Simulation verification of adaptive dead time control for radiation detectors based on reinforcement learning and FPGA

Abstract Addressing the critical engineering challenge of severe data loss caused by system dead time in dynamic high-flux scenarios within radiation monitoring environments, this paper proposes an adaptive dead-time control system based on circuit-algorithm-hardware co-design. The research work initially constructs a deep simulation model of the signal conditioning circuit in the LTspice (Linear Technology SPICE) environment. The model integrates multi-source non-ideal factors such as power supply ripple and transient noise, digital crosstalk, temperature effects, and PCB (Printed Circuit Board) parasitic parameters. Through dual-pulse tests and Monte Carlo simulations, the statistical distributions of key parameters including system dead time and timing jitter are quantitatively extracted, providing physical constraint boundaries for subsequent algorithm design. Secondly, the circuit physical constraints are encoded into a Markov Decision Process (MDP), and a RL (reinforcement learning) model is trained in a realistic simulation environment to master the adaptive dead-time control strategy. Furthermore, via policy distillation, the complex neural network strategy is compressed into a hardware-friendly finite state machine lookup table, significantly reducing computational complexity while maintaining the original strategy’s performance. Finally, system verification of the adaptive controller is completed within the FPGA (Field-Programmable Gate Array development environment, demonstrating minimal resource utilization and timing convergence. The research results indicate that at 1.5 MHz, the total efficiency is improved by approximately 21.07 % compared to the fixed dead-time strategy. The “circuit-aware, algorithm-optimized, hardware co-design” system closed loop established in this study allows the proposed scheme to be directly integrated into existing nuclear monitoring architectures, demonstrating significant engineering application value.