Few-shot Learning on AMS Circuits and Its Application to Parasitic Capacitance Prediction

Shan Shen 1,2 , Yibin Zhang 2 , Hector Rodriguez Rodriguez 2 , Wenjian Yu 2 · 2025

Read Few-shot Learning on AMS Circuits and Its Application to Parasitic Capacitance Prediction by Shan Shen 1,2 , Yibin Zhang 2 , Hector Rodriguez Rodriguez 2 , Wenjian Yu 2 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-06538.txt.

Introduction

Shan Shen, Yibin Zhang, and colleagues predict parasitic effects in analog chip design from very few examples in a 2025 engineering study. Parasitic capacitance, an unwanted by-product of wiring, shifts how mixed-signal circuits behave, and designers want estimates before layout is final. Circuit design data are scarce and guarded, so deep models starve.

CircuitGPS turns a circuit's netlist into a mixed-type graph, with the target capacitance as a link to predict. Pretraining followed by few-shot adaptation squeezes signal from limited labeled designs. The method aims to give chip teams early warnings about parasitics without a mountain of past designs.

Catalog record

CREATORShan Shen 1,2 , Yibin Zhang 2 , Hector Rodriguez Rodriguez 2 , Wenjian Yu 2
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE33 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-06538

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Few-shot Learning on AMS Circuits and Its Application to Parasitic Capacitance Prediction Thanks: This work is supported by the National Natural Science Foundation of China (NSFC) under grant (No. 62204141) and Beijing Natural Science Foundation (No. Z230002). W. Yu is the corresponding author. The code and the small datasets are available at https://github.com/ShenShan123/CirGPS.git Shan Shen 1,2 , Yibin Zhang 2 , Hector Rodriguez Rodriguez 2 , Wenjian Yu 2 Abstract Graph representation learning is a powerful method to extract features from graph-structured data, such as analog/mixed-signal (AMS) circuits. However, training deep learning models for AMS designs is severely limited by the scarcity of integrated circuit design data. In this work, we present CircuitGPS, a few-shot learning method for parasitic effect prediction in AMS circuits. The circuit netlist is represented as a heterogeneous graph, with the coupling capacitance modeled as a link. CircuitGPS is pre-trained on link prediction and fine-tuned on edge regression. The proposed method starts with a small-hop sampling technique that converts a link or a node into a subgraph. Then, the subgraph embeddings are learned with a hybrid graph Transformer. Additionally, CircuitGPS integrates a low-cost positional encoding that summarizes the positional and structural information of the sampled subgraph. CircuitGPS improves the accuracy of coupling existence by at least 20% and reduces the MAE of capacitance estimation by at least 0.067 compared to existing methods. Our method demonstrates strong inherent scalability, enabling direct application to diverse AMS circuit designs through zero-shot learning. Furthermore, the ablation studies provide valuable insights into graph models for representation learning. I Introduction Deep learning (DL) models rely on the availability of large-scale labeled datasets to train deep neural networks with millions of parameters. In many real-world scenarios, acquiring such extensive labeled datasets is impractical or impossible. For example, obtaining practical and real integrated circuit (IC) designs from the industry is notoriously difficult due to a combination of intellectual property concerns, security risks, and the complexity of the designs. IC designs are the result of extensive research and substantial investment. Strict intellectual property (IP) laws prevent the sharing of proprietary designs without proper authorization. Some IC designs are critical for defense and infrastructure, making them sensitive from a security perspective. Another reason for data scarcity in the IC field is the high cost of labeling data for tasks such as rare event prediction, layout optimization, or parasitic extraction. These tasks demand expert knowledge and the use of time-consuming EDA tools. Consequently, there is a growing need for models that can learn effectively from limited amounts of design data. Few-shot learning (FSL) addresses this challenge by enabling models to generalize to new tasks or classes using only a small number of labeled examples [ 1 ] . Inspired by the human ability to learn new concepts from minimal information, FSL reduces the dependency on large datasets to make machine learning models more adaptable and efficient. This paradigm shows great potential in the EDA field [ 2 ] . In this work, we leverage several FSL techniques in the application of analog/mixed-signal (AMS) circuit design automation. AMS circuit design often demands substantial manual intervention. It relies heavily on manually selecting circuit topology and component sizes. In conventional AMS design workflows, circuit designers rely on pre-layout simulations for design optimization, followed by post-layout verification. However, as process technologies scale to smaller transistors and lower operating voltages, the reduced transistor driving capability exacerbates sensitivity to parasitic effects. For example, the coupling capacitance has become too significant to be overlooked in simulations, producing a substantial disparity between pre-layout and post-layout performance metrics [ 3 , 4 ] . To address the challenges described above, we present a g eneral, p owerful, s calable circuit graph learning method ( CircuitGPS ) to perform various parasitic-related tasks in AMS circuits. The circuits are modeled as heterogeneous graphs, while the coupling effects are treated as links to be predicted. Inspired by SEAL [ 5 ] , a small-hop subgraph sampling technique is proposed to sample target links from a large circuit graph. Then, we customize a graph Transformer (GT) [ 6 ] to extract the meta-embeddings at the subgraph level. The pre-trained meta-learner, combined with its learned embeddings, is subsequently adapted for the downstream task of coupling capacitance regression. Our major contributions are summarized as follows: • A small-hop subgraph sampling technique is proposed to convert a link/node into a subgraph. This universal and scalable operation enables FSL by decoupling the target links/nodes from the specific circuit graph. Therefore, we can leverage more powerful and sophisticated DL models aimed at graph-level representation learning. • A low-cost position encoding, d ouble-anchor s hortest p ath d istance ( DSPD ), is defined to measure the relative position w.r.t the anchors of the subgraph. DSPD improves the model expressivity and leads to better generalization. • A powerful framework, CircuitGPS , is built on the top of the GraphGPS framework [ 6 ] , which can perform multiple tasks on large circuit graphs. The framework was complemented with a fine-tuning strategy for the downstream capacitance regression. • Through our ablation studies, we summarize two key insights into DL graph models trained on the AMS circuits: ( 1 ) involving the original node features during model training for link prediction will reduce generalization of the graph model; ( 2 ) classic message-passing graph neural networks (MPNNs) can achieve highly competitive performance, even outperforming GTs. II Background Few-shot learning . FSL can be formally defined as a machine learning problem where a model must quickly adapt to a new task using only a few training examples or “shots” per class [ 7 ] . Typically, FSL involves a “meta-learning” phase, where the model learns how to learn across a variety of tasks, and an “adaptation” phase, where it applies this knowledge to a previously unseen task with limited data. Subgraph sampling . Zhang and Chen [ 8 ] first extracted local enclosing subgraphs around links as the training data, and used a fully-connected neural network to learn which enclosing subgraphs correspond to link existence, which is called Weisfeiler-Lehman Neural Machine. The enclosing subgraph for a node pair (m,n) is the subgraph induced from the network by the union of the neighbors of m and n up to h hops. These enclosing subgraphs are very informative for link prediction – all first-order heuristics such as common neighbors can be directly calculated from the 1-hop enclosing subgraphs. However, it is shown that high-order heuristics often have much better performance than first and second-order ones [ 9 ] , but the time and memory consumption are unsuitable for most practical applications. In work [ 5 ] , Zhang et al. found that most high-order heuristics can be unified by a \gamma -decaying theory, and that even a small h can safely learn good high-order features. Graph Transformers (GT) . Considering the great successes of Transformers in natural language processing [ 10 ] and also in computer vision [ 11 ] , it is natural to study their applicability in the graph domain. A fully-connected GT [ 12 ] was first introduced by Dwivedi et al. SAN [ 13 ] further improved this work by implementing invariant aggregation of the positional encoding

Read the complete text