Machine Learning for Automated Analog Circuit Design

63 concepts53 papers1 proposal

Digital integrated circuit design has benefited from decades of rigorous automation. Engineers write high-level code in hardware description languages, and compilers deterministically map that code to discrete, standardized cell libraries. Analog circuit design, by contrast, operates in a continuous domain where every wire length, transistor width, and component placement directly alters electrical behavior. Because analog design lacks rigid logical abstractions, it has historically remained a manual, expert-driven "dark art" relying heavily on human intuition and slow, iterative simulations (Jingyu Pan et al. in A Survey of Research in Large Language Models for Electronic Design Automation; Wenlong Lyu et al. in Batch Bayesian Optimization via Multi-objective Acquisition Ensemble for Automated Analog Circuit Design).

As transistor counts scale and performance constraints tighten, this manual bottleneck becomes unsustainable (Lidong Guo in The Analysis of Applications of Microcomputers in Information Engineering: From Automatic Control Electromechanical Systems to LSI Circuit Design). To bridge this automation gap, a major research direction has emerged that leverages machine learning to automate the three core pillars of analog design: topology discovery, parameter sizing, and physical layout generation.

Topology Discovery as Sequence Generation

The first challenge in analog design is determining the topology—the specific structural arrangement of transistors, capacitors, and resistors that can satisfy a given performance specification. Because circuits are fundamentally graphs, generating them directly is computationally difficult due to graph permutation invariance and complex structural constraints.

The core insight driving recent breakthroughs is that circuit graphs can be flattened into specialized sequences, allowing generative models to treat circuit design as a language translation or sequence modeling task.

[Performance Specs] ───> [ Generative Model ] ───> [ Eulerian / Token Sequence ] ───> [ Circuit Netlist ]

For example, Chen-Chia Chang et al. introduced LaMAGIC, which fine-tunes Large Language Models (LLMs) to map high-level performance specifications directly to circuit netlists represented as token sequences (LaMAGIC: Language-Model-based Topology Generation for Analog Integrated Circuits). To overcome the quadratic token complexity that typically plagues large graph descriptions, the authors later developed LaMAGIC2, which utilizes unique component identifiers and direct float-input embeddings to achieve linear token scaling (LaMAGIC2: Advanced Circuit Formulations for Language Model-Based Analog Topology Generation).

Similarly, Jian Gao et al. bypassed traditional graph complexities in AnalogGenie by representing circuits as pin-level Eulerian sequences (AnalogGenie: A Generative Engine for Automatic Discovery of Analog Circuit Topologies). By solving the "Chinese Postman Problem" to minimize sequence length, their framework allows transformers to discover entirely novel topologies. This approach was further optimized in AnalogGenie-Lite to scale to complex designs containing hundreds of devices (AnalogGenie-Lite: Enhancing Scalability and Precision in Circuit Topology Discovery through Lightweight Graph Modeling). To ensure these generated sequences represent physically viable circuits, Prashanth Vijayaraghavan proposed combining instruction-tuned language models with reinforcement learning (RL), using automated AI feedback to iteratively refine the structural validity of the generated topologies (AUTOCIRCUIT-RL: Reinforcement Learning-Driven LLM for Automated Circuit Topology Generation).

Sample-Efficient Parameter Sizing

Once a topology is selected, designers must perform sizing—meticulously tuning the physical dimensions (widths and lengths) of every transistor to balance conflicting objectives like gain, noise, speed, and power consumption. Sizing is a high-dimensional, non-linear optimization problem. Evaluating a single configuration requires running a high-fidelity SPICE simulation, which is computationally expensive and creates a massive bottleneck when multiplied across thousands of design iterations.

To make this search practical, researchers focus on sample efficiency—maximizing performance gains while minimizing simulation counts. One approach is parallelized exploration. Wenlong Lyu et al. developed MACE, a batch Bayesian optimization framework that treats multiple acquisition functions (different mathematical strategies for picking the next test point) as competing objectives (Batch Bayesian Optimization via Multi-objective Acquisition Ensemble for Automated Analog Circuit Design). By optimizing these acquisition functions simultaneously, MACE identifies a diverse batch of high-quality candidate points to evaluate in parallel, accelerating the discovery of the Pareto-optimal design front.

Another approach infuses domain knowledge directly into reinforcement learning agents to constrain the search space. Xuan Zhang et al. embedded circuit topology rules and specification trade-offs into a multimodal neural network, allowing an RL agent to navigate the parameter space with human-like intuition (Domain Knowledge-Infused Deep Learning for Automated Analog/Radio-Frequency Circuit Parameter Optimization). To handle massive, industrial-scale circuits, Ahmet F. Budak et al. introduced DNN-Opt, an actor-critic framework that uses specialized data augmentation to drastically boost sample efficiency, enabling deep neural networks to optimize black-box circuit parameters with minimal simulation overhead (DNN-Opt: An RL Inspired Optimization for Analog Circuit Sizing using Deep Neural Networks).

Traditional Optimization:  [New Circuit] ──> [Train Model From Scratch] ──> [Sizing Solution]
Transferable GCN-RL:       [New Circuit] ──> [Graph Neural Network] ───────> [Instant Sizing Intuition]

Crucially, traditional optimization methods must be retrained from scratch for every new circuit. To enable models to generalize, Hanrui Wang et al. combined Graph Convolutional Networks (GCNs) with RL (GCN-RL Circuit Designer: Transferable Transistor Sizing with Graph Neural Networks and Reinforcement Learning). Because the GCN operates directly on the circuit's underlying graph structure rather than a static list of parameters, the model learns generalized relationships between topology and sizing. This allows the system to transfer its learned design intuition across entirely different circuit structures and technology nodes (e.g., translating knowledge from a 180nm process to a 65nm process).

Closing the Physical Layout Loop

Even a perfectly sized schematic can fail post-fabrication due to layout parasitics—unintended resistances and capacitances introduced by the physical geometry and proximity of wires on the silicon chip. Traditional automated tools often rely on a single-pass generation flow, which lacks a feedback mechanism. If a layout fails post-simulation checks, human experts must step in to manually adjust routing channels and placement (Giuseppe Chiari et al.).

To close this loop, researchers are developing parasitic-aware, closed-loop automation frameworks. The authors of OSIRIS addressed the lack of high-quality training data by creating an open-source pipeline that generates massive datasets of circuit variants, enabling machine learning models to learn from iterative, parasitic-aware layout extractions (OSIRIS: Bridging Analog Circuit Design and Machine Learning with Scalable Dataset Generation).

To capture the implicit geometric wisdom that human designers use to mitigate parasitics, Keren Zhu 0001 et al. developed TAG (TAG: Learning Circuit Spatial Embedding From Layouts). By pre-training a model on unlabelled legacy layouts using graph neural networks and self-attention, TAG learns to map electrical connectivity to spatial intent, significantly improving downstream tasks like parasitic prediction.

For the physical placement phase, Davide Basso et al. combined Relational Graph Neural Networks (R-GCNs) with RL to create a floorplanning system that learns the underlying spatial logic of analog placement, allowing it to generalize across diverse circuit types (Effective Analog ICs Floorplanning with Relational Graph Neural Networks and Reinforcement Learning). To accelerate the subsequent routing phase, Davide Basso integrated Steiner tree algorithms with RL, reducing early layout template generation times from hours to under a minute (Fast ML-driven Analog Circuit Layout using Reinforcement Learning and Steiner Trees).

Towards End-to-End Co-Optimization

Because topology, sizing, and layout are deeply interdependent, optimizing them in isolation often yields suboptimal results. The frontier of this research direction lies in unified, end-to-end co-optimization frameworks.

Sequential Flow:   [Topology Selection] ──> [Transistor Sizing] ──> [Physical Layout] (Parasitics fail ──┐)
                               ▲────────────────────────────────────────────────────────────────────────┘
Co-Optimization:   [ Differentiable Pipeline / Diffusion Model ] ───> [Simultaneous Layout-Ready Design]

Early unified approaches used multi-stage pipelines. Jiwoo Hong et al. proposed a framework that pairs evolutionary algorithms for topology generation with reinforcement learning for transistor sizing, demonstrating significant performance gains over manual designs (An Automatic Circuit Design Framework for Level Shifter Circuits).

Recent work seeks to collapse these stages into a single, simultaneous step. Asal Mehradfar et al. introduced FALCON, a framework that utilizes a differentiable optimization pipeline to co-optimize topology and sizing while explicitly enforcing physical layout constraints from the outset (FALCON: An ML Framework for Fully Automated Layout-Constrained Analog Circuit Design). Taking this concept to its logical conclusion, Chengjie Liu et al. developed DiffCkt, which applies generative diffusion models to treat end-to-end circuit synthesis as an inverse problem (DiffCkt: A Diffusion Model-Based Hybrid Neural Network Framework for Automatic Transistor-Level Generation of Analog Circuits). By generating topology configurations and transistor sizes simultaneously in a single unified framework, DiffCkt bypasses the need for slow, iterative optimization loops, producing layout-ready analog circuits with unprecedented computational efficiency.

Key Foundational Work

System-on-Chip: Reuse and Integration (2006) — contributed 1 concept elements to this direction.
VHDL based design methodology for hierarchy and component re-use (1995) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 50 papers
  2. Version 1 · 2026-10-04 · 88 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_analog_design_automation_gap_d83815e4,
  title = {analog design automation gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/analog-design-automation-gap/rdr_d83815e438f5417f982089c66b840ccf},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_analog_design_automation_gap_d83815e4_v2,
  title = {analog design automation gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/analog-design-automation-gap/rdr_d83815e438f5417f982089c66b840ccf/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake