Source-linked AI summary
Inspector: Conversational and Lightweight Analyzer of Analog Circuit Layouts Using LLM and CNNs
Abril Cano Castro, Giuseppe Chiari, Michele Piccoli, Federico Viola, Davide Zoni
TL;DR
GDSII layout analysis remains dependent on manual inspection or rigid, costly scripts, limiting access to high-level semantic information. Inspector combines fine-tuned LLMs and CNNs to support natural-language queries over analog layouts, and reports up to 81% higher accuracy than general-purpose VLMs across four realistic tasks.
Problem
GDSII analysis still relies on manual inspection or computationally expensive rule-based scripts, limiting extraction of semantic information such as device counts and sub-circuit topologies.
Source: §I · p. 1Method
Inspector combines CNNs for GDSII image recognition with an LLM for user interaction and task and result interpretation.
Source: §I · p. 1Results
Up to 81% higher accuracy than state-of-the-art general-purpose VLMs is reported across four realistic design tasks.
Source: §V · p. 4Takeaways & Limitations
Inspector provides a natural-language interface for querying analog layout information directly from GDSII data.
Source: §V · p. 4The reported GPT-5.2 parameter count is an estimate because its exact model size has not been publicly disclosed.
Source: C. Metrics Definition · p. 4
Abstract
from arXiv · showhide
The integration of artificial intelligence into computer-aided design frameworks has sparked a shift in the design of analog integrated circuits (ICs), transitioning the field from using manual and algorithmic-based solutions to adopting automated and intelligent paradigms. In this scenario, the GDSII file represents the industry-standard database containing the ultimate and most accurate source of information of the analog circuit, encapsulating the complex physical geometries and parasitic realities that define tape out performance. This paper proposes a novel framework that combines fine-tuned LLMs and CNNs to analyze GDSII files of analog circuits, enabling a conversational interface between the tool and the designers. Experimental results using thousands of analog designs across four realistic tasks demonstrate that the proposed solution outperforms state-of-the-art general-purpose massive VLMs by a significant margin (up to 81%), thus providing a lightweight solution to the problem of GDSII analysis.
I. INTRODUCTION
GDSII analysis remains reliant on manual inspection or expensive rule-based scripts, limiting extraction of semantic layout information. Inspector combines fine-tuned LLMs and CNNs for conversational analysis and reports up to 81% higher task accuracy than general-purpose VLMs.
I. INTRODUCTION
- GDSII analysis still relies on manual inspection or computationally expensive rule-based scripts, limiting extraction of semantic information such as device counts and sub-circuit topologies.The paper situates this gap in the back-end phase of analog design.Source: §I · p. 1
- Up to 81% higher task accuracy than state-of-the-art general-purpose VLMs is reported for Inspector’s four realistic design tasks.The evaluation uses real layouts and representative DRC-free, LVS-compliant analog circuits.Source: §I · p. 1
- Inspector’s lightweight conversational model is fine-tuned for analog-layout understanding and natural-language interaction.CNNs recognize GDSII images, while the LLM handles interaction and interprets tasks and results.Source: §I · p. 1
- The work contributes an open dataset and end-to-end workflow with layout images, question-answer conversations, and a reproducible training and evaluation pipeline.Source: §I · p. 1
II. STATE OF THE ART
Prior analog-design AI work has concentrated on front-end schematic, topology, and optimization tasks. Physical-layout analysis is less explored, with limited direct semantic understanding of GDSII.
II. STATE OF THE ART
- Front-end research covers schematic recovery with CNNs and transformers, topology synthesis with LLM- and code-based representations, and autonomous design agents.Examples include Netlistify, LaMAGIC, AnalogGenie, AnalogCoder [13], Artisan, AnalogXpert, ADO-LLM, and LEDRO.Source: §II · p. 1
- Back-end physical-layout analysis remains less explored, and direct semantic understanding of GDSII is limited.Related work includes layout-script generation, lithography-hotspot detection, and translating textual design rules into verification code: SOLOMON, LLM-HD, and DRC-Coder.Source: §II · p. 1
- Analog layout automation includes generator- and algorithm-driven flows as well as neural, Bayesian, reinforcement-learning, and graph-neural-network approaches.Examples named are BAG,, ALIGN, and MAGICAL.Source: §II · p. 2
A. Dataset Generation
Inspector prepares separate training data and workflows for visual detection and language-model interaction. The LLM is fine-tuned on templated Q/A pairs, while task-specific CNNs train on augmented layout images.
A. Dataset Generation
- GDSII layouts are rasterized to PNG, graphically altered, and automatically labeled with component classes and bounding boxes before task-specific dataset splitting.The resulting image-annotation pairs train CNN visual detectors.Source: A. Dataset Generation · p. 2
B. Inspector Training
- The LLM is fine-tuned on Q/A pairs for task identification and result reconstruction, while a dedicated CNN trains on the augmented images for its assigned task.The parallel workflows match each model to its respective semantic or visual learning objective.Source: B. Inspector Training · p. 2
C. Inspector Deployment
At deployment, Inspector uses the prompt to identify the requested task, routes task-specific requests to a corresponding CNN, and turns detections into natural-language responses.
C. Inspector Deployment
- The LLM classifies prompts as generic or task-specific; it answers generic queries directly and assigns a task label to task-specific ones.Source: C. Inspector Deployment · p. 2
- The dataset-composition table reports circuit types, layout variants, average device counts, and per-device-type counts; task descriptions are organized into Easy, Medium, and Hard tiers.Source: Table I · p. 3, Table II · p. 3
- The task label selects a CNN that returns component classes and spatial information, which the LLM interprets in the original prompt’s context to produce an answer.Source: C. Inspector Deployment · p. 3
IV. EXPERIMENTAL EVALUATION
The evaluation uses layouts from several circuit-complexity levels and compares Inspector with two representative VLM baselines. Four tasks progress from individual-device identification to component counting in mixed circuits.
A. Experimental Setup
- The CNN training dataset spans single devices, basic analog blocks, and mixed topologies, all implemented in the SkyWater 130nm PDK and meeting DRC and LVS requirements.InternLM-4KHD and GPT-5.2 serve as representative VLM baselines.Source: A. Experimental Setup · p. 3
B. Tasks Definition
- Tasks A–C identify individual devices, recognize base circuit topologies, and enumerate their components, respectively.The evaluation compares Inspector with InternLM-4KHD and GPT-5.2.Source: B. Tasks Definition · p. 3, A. Experimental Setup · p. 3
- Task D counts components in mixed circuits composed of multiple interconnected blocks.The HPF example uses automatically generated component labels that contain each device’s whole bounding box.Source: B. Tasks Definition · p. 3, Figure 3 · p. 3
C. Metrics Definition
The evaluation defines metrics for CNN detection and LLM task handling, and reports model-level accuracy, runtime, and parameter count. Figure 4 presents their trade-off, with lower-is-better metrics shown as reciprocals.
C. Metrics Definition
- CNN detection is measured with mAP@0.5, mAP@0.5:0.95, precision, and recall; LLM metrics measure task routing and response reconstruction.Task Identification (T-ID) is the percentage of prompts routed to the appropriate pipeline; Results Reconstruction (R-R) is the percentage of responses reconstructed from CNN structured outputs.Source: C. Metrics Definition · p. 3
- Table III compares Inspector and standalone VLMs by parameters, runtime, and accuracy, while Tables IV and V report CNN detection and LLM performance metrics.Figure 4 summarizes the triangular trade-off among accuracy, execution time, and parameter count; lower-is-better metrics are shown as reciprocals.Source: Table III · p. 4, Table IV · p. 4, Table V · p. 4
- GPT-5.2’s parameter count is an estimate because its exact model size has not been publicly disclosed.Source: C. Metrics Definition · p. 4
D. Experimental Results
Inspector combines high task accuracy with second-scale inference, outperforming the standalone VLMs described in the comparison. Its CNN detection and LLM metrics are also reported separately.
D. Experimental Results
- 91–99% accuracy with second-scale inference is reported for Inspector using Llama-3.2 and lightweight task-specific YOLOv8 detectors, while InternLM-4KHD remains below 41%.GPT-5.2 falls from perfect Task A accuracy to 26% on Task D.Source: D. Experimental Results · p. 4
- Above 99% mAP@0.5 is reported for Inspector across all tasks, with near-perfect precision and recall.Source: D. Experimental Results · p. 4
- T-ID and R-R range from approximately 92% to 100%, with second-scale inference.Source: D. Experimental Results · p. 4
E. Ablation Study
The ablation examines how task accuracy changes with training-data availability. Task A needs less data to reach near-optimal performance than Tasks B–D, with Task D the most sensitive to data reduction.
E. Ablation Study
- Task A reaches near-optimal performance with only 50% of the training data, while Tasks B–D improve more progressively as the dataset grows.Source: E. Ablation Study · p. 4
- Task D shows the greatest sensitivity to data reduction, which the authors relate to the structural complexity and variability of higher-level reasoning tasks.Source: E. Ablation Study · p. 4
V. CONCLUSIONS
Inspector combines an LLM and CNNs for conversational analysis of analog circuit layouts from GDSII data. Across multiple tasks and thousands of designs, it significantly outperforms general-purpose VLMs while maintaining a lightweight architecture.
V. CONCLUSIONS
- Up to 81% higher accuracy than state-of-the-art general-purpose VLMs across multiple tasks and thousands of designs.Inspector maintains a lightweight architecture.Source: §V · p. 4
- Inspector combines semantic reasoning with specialized visual detection to analyze analog circuit layouts directly from GDSII data.Designers can query layout information through a natural-language interface.Source: §V · p. 4
- The framework is evaluated on multiple tasks and thousands of designs while maintaining a lightweight architecture.Source: §V · p. 4