💻 computer science

AnalogRetriever: Learning Cross-Modal Representations for Analog Circuit Retrieval

AnalogRetriever is a unified tri-modal retrieval framework that enables cross-modal searching of analog circuits (schematics, descriptions, and netlists) by mapping them into a shared embedding space, significantly improving both retrieval accuracy and the performance of downstream agentic design tasks.

Original authors: Yihan Wang, Lei Li, Yao Lai, Jing Wang, Yan Lu

Published 2026-04-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yihan Wang, Lei Li, Yao Lai, Jing Wang, Yan Lu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a chef in a massive, chaotic professional kitchen. You need to make a very specific type of sauce, but you don't have a recipe. You have three things to work with:

A handwritten note that says: "A creamy, spicy sauce with a hint of garlic." (This is the Text).
A blurry photo of a finished plate of food. (This is the Schematic/Image).
A complex list of chemical ingredients and exact measurements like: "50g emulsified fat, 2g capsaicin, 0.5g allicin..." (This is the SPICE Netlist/Code).

In a normal kitchen, if you search for "spicy sauce" using only the ingredient list, you might find nothing because the list doesn't use the word "spicy." If you search using the photo, you might find something that looks similar but tastes completely different.

AnalogRetriever is like a "Super-Sous-Chef" who has learned to understand all three of these things at the exact same time.

The Problem: The "Language Barrier" of Circuits

Analog circuit design is incredibly difficult because the same "idea" (like an amplifier) can be described in three totally different ways: a picture, a paragraph of text, or a wall of mathematical code.

Currently, engineers have to manually hunt through old textbooks and files, trying to match keywords. It’s like trying to find a specific song by searching for the lyrics, the sheet music, and a picture of the album cover separately—and hoping they all lead to the same tune.

The Solution: The "Universal Translator"

The researchers built AnalogRetriever, a system that creates a "Shared Mental Map" for these three different formats.

The Vision Expert (CLIP): It looks at the circuit drawings (the "photos") and understands the shapes and symbols.
The Librarian (Transformer): It reads the descriptions (the "text") and understands the functional goals.
The Structural Engineer (RGCN): This is the secret sauce. Instead of just reading the code like a book, it understands the connections. It knows that in a circuit, it matters deeply whether a wire is plugged into the "input" or the "output" of a component—just like in a recipe, it matters if you add salt before or after the water boils.

By training these three experts together, the system learns that a specific drawing, a specific paragraph, and a specific block of code all point to the exact same "concept."

The "Training Camp" (Curriculum Learning)

You wouldn't throw a toddler into a high-stakes cooking competition on day one. You teach them how to chop vegetables first, then how to follow a simple recipe, and finally how to create a five-course meal.

The researchers did the same thing. They didn't try to teach the system everything at once. They started by teaching it how to understand the "code," then slowly introduced the "images" and "text," and finally gave it "hard tests"—forcing it to distinguish between two circuits that look almost identical but behave differently.

Why does this matter? (The "Robot Chef" Assistant)

The researchers didn't just build a search engine; they plugged it into an AI "Designer" (called AnalogCoder).

Think of the AI Designer as a talented but sometimes "hallucinating" chef. Sometimes, the chef tries to make a sauce but forgets the salt or uses the wrong oil, creating a mess. By giving the chef AnalogRetriever, the chef can now say: "Wait, before I start cooking, let me look at a real, proven recipe first."

Because the AI can now "look up" real, working designs, it makes far fewer mistakes. It turns "failed attempts" into "successful masterpieces."

Summary in a Nutshell

AnalogRetriever is a bridge. It connects the words we use, the pictures we draw, and the math we write, allowing engineers to find exactly what they need in a split second, making the complex world of electronic design faster, smarter, and much more reliable.

Technical Summary: AnalogRetriever

AnalogRetriever is a novel tri-modal retrieval framework designed to bridge the gap between heterogeneous representations in analog circuit design: natural language descriptions (text), schematic diagrams (images), and SPICE netlists (graph-structured code).

1. The Problem: Cross-Modal Semantic Gap in Analog Design

Analog circuit design relies heavily on reusing existing Intellectual Property (IP). However, engineers face a significant "search bottleneck":

Heterogeneity: The same circuit exists in three distinct formats (text, image, netlist), but existing tools only support single-modality keyword matching.
Domain Gap: Standard Vision-Language Models (like CLIP) fail on circuit schematics because schematics are abstract line drawings with domain-specific symbols that fall outside the distribution of natural photographs.
Structural Complexity: SPICE netlists are graph-structured. Standard text encoders cannot capture the topological semantics (e.g., the difference between a gate and a drain connection), leading to poor retrieval accuracy.
Data Quality: Existing datasets (like MASALA-Chai) suffer from low compilation and simulation (DC pass) rates, making them unsuitable for training robust models.

2. Methodology

The authors propose a unified framework that maps all three modalities into a shared $d$ -dimensional embedding space via tri-modal contrastive learning.

A. Modality-Specific Encoders:

Vision & Text: Uses a pretrained CLIP (ViT-L/14) model. To adapt to the circuit domain without "catastrophic forgetting," they freeze the bottom 16 layers and fine-tune only the top 8 layers.
Netlist (Graph): Employs a Port-Aware Relational Graph Convolutional Network (RGCN). Unlike standard GCNs, the RGCN uses separate learnable weights for 20 different edge types (e.g., MOSFET drain/gate/source/bulk). This allows the model to distinguish the electrical function of different terminals on the same component. A learnable attention pooling mechanism then aggregates node embeddings into a graph-level vector.

B. Training Strategy: Three-Phase Curriculum:
To prevent the randomly initialized RGCN from disrupting the pretrained CLIP alignment, the authors use a curriculum:

Phase 1 (Warm-up): Train only the RGCN to align with the frozen CLIP space using code-involved directions.
Phase 2 (Transition): Unfreeze CLIP and enable full 6-way tri-modal contrastive loss with random sampling.
Phase 3 (Hard Negative Mining): Introduce "hard negatives" (circuits that are functionally similar but topologically different) using a linear increase in the hard-negative ratio to sharpen discrimination.

C. Dataset Refinement:
The authors developed a two-stage LLM-based repair pipeline using Ngspice as an oracle. They used an LLM to fix syntax/model errors and an iterative feedback loop to ensure every netlist compiles and passes DC operating-point simulations, raising the dataset's DC pass rate from 11.4% to 99.7%.

3. Key Contributions

Tri-Modal Topology-Aware Retrieval: The first framework to enable seamless search across text, schematics, and SPICE netlists.
Port-Aware RGCN: A specialized graph encoder that captures fine-grained electrical semantics.
Curriculum-Guided Training: A stable optimization strategy that progressively increases task difficulty.
High-Quality Dataset: A refined, verified dataset of 6,354 triplets.

4. Results

Retrieval Performance: AnalogRetriever achieved an average Recall@1 of 75.2% across all six cross-modal directions, outperforming the strongest baseline (CROP) by over 15×.
Mutual Enhancement: Interestingly, adding the netlist modality improved the performance of the original Text $\leftrightarrow$ Image directions, proving that topological cues regularize the entire shared embedding space.
RAG Integration (AnalogCoder): When used as a Retrieval-Augmented Generation (RAG) module for an LLM agent, it improved the functional correctness of eight different LLMs by an average of +5.6%, reaching a state-of-the-art 86.7% on Claude Sonnet 4.6.

5. Significance

AnalogRetriever represents a major step toward Automated Analog Design. By providing a way to "query" a library of verified designs using natural language or sketches, it reduces the manual search burden on engineers. Furthermore, by grounding generative LLMs with retrieved, verified netlists, it mitigates the "hallucination" problem (generating invalid topologies), moving the field closer to reliable, AI-driven circuit synthesis.