. (2026). Optimization of Single Photon Avalanche Diode Design using Surrogate-Reinforcement Learning
. IEEE Photonics Journal, 10.1109/JPHOT.2026.3710044
This paper presents the optimization of Single-Photon Avalanche Diode (SPAD) device design parameters using surrogate-assisted reinforcement learning (RL) framework. Both Sentaurus Technology Computer-Aided Design (TCAD) and MATLAB simulations are used to generate datasets that capture the relationship between SPAD geometry, doping concentrations, and avalanche triggering probability. Avalanche triggering probability significantly affects the key performance metric of SPAD devices, including photon detection probability (PDP), dark count rate (DCR), and breakdown voltage. Upon training, we evaluate four machine learning (ML) models using our dataset and choose Extreme Gradient Boosting (XGBoost) as high performing surrogate model that accurately predicts the avalanche triggering probability from SPAD design parameters. The trained surrogate is integrated into a RL environment, enabling multiple RL agents, including Proximal Policy Optimization (PPO), Advantage Actor-Critic (A2C), Soft Actor-Critic (SAC), and Temporal-Difference Model Predictive Control 2 (TD-MPC2) to efficiently explore and optimize the SPAD design space without repeated TCAD simulations. The proposed optimization framework successfully identifies multiple high performance SPAD configurations, achieving avalanche triggering probabilities above 0.55 with TCAD validation errors below 4%. Both single-stage and multi-stage learning strategies are investigated, demonstrating improved convergence and optimization stability through staged exploration and exploitation. In addition, dark count rate (DCR) performance, effective noise of SPAD devices, is evaluated confirming that the optimized designs maintain favorable noise performance. Additionally, the proposed surrogate assisted RL-based SPAD optimization substantially reduces the computational cost compared to conventional TCAD-driven optimization. The results demonstrate that surrogate-assisted RL framework provides an efficient and viable alternative to conventional TCAD-driven SPAD design optimization.
