EEschematic: Multimodal-LLM Based AI Agent for Schematic Generation of Analog Circuit

Chang Liu, Danial Chitnis · 2025

Read EEschematic: Multimodal-LLM Based AI Agent for Schematic Generation of Analog Circuit by Chang Liu, Danial Chitnis (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-17002.txt.

Introduction

Chang Liu and Danial Chitnis automate drawing analog circuit diagrams in a 2025 paper. A schematic is the drawing engineers use to check a circuit. Text lists of connections, called netlists, hide the visual layout. Designers still need clear diagrams for review and discussion.

Their agent, EEschematic, uses a multimodal language model. Multimodal means it reads text, images, and symbols together. Six example substructures teach it typical placements. A visual step-by-step strategy refines placement and wiring repeatedly. The output stays in a format humans can edit afterward. Tests covered an inverter and two amplifier designs. The generated diagrams showed high visual quality and correct structure. The tool could speed a tedious step in chip design.

Catalog record

CREATORChang Liu, Danial Chitnis
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE17 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-17002

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

EEschematic: Multimodal-LLM Based AI Agent for Schematic Generation of Analog Circuit Thanks: Chang Liu is sponsored by Peter Denyer’s PhD Scholarship at The University of Edinburgh Chang Liu, Danial Chitnis Abstract Circuit schematics play a crucial role in analog integrated circuit design, serving as the primary medium for human understanding and verification of circuit functionality. While recent large language model (LLM)-based approaches have shown promise in circuit topology generation and device sizing, most rely solely on textual representations such as SPICE netlists, which lack visual interpretability for circuit designers. To address this limitation, we propose EEschematic, an AI agent for automatic analog schematic generation based on a Multimodal Large Language Model (MLLM). EEschematic integrates textual, visual, and symbolic modalities to translate SPICE netlists into schematic diagrams represented in a human-editable format. The framework uses six analog substructure examples for few-shot placement and a Visual Chain-of-Thought (VCoT) strategy to iteratively refine placement and wiring, enhancing schematic clarity and symmetry. Experimental results on representative analog circuits, including a CMOS inverter, a five-transistor operational transconductance amplifier (5T-OTA), and a telescopic cascode amplifier, demonstrate that EEschematic produces schematics with high visual quality and structural correctness. I Introduction Machine learning based methods have significantly accelerated research in analog circuit design, particularly in topology generation and sizing. For example, research [ 1 ] develop an analog Predefined Building Block Library (PBBL) for analog sub-circuits and leverages Deep Reinforcement Learning (DRL) to generate building block sequences and decode as Python code to select topologies. AnalogCoder uses LLMs to generate Python code for PySpice, developing a feedback loop to enrich the circuit tool library to improve LLMs’ ability in analog design [ 2 ] . AmpAgent [ 3 ] employs a Retrieval-Augmented Generation (RAG) method, manually converting the circuit images from research papers to SPICE netlists to enhance LLM’s knowledge of analog topology. Most approaches represent circuits as SPICE netlists and Python code, describing connectivity in a textual format. However, analog circuit designers typically rely on schematic diagrams to understand signal paths and circuit topologies, rather than textual descriptions. Consequently, even when ML models generate optimal circuit netlists, they are often complex for designers to interpret and verify, and converting them into schematics is time-consuming and prone to mistakes [ 4 ] . Therefore, an automatic approach for converting analog circuit netlists to visual schematics is desired. Research on automated netlist-to-schematic generation is mainly focused on digital circuits. Existing synthesis and implementation tools such as Vivado [ 5 ] and Synopsys Design Compiler [ 6 ] support automated design and visualization. However, these gate-level approaches are not well-suited for transistor-level analog schematic generation, as analog schematics emphasise factors including symmetry and visual clarity, rather than sequential gate connections. Few studies have explored analog schematic generation. Early works used symmetry-based placement [ 7 ] and net-crossing minimization [ 8 ] , but these methods failed to generalize across circuit types. RL-based approaches generate schematics using building block classification, successfully handling a few analog circuits, yet still rely on comprehensive libraries of subcircuits [ 9 ] . Schemato used domain-specific LLMs to convert netlists to schematics based solely on textual information, achieving a 76% compilation success rate on their dataset. However, it struggles to generate schematics for circuits with more than five components with accurate connectivity [ 4 ] . Recently, Multi-modal Large Language Models (MLLMs) such as Gemini 2, GPT 4, and Claude 4 have demonstrated outstanding abilities in integrating textual and visual understanding and reasoning. MLLMs allow flexible modelling of relationships across modalities, which suggests they have the potential to generate schematics by integrating multimodal information, such as textual descriptions, circuit diagrams, and symbolic netlists, into coherent circuit representations. In this work, we propose EEschematic, a multimodal AI agent for automatic analog schematic generation. We reduce reliance on schematic libraries by introducing a few basic analog substrcuture examples for a few-shot guidance during the initial placement stage. Additionally, we integrate a Visual Chain-of-Thought (VCoT) prompting for placement and wiring, enabling iterative visual reasoning over schematic images and JSON representations. Experimental results on different analog circuits, including the 5T-OTA and the telescopic cascode amplifier, demonstrate that EEschematic can generate visually clear and structurally correct schematics. II Methodology II-A Circuit Representations Fig. 1 : Different modalities of circuit representation, including natural language description, SPICE netlist, circuit schematic diagram, and corresponding position and wiring information within a JSON schema. In this work, circuit schematic generation is formulated as a multimodal task that involves constructing a natural language description, an SPICE netlist, a circuit schematic diagram, and corresponding position and wiring information within a JSON file for the schematic renderer, as shown in Fig. 1 . The SPICE netlist encodes the electrical connectivity and component parameters of the circuit in a machine-readable format. The natural language description provides a human-understandable explanation of circuit functionality, summarizing the roles of key components and their interactions. The schematic diagram offers a visual representation of the circuit topology, enabling intuitive analysis and refinement. The position and wiring data specify the spatial information and interconnections, ensuring consistent schematic rendering. Together, these modalities establish a shared context that enables MLLMs to associate textual, symbolic, and geometric information for more accurate schematic generation. II-B Sub-circuit Examples TABLE I : Sub-circuit examples predefined as LLM’s context for schematic generation. The JSON file is represented in images, which are generated from the schematic renderer to show the spatial structure of the circuits. Single Cascode Single Current Source JSON schema represented schematic Natural Language Description Single transistor. The source and drain are connected… Single transistor. The source should be connected to… Diode Connected Two-Transistor Cascode JSON schema represented schematic Natural Language Description Single transistor with gate and drain shorted together… Two transistors with gates connected together. The left one … Differential Pair Current Mirror JSON schema represented schematic Natural Language Description Three transistors with two sources connected to … Two PMOS transistors with gates connected together…. LLMs’ ability to reason from few-shot examples makes it possible to learn the underlying principles of analog structure from a limited set of examples and generalize this understanding to generate correct placement, demonstrating that it is not necessary to maintain a comprehensive example set containing every possible structure variant, such as both PMOS and NMOS-based implementations. In this work, we manually constructed a reduced analog substructure example set based on the building block library [ 1 ] to store instructions and examples for the placement process. This set comprises three key collections: First, substructure placement descriptions, which serve as placement guidelines. Second, the corresponding SPICE netlists form the textual foundation for both structural recognition and simulation-driven reasoning by the LLM. Third, a JSON schema encodes positional information, which provides examples for reference. An overview of all the examples is presented in Table I . II-C Initial Placement and Wiring The initial placement aims at establishing a foundational structure that ensures proper symmetry, flipping, and rotation of components. The LLM-guided process begins by analyzing the SPICE netlist to identify circuit substructures, with examples provided as context for the LLM. The LLM first analyzes the netlist to identify substructures, then determines their spatial relationships and orientations to generate an initial schematic placement that serves as the basis for

Read the complete text