AnalogRetriever: AI Framework for Cross-Modal Analog Circuit Search
AnalogRetriever serves as a comprehensive tri-modal retrieval system designed for analog circuit searches, tackling the issue of navigating diverse representations, including SPICE netlists, schematics, and functional descriptions. Current approaches are restricted to precise matches within a single modality, which limits their ability to understand cross-modal semantic connections. This framework enhances a high-quality dataset built on Masala-CHAI through a two-step repair process, increasing the netlist compilation rate from 22% to 100%. AnalogRetriever utilizes a vision-language model to encode schematics and descriptions, while employing a port-aware relational graph convolutional network for netlists, aligning all three modalities into a unified embedding space through curriculum contrastive learning. Experimental results indicate that AnalogRetriever achieves an a.
Key facts
- AnalogRetriever is a unified tri-modal retrieval framework for analog circuit search.
- It searches across SPICE netlists, schematics, and functional descriptions.
- Existing methods are limited to exact matching within a single modality.
- A two-stage repair pipeline raises the netlist compile rate from 22% to 100%.
- The dataset is built on top of Masala-CHAI.
- Schematics and descriptions are encoded with a vision-language model.
- Netlists are encoded with a port-aware relational graph convolutional network.
- All three modalities are mapped into a shared embedding space via curriculum contrastive learning.
Entities
—