Analog-circuit design repeatedly evaluates expensive physical-design and simulation steps. This research studies whether a layout representation can be combined with inexpensive pre-layout measurements to predict post-placement performance before the full evaluation flow completes.
The work combines GDSII-derived layout images, a convolutional variational autoencoder (VAE), and target-specific regressors. The public repository is a research portal: it documents the problem, data representation, model family, evaluation protocol, and publications without releasing the implementation or proprietary datasets.
Release status: documentation, sanitized examples, and the latest reported benchmark summary are available now.
Let x_layout represent the physical layout image and y_pre represent
inexpensive pre-layout performance measurements. The goal is to estimate
post-placement targets y_post:
f(x_layout, y_pre) → y_post
The predicted targets include circuit-level quantities such as current, small-signal gain, bandwidth, phase margin, and offset, depending on the dataset and experiment. The intended use is design-space exploration: rank or filter candidate layouts before committing to the most expensive downstream analysis.
- A layout-to-image representation for multi-layer analog-circuit geometry.
- A convolutional VAE that compresses a 256×256 layout image into a compact latent representation.
- Regressors that fuse the layout latent with pre-layout performance features.
- Comparisons across grayscale/RGB representations, latent sizes, scaling methods, direct prediction, residual prediction, fine-tuning, ensembles, and transfer-learning variants.
- A reproducible documentation record for the datasets, model family, and evaluation methodology.
flowchart LR
A[Candidate circuit layout\nGDSII geometry] --> B[Layer-aware rasterization\nmonochrome or RGB]
B --> C[256 x 256 image tensor]
C --> D[Convolutional VAE encoder]
D --> E[Latent mean vector μ]
F[Pre-layout metrics\nscaled feature vector] --> G[Feature fusion]
E --> G
G --> H[Target-specific MLP ensemble]
H --> I[Predicted post-placement metrics]
I --> J[Design-space ranking\nand early screening]
The VAE decoder is used for representation learning and reconstruction
analysis. Regression experiments generally use the encoder mean μ as a
deterministic layout feature.
| Area | Technology |
|---|---|
| Language | Python |
| Deep learning | PyTorch, torchvision |
| Image processing | gdspy, Pillow, Matplotlib |
| Data processing | pandas, NumPy |
| Hyperparameter search | Optuna with SQLite studies |
| Classical baseline | scikit-learn Random Forest regressors |
| Experiment tracking | MLflow for parameters, metrics, tags, and artifacts |
| Data versioning | DVC for private datasets and large artifacts |
| Reproducible environment | Docker and Docker Compose |
| Documentation | Markdown and Mermaid diagrams |
The experiments use two related analog-circuit layout collections. Each sample joins a layout identifier with pre-placement and post-placement performance measurements. The raw GDSII files, simulation outputs, generated image sets, and complete CSV tables are not included in this public repository.
| Collection | Generated/source records | Usable paired records | Typical targets |
|---|---|---|---|
| Layout collection 1 | 10,192 | 4,055 | GBW, gain, current, offset, phase margin |
| Layout collection 2 | 5,469 | 5,468 | GBW, gain, current, phase margin |
Counts reflect the filtering rules used by the experiments, including finite metric checks and collection-specific validity constraints. They should not be interpreted as a license to redistribute the underlying data.
More detail is available in docs/dataset.md.
The example layout image is a sanitized geometry rendering. The accompanying numbers are illustrative and are not benchmark claims.
| Input | Example value |
|---|---|
| Layout representation | 256×256 RGB tensor |
| Pre-layout bandwidth | 9.8 MHz |
| Pre-layout gain | 27.4 dB |
| Pre-layout current | 0.17 µA |
| Pre-layout phase margin | 50.1° |
| Output | Example value |
|---|---|
| Predicted post-placement bandwidth | 0.46 MHz |
| Predicted post-placement gain | 29.7 dB |
| Predicted post-placement current | 0.13 µA |
| Predicted post-placement phase margin | 19.5° |
The input/output contract and units are documented in
docs/examples.md.
The latest reported results are summarized below. Values and emphasis are reproduced from the supplied results table; detailed split and protocol metadata should be added alongside any future benchmark expansion.
| Model | Metric | I_dd [µA] |
G_dc [dB] |
GBW [Hz] | PM [°] |
|---|---|---|---|---|---|
| Layer-aware approach | MAE | 1.28 | 0.022 | 926 × 10³ | 0.251 |
| Layer-aware approach | MAPE [%] | 1.32 | 0.05 | 1.05 | 0.38 |
| Layer-aware approach + pre-layout fusion | MAE | 0.022 | 0.022 | 121 × 10³ | 0.132 |
| Layer-aware approach + pre-layout fusion | MAPE [%] | 0.02 | 0.05 | 0.09 | 0.21 |
See docs/results.md for the table and reporting guidance.
The research workflow uses Git for code and documentation revisions, DVC for dataset and large-artifact lineage, MLflow for experiment tracking, and Docker for a repeatable tooling environment. The public repository contains the documentation contract and neutral tooling configuration; private source code, datasets, and restricted run artifacts remain outside this repository.
See docs/reproducibility.md for the workflow and
local Docker/MLflow commands.
The transfer-learning learning-curve protocol, seven-day tuning campaign, and
acceptance tests are documented in
docs/transfer-learning-campaign.md.
This work has led to the following publications and research outputs:
- Leveraging Convolutional Autoencoders for Post-Layout Performance Estimation of Analog ICs — ICECS 2025
- On the Exploration of Convolutional Variational Autoencoders for Analog Integrated Circuit Post-Placement Performance Regression — SMACD25
- Combining Layer-Aware Images and Pre-Layout Fusion for Accurate Analog IC Post-Layout Performance Prediction — SMACD26
- Generative AI for Next Generation Computer Design — book chapter, Springer Nature
Author publication record: Google Scholar profile
See docs/publications.md for the publication record.
This repository does not contain training code, inference code, model weights, raw layout files, proprietary algorithms, private infrastructure, or complete datasets. The companion private repository contains the implementation and internal experiment inventory.
The documentation and approved public assets are released under the MIT
License. See LICENSE.
Citation metadata, author list, institutional affiliations, project URL, and contact information will be completed before the public release is announced.