Trinet-Based Deep Learning Framework For Automated Analog Circuit Parameter Prediction From Design Specifications
Keywords:
analog circuit automation, deep learning, TriNet, parameter prediction, residual learning, feasibility classification, fixed-point inference, Verilog.Abstract
Analog integrated-circuit design requires repeated sizing, simulation, verification, and manual refinement because device dimensions and passive-component values interact nonlinearly with gain, bandwidth, phase margin, slew rate, power, noise, and output swing. This paper presents a TriNet-based deep learning framework that directly predicts implementable analog circuit parameters from user-defined performance specifications. The proposed pipeline combines input normalization, feasibility screening, three progressive learners, residual refinement, output accumulation, and inverse scaling. The base learner captures dominant low-order relationships, the intermediate learner models residual dependencies left by the first stage, and the advanced learner learns fine-grained nonlinear corrections. A validity classifier prevents unrealistic specification combinations from entering the regression path, thereby reducing physically meaningless predictions. For hardware-oriented deployment, the network can be represented using fixed-point arithmetic and synthesizable Verilog modules with deterministic control, bounded latency, and reusable multiply-accumulate resources. The accumulated normalized prediction is converted into practical design variables such as transistor widths and lengths, bias currents, resistance values, and compensation capacitances. The framework is intended to reduce the number of circuit-simulator iterations while preserving design accuracy and supporting rapid design-space exploration. The paper describes the complete methodology, mathematical formulation, training strategy, hardware inference organization, and evaluation protocol. Experimental results can be inserted after simulation and implementation, including prediction error, feasibility-classification accuracy, resource utilization, timing, power, and comparison with single-stage neural models and conventional optimization techniques.
