Speaker
Description
Custom-designed analog circuits are fundamental to many particle-detector readout systems, but transistor-level circuit design remains a largely expert-driven and iterative process. We present Open-Source Analog design Machine(OSAM), a framework for AI-assisted transistor-level analog design that combines (g_m/I_D)-based sizing, SPICE simulation, and large-language-model-guided design refinement. Given circuit-level performance specifications, OSAM uses (g_m/I_D) lookup tables derived from simulated transistor characteristics together with analytical design relations to determine initial operating points, bias conditions, and device sizes. DC, AC, and transient simulations are performed using ngspice. The resulting circuit performance and transistor operating information are provided to the LLM, which proposes constrained modifications to circuit parameters. These modifications are then evaluated through subsequent SPICE simulations, forming a closed-loop design process that combines physics-based initialization, numerical simulation, and LLM-guided refinement.
The framework is demonstrated in the open-source Sky-130 process using a Miller-compensated two-stage operational amplifier, with open-loop gain, unity-gain bandwidth, phase margin, and slew rate specified as design targets. The target specifications are achieved within a small number of closed-loop LLM refinement iterations.
We further extend the methodology to 77 K cryogenic operation using cryogenic transistor models developed at UTA. This enables the same physics-guided methodology to generate and evaluate transistor-level circuits using cryogenic device models, providing a path toward automated design of custom analog electronics for detector systems spanning room-temperature and cryogenic operation.