AMSnet 2.0: A Large AMS Databasenwith AI Segmentation for Net Detection

Yichen Shi¹*, Zhuofu Tao²*, Yuhao Gao¹, Li Huang¹*, Hongyang Wang¹, Zhiping Yu³, Ting-Jung Lin¹**, Lei He¹**,

¹Ningbo Institute of Digital Twin, Eastern Institute of Technology, Ningbo, China ²University of California, Los Angeles, USA ³Tsinghua University, Beijing, China
^*Indicates Equal Contribution
^**Indicates Corresponding authors

Paper 🤗 Dataset Supplementary Code arXiv

Pipeline Overview

Overview of the AMSnet 2.0 dataset construction and automated generation pipeline.

Abstract

Multimodal large language models (MLLM) struggle to understand circuit schematics due to their limited recognition capabilities. This could be attributed to the lack of high quality schematic-netlist training data. Existing work such as AMSnet applies schematic parsing to generate netlists. However, these methods rely on hard-coded heuristics and are difficult to apply to complex or noisy schematics in this paper. We therefore propose a novel net detection mechanism based on segmentation with high robustness. The proposed method also recovers positional information, allowing digital reconstruction of schematics. We then expand the AMSnet dataset with schematic images from various sources and create AMSnet 2.0. AMSnet 2.0 contains 2,686 circuits with schematic images, Spectre-formatted netlists, OpenAccess digital schematics, and positional information for circuit components and nets, whereas AMSnet only includes 792 circuits with SPICE netlists but no digital schematics.

Our Workflow

Our core contribution is the development of an end-to-end automated pipeline. It takes schematic images as input and, through a series of AI-based parsing and processing steps, transforms them into machine-readable netlists and human-editable digital schematics. The diagram below provides a high-level overview of this process.

Our pipeline automates schematic analysis:

Annotation: Experts label components and nets on our platform.

Detection: A YOLO model identifies components and distinguishes between wire junctions and crossovers.

Segmentation: A U-Net isolates wire pixels, which are then separated into distinct nets.

Generation: The system analyzes connectivity to create a final Spectre netlist and an editable schematic.

Motivation: The Challenge for AI in Circuit Understanding

Although Multimodal Large Language Models (MLLMs) have shown powerful capabilities in many fields, they struggle to understand and parse professional circuit schematics. This is largely due to the lack of high-quality, large-scale "schematic-netlist" paired data for training. Existing methods for schematic parsing, such as those used in AMSnet 1.0, often rely on hard-coded heuristics, making them difficult to apply to complex or noisy real-world schematics.

Noisy schematics with markings and highlights

Fig. 2: Examples of noisy schematics with overlaid markings (a) and partial highlighting (b), which are challenging for traditional methods.

Our Approach: A Robust Segmentation-Based Pipeline

To overcome the limitations of traditional methods, we propose a novel, deep-learning-based pipeline for schematic parsing. This pipeline not only generates netlists but also recovers positional information to reconstruct editable digital schematics.

Core Technique: Two-Stage Net Detection

The core of our method is an innovative two-stage approach for net detection, which significantly improves the accuracy and robustness of identifying wires from complex images. In the first stage, we use a semantic segmentation network (U-Net) to produce a "mask" of all wire pixels, separating them from the background and other elements. In the second stage, we post-process this wire mask, especially at intersection points, using intelligent split-and-merge operations to precisely delineate wire segments belonging to different nets. This approach eliminates the need for bounding box proposals, fundamentally improving segmentation accuracy.

Fig. 4: Our proposed two-stage net detection method. (a) Semantic segmentation to extract all wires. (b) Split and merge operations at intersection points to identify individual nets.

Handling Noise with Data Augmentation

To enable our model to handle real-world schematics with annotations or markings, we employ a data augmentation technique using mocked markings. During training, we randomly generate rectangles and text markings over clean schematics while keeping the net mask labels unchanged. After this augmented training, our model can successfully "see through" these mocked markings and accurately detect the underlying wires, greatly enhancing its robustness in practical application scenarios.

The AMSnet 2.0 Dataset

We construct and release AMSnet 2.0, a large-scale, multimodal dataset for analog and mixed-signal circuits. It contains 2,686 circuits, providing schematic images, Spectre-formatted netlists, OpenAccess digital schematics, and positional information for all components and net segments. In comparison, the original AMSnet only included 792 circuits with SPICE netlists and no digital schematics. This dataset will provide a solid foundation for training more capable AI models for circuit understanding.

Fig. 8: Distribution of schematic complexity by the number of elements and nets.

Fig. 9: Distribution of schematic element types in the dataset.

Experimental Results

We evaluated our method on a test set containing 80 easy, 70 medium, and 50 hard schematics. We use the F1 score as the evaluation metric and compare our U-Net-based approach against the YOLOv11-seg model, which performs direct instance segmentation. As shown in the table below, our method achieves significantly better performance across all difficulty levels.

Netlist Generation F1 Score (%)

Model	Easy	Medium	Hard
U-Net (Ours)	90.19	84.17	80.39
YOLOv11-seg	83.62	71.10	73.11

Table I: F1 score comparison for the netlist generation task.

AMSnet 2.0 Results on Easy, Medium, and Hard splits

Fig. 5: Results for element detection, net detection, and Spectre netlist generation from easy (a), medium (b), and hard (c) splits.

Reconstructed Schematics in OpenAccess format

Fig. 6: The digital schematics automatically reconstructed by our flow and displayed in Cadence Virtuoso, corresponding to the examples in Fig. 5.

Paper

BibTeX

@misc{shi2025amsnet20largeams,
        title={AMSnet 2.0: A Large AMS Database with AI Segmentation for Net Detection}, 
        author={Yichen Shi and Zhuofu Tao and Yuhao Gao and Li Huang and Hongyang Wang and Zhiping Yu and Ting-Jung Lin and Lei He},
        year={2025},
        eprint={2505.09155},
        archivePrefix={arXiv},
        primaryClass={cs.CV},
        url={https://arxiv.org/abs/2505.09155}, 
  }