Efficient RF Passive Components Modeling with Bayesian Online Learning and Uncertainty Aware Sampling

TL;DR

RF passive component modeling is hampered by the need for extensive EM simulations across geometry and frequency. The authors propose a Bayesian online learning framework with a reconfigurable Bayesian neural network and uncertainty-aware sampling to jointly optimize geometry and frequency data collection. Key contributions include a reusable BNN backbone with heads for point- and vector-mode representations, and an uncertainty-guided adaptive sampling strategy that reduces EM simulations while maintaining accuracy. Validation on three RF components shows up to $35\times$ speedup with only $2.86\%$ of EM time and significant RMSE improvements (e.g., up to ~32\% for spiral inductors), demonstrating practical impact for rapid RF design.

Abstract

Conventional radio frequency (RF) passive components modeling based on machine learning requires extensive electromagnetic (EM) simulations to cover geometric and frequency design spaces, creating computational bottlenecks. In this paper, we introduce an uncertainty-aware Bayesian online learning framework for efficient parametric modeling of RF passive components, which includes: 1) a Bayesian neural network with reconfigurable heads for joint geometric-frequency domain modeling while quantifying uncertainty; 2) an adaptive sampling strategy that simultaneously optimizes training data sampling across geometric parameters and frequency domain using uncertainty guidance. Validated on three RF passive components, the framework achieves accurate modeling while using only 2.86% EM simulation time compared to traditional ML-based flow, achieving a 35 times speedup.

Figures (6)

• Figure 1: An illustration of conventional ML-assisted RF passive components modeling flow and proposed Bayesian online learning flow.
• Figure 2: Relationship between test accuracy (evaluated using MSE, R-squared, and PSNR) and training dataset size for a spiral inductor.
• Figure 3: (a) Our proposed Bayesian neural network (BNN) architecture with reconfigurable heads, including fully-connected heads for point mode and convolutional heads for vector mode. (b) Probabilistic graphical model of the proposed BNN structure, where green and blue components represent the fully-connected and convolutional head configurations, respectively.
• Figure 4: (1) Estimated uncertainty distribution across the geometric domain for the spiral inductor test case. (2) BNN inference results showing: (i) sampled transmission line and bandpass coupled-line filter responses, and (ii) their corresponding sampling frequencies. The blue shaded region indicates the min-max range of predicted S-parameters, while the dashed line represents the mean response across all samples.
• Figure 5: Three-dimensional models and top-view layouts of the experimental test structures: (1) bandpass coupled-line filter, (2) spiral inductor, and (3) microstrip transmission line.