⚡ electrical engineering

Inverse Design of Multi-Layer Sub-Pixel-Resolution RF Passives Through Grayscale Diffusion with Flexible S-Parameter Conditioning

This paper introduces a novel inverse design framework utilizing grayscale diffusion with annealed Langevin projection to generate manufacturable, two-layer RF passive components with sub-pixel resolution directly from partial S-parameter specifications, achieving high accuracy and flexibility across a broad frequency range.

Original authors: Tommaso Dreossi, Christopher M. Bryant, Hao Liu, Nathan Mirman, Noah Kessler, Michael Frei, Harish Krishnaswamy

Published 2026-05-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tommaso Dreossi, Christopher M. Bryant, Hao Liu, Nathan Mirman, Noah Kessler, Michael Frei, Harish Krishnaswamy

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 an architect, but instead of designing houses, you are designing tiny, invisible highways for radio waves. These highways are called RF passives (like filters and antennas), and they are crucial for everything from your Wi-Fi router to your smartphone.

Traditionally, designing these circuits is like trying to solve a massive, 3D jigsaw puzzle blindfolded. Engineers have to guess a shape, run a super-slow computer simulation to see if it works, realize it's wrong, tweak the shape, and repeat this process for days or even weeks.

This paper introduces a new "AI Architect" that flips the script. Instead of starting with a shape and asking, "What does this do?", you tell the AI, "I need a circuit that does this specific job," and the AI instantly draws the blueprint for you.

Here is how their new system works, broken down into simple concepts:

1. The "Magic Sketchpad" (Grayscale Diffusion)

Old AI design tools were like a black-and-white stamp: they could only put down a square of metal or leave it empty. If the design needed a gap between two wires that was smaller than the size of a single "pixel" on the grid, the old tools failed.

This new system uses Grayscale Diffusion. Think of it like a digital paintbrush that can paint with any shade of gray, not just black or white.

The Analogy: Imagine a digital photo where a single pixel isn't just "on" or "off." Instead, a pixel can be 30% metal, 70% empty space. This allows the AI to draw incredibly fine details—like the tiny, hair-thin gaps needed for high-speed radio waves—without needing a super-high-resolution grid. It captures the "fuzziness" of real-world physics that binary (on/off) tools miss.

2. The "Two-Floor House" (Multi-Layer & Vias)

Previous AI tools could only design circuits on a single flat sheet of paper. Real-world electronics, however, are often like two-story buildings. They have a top layer and a bottom layer of copper, connected by tiny vertical tunnels called vias (like elevators for electricity).

The Innovation: This is the first AI system that can design these "two-story" circuits with the elevators (vias) included. It understands that sometimes you need to route a wire down to the basement floor to get around an obstacle.

3. The "Flexible Client" (Partial Conditions)

Sometimes, a client (the engineer) knows exactly what they want. Sometimes, they only have a vague idea.

The Flexibility: This AI is a flexible client. You can give it a full set of instructions (the exact radio signal you want to block or pass), or you can just say, "I need a filter for this frequency, but you figure out the rest." It can also handle different types of board materials (like different types of plastic) and change where the wires connect, all without needing to be retrained.

4. The "Safety Inspector" (Hard Constraints)

In real life, you can't just draw a wire anywhere; it has to connect to a specific pin on a chip. If the AI draws a wire that doesn't touch the pin, the circuit is useless.

The Fix: The system uses a "Safety Inspector" (called Langevin projection). As the AI draws the circuit, this inspector constantly checks: "Is the wire touching the required pin?" If not, it forces the wire to snap into place. This guarantees that every design the AI produces is physically buildable.

5. The "Speedy Preview" (Forward Model)

Usually, checking if a design works takes days of computer simulation. This paper uses a special AI "Speedy Preview" (a Vision Transformer) that looks at the blueprint and predicts how it will perform in milliseconds.

The Process: The AI generates 100 different blueprints in seconds. The "Speedy Preview" checks them all, picks the best ones, and ranks them.

The Real-World Proof

The authors didn't just simulate this; they built two actual circuits on a board the size of a postage stamp (8mm x 8mm) and tested them in a lab:

The Hairpin Filter Fix: They took an existing filter design that was impossible to build because the gaps between wires were too tiny for standard manufacturing. The AI redesigned it to be "manufacturable" while keeping the same performance. It worked perfectly at 17 GHz.
The "From Scratch" Filter: They gave the AI only a target signal (a combline filter) and no reference design. The AI invented a brand-new circuit with vias and complex shapes from nothing. It worked at 9.5 GHz.

In Summary

This paper presents a new way to design radio circuits. Instead of a slow, manual guessing game, it uses an AI that can:

Draw with "gray" pixels for ultra-fine details.
Build multi-layer circuits with connecting tunnels.
Guarantee that wires connect to the right spots.
Generate dozens of working designs in seconds.

It bridges the gap between "cool computer ideas" and "things we can actually build in a factory."

Technical Summary: Inverse Design of Multi-Layer Sub-Pixel-Resolution RF Passives Through Grayscale Diffusion

Problem Statement
The inverse design of RF passive components from S-parameters is a high-dimensional, ill-posed problem. Conventional design relies on iterative electromagnetic (EM) simulations and expert tuning, while topology optimization requires days to converge. Existing machine learning approaches, including generative models like GANs and prior diffusion frameworks (e.g., DALL-EM), face significant limitations for practical Printed Circuit Board (PCB) design. These methods are typically restricted to:

Single-layer structures without vias.
Binary (metal/no-metal) pixel representations, which lack the resolution to model tightly coupled structures where gaps are smaller than the grid size.
Electrically small boards with fixed port configurations.
Requirements for fully specified S-parameter targets and specific dielectric properties.

Methodology
The authors propose a conditional diffusion framework that generates two-layer PCB layouts with sub-pixel resolution and via inclusion. The system operates as a "generate-and-rank" pipeline comprising two primary neural networks:

Inverse Designer (Generative Model):
- Architecture: A conditional denoising diffusion probabilistic model (DDPM) based on an Attention U-Net.
- Outputs: Two channels: continuous metallization density (grayscale values in [0, 1]) and a dedicated via presence layer.
- Conditioning: The model accepts multi-modal inputs encoded into spatial feature maps and concatenated with the noisy layout:
  - S-Parameters: Target frequency response (1–20 GHz, 51 points), supporting partial specifications (e.g., constraining only $|S_{21}|$ ).
  - Template: Reference topology (e.g., hairpin, combline) or a "null" embedding for topology-free generation.
  - Feed Locations: Physical coordinates for ports, supporting variable port counts.
  - Dielectric Properties: Substrate parameters ( $\epsilon_r$ , $\tan \delta$ , thickness).
- Sampling Strategies:
  - DPM-Solver++: Used for rapid generation (seconds) without hard constraints.
  - Annealed Langevin Dynamics: Used when hard physical constraints are required. This method employs per-step projection to clamp metallization at specified feed points, ensuring physical realizability.
- Post-Processing: A CenterNet-based detection network converts the grayscale rasterized metal layer into parameterized rectangular primitives (e.g., for Gerber export).
Forward Model (Ranking):
- Architecture: A Vision Transformer (ViT) with an MLP head.
- Function: Predicts S-parameters for each candidate layout generated by the diffusion model. Candidates are ranked by the Root-Mean-Square Error (RMSE) and Weighted Mean Absolute Error (WMAE) against the target specifications.

Key Contributions

Grayscale Sub-Pixel Resolution: Unlike prior binary approaches, this framework generates continuous metallization densities. This allows the representation of geometric features (such as coupling gaps) finer than the 0.125 mm grid resolution (64×64 pixels on an 8×8 mm board), which is critical for tightly coupled structures.
Multi-Layer and Via Support: It is the first generative ML system to handle two-layer stackups with vias, enabling complex ground structures and inter-layer connections.
Flexible Conditioning: The system supports partial S-parameter conditioning, variable port placement, and diverse substrate materials, allowing for design exploration without fully specified targets.
Hard Constraint Enforcement: Through annealed Langevin projection, the framework guarantees metallization at feed points, a critical physical constraint often ignored in generative design.

Results
The framework was validated on an 8×8 mm board discretized at 64×64 pixels across 1–20 GHz.

Performance Metrics: The ViT forward model predicts S-parameters with a weighted mean absolute error (WMAE) of $0.77 \pm 1.28$ dB.
Sampling Efficiency:
- Langevin Dynamics: Achieved an 89.1% valid board rate (due to feed projection) with a WMAE of $0.93 \pm 1.38$ dB.
- DPM-Solver++: Faster (1.9s vs. 13.8s per sample) but lower valid board rate (8.5%).
Fabricated Validations (RO4003C Substrate):
1. Hairpin Filter Redesign: The system generated a manufacturable alternative to a hairpin filter whose original coupling gaps violated fabrication rules. The fabricated design achieved a passband centered at ~17 GHz with 6.2 dB insertion loss (target: 1.5 dB), showing good agreement with full-wave simulation.
2. Combline Bandpass Filter (From Scratch): Designed solely from target S-parameters without a reference topology. The resulting design included via-connected ground structures, achieving a 9.5 GHz passband with 4.6 dB insertion loss (target: 1.1 dB).

Significance
The paper claims this work represents the first generative RF design framework to simultaneously support via-inclusive multi-layer structures, grayscale continuous metallization, hard physical constraints via Langevin projection, and partial S-parameter conditioning. By enabling sub-pixel feature resolution on electrically large boards (approximately one guided wavelength at 20 GHz), the approach bridges the gap between generative design and manufacturability, offering a pathway to automate the design of complex RF passives that were previously limited by binary resolution or single-layer constraints.