面向两级米勒补偿运算放大器的解析模型引导残差学习代理辅助多目标优化方法
Analytical-Model-Guided Residual Learning Surrogate-Assisted Multi-Objective Optimization for Two-Stage Miller-Compensated Operational Amplifiers
韩庆奇 1李政君 1于洪仕1
作者信息
- 1. （1.辽宁工程技术大学电子与信息工程学院，辽宁葫芦岛，125105）
- 折叠
摘要
针对两级米勒补偿运算放大器多目标优化中仿真代价高、小样本下代理模型泛化能力不足的问题，提出一种解析模型引导的残差学习代理辅助优化框架。首先进行物理先验，以混合集成回归器学习解析模型与SPICE仿真之间的残差，仅需100个训练样本即可达到决定系数不低于0.92的预测精度。其次，设计混合边界置信域约束，有效抑制代理模型在数据稀疏区域的外推误差。在此基础上，构建四轮主动学习循环，将训练样本逐步富集至帕累托前沿附近，最终代理模型的测试决定系数分别达到0.893、0.913、0.809和0.980。最后，以NSGA-III驱动大规模搜索，仅使用616次晶体管级仿真即获得30个高质量非支配设计解，验证成功率达93.75%，各指标平均绝对百分比误差分别为3.9%、12.4%、8.7%和1.8%，相较全仿真驱动优化加速约58倍。帕累托解集的参数分布分析为版图设计提供了明确的优选方向。
Abstract
To address the challenges of high simulation cost and insufficient surrogate model generalization under limited data in multi-objective optimization of two-stage Miller-compensated operational amplifiers, an analytical-model-guided residual learning surrogate-assisted optimization framework is proposed. Physical prior knowledge is first introduced, and a hybrid ensemble regressor learns the residuals between the analytical model and SPICE simulations, achieving a coefficient of determination no less than 0.92 with only 100 training samples. A hybrid boundary confidence region constraint is then designed to effectively suppress extrapolation errors of the surrogate model in data-sparse regions. On this basis, a four-round active learning loop is constructed to gradually enrich training samples toward the Pareto front, resulting in testcoefficient of determination values of 0.893, 0.913, 0.809, and 0.980 for DC gain, unity?gain bandwidth, phase margin, and power consumption, respectively. Finally, driven by NSGA-III, 30 high-quality non-dominated design solutions are obtained using only 616 transistor-level simulations, achieving a verification success rate of 93.75%, mean absolute percentage errors of 3.9%, 12.4%, 8.7%, and 1.8% for the four metrics, and a speedup of approximately 58× over fully simulation?driven optimization. Analysis of the parameter distributions in the Pareto?optimal set provides clear guidance for layout design.关键词
模拟集成电路设计/代理模型辅助优化/残差学习/主动学习/NSGA-III/帕累托前沿Key words
Analog IC design/surrogate-assisted optimization/residual learning/active learning/NSGA III/Pareto front引用本文复制引用
韩庆奇,李政君,于洪仕.面向两级米勒补偿运算放大器的解析模型引导残差学习代理辅助多目标优化方法[EB/OL].(2026-07-03)[2026-07-05].http://www.paper.edu.cn/releasepaper/content/202607-5.学科分类
微电子学、集成电路
