How Much of PCB Design Will AI Take Over? From Auto-routing to Physical Constraints and Verification

Efforts to utilize AI in circuit design are occurring successively across research, EDA vendors, design services, and individual development. These include the search for high-frequency power circuits, agents that call design tools to repeat verification and re-execution, product concepts connecting PCBs with advanced packaging, and the prototyping of a Linux computer using 2 PCB boards of 843 components. The processes AI can touch are extending beyond auto-routing. Reported by EE Times Japan Reported by TECH+ Announcement by Cadence Quilter case within the lab

The background of this expansion is the need to simultaneously handle signal, power, thermal, and manufacturability conditions as PCBs become higher density and higher speed. In designs that oscillate between not only schematics and bills of materials but also board outlines, footprints, netlists, and analysis results, the iterations of testing candidates, verifying, and feeding results back increase. The emergence of mechanisms that take existing EDA inputs and link component candidate searching, placement/routing, and analysis has expanded the scope of application. Quilter implementation procedure PCB-Bench Survey on PCB design automation Reported by Siemens

The question to be asked is not simply whether AI can draw a PCB, but which processes it can undertake, under which inputs and constraints, and up to what level of verification. The constraints and passing criteria decided by the designer, as well as the final decision passed to manufacturing, remain elsewhere.

From Routing Automation to the Decomposition of Design Processes

Research from Chiba University and Tokyo University of Science reported an attempt to automate the design of high-frequency power circuits by combining machine learning, multiphysics analysis, and multi-objective optimization. This search simultaneously handles electrical characteristics and component sizes for 1kW and 1MHz class LLC converters. Reported by EE Times Japan

Conditions to increase efficiency up to 96.1% and conditions to reduce the volume of magnetic components by approximately 41% have been shown. The target timing for practical application should be read as the research side's outlook. Reported by EE Times Japan

The AI agents described by Siemens EDA envision a flow where the agent calls design tools, verifies the results, and corrects and re-executes them if necessary. Siemens itself indicates a division of roles where humans determine the architecture and intent, and the agent handles the repetitive tasks. Reported by TECH+

Cadence's AuraStack AI Super Agent is a concept that treats the PCB and advanced packaging as a single target, linking planning, constraints, structure, placement/routing, manufacturability, and signal/power/thermal/mechanical analysis. Cadence explains that this can reduce time-to-market by up to 1 (one-half) and increase productivity by up to 15 times. These are figures and functional descriptions based on corporate announcements and should be treated separately from independent benchmark results. Announcement by Cadence

Siemens is also attempting to integrate schematics, layout, cloud, and AI-driven automation in PADS Pro Essentials and Xpedition Standard for small and medium-sized enterprises. This is a movement to place AI at the entrance of everyday EDA, not just for large-scale advanced packaging. Announcement by Siemens

What is common here is that AI is not placed as a function dedicated solely to drawing lines, but is connected to parts of the design process, from loading circuits and constraints to the iteration of analysis.

The numbers claiming speed are within the conditions

Quilter reported that it designed a dual-board Linux computer using 843 components and reached the first boot in approximately 1 weeks. Quilter Case Study in Lab

According to the explanation, the human labor time was 38.5 hours, while conventional work estimated for comparison was 428 hours. Quilter Case Study in Lab

The routing rate was approximately 98%, and DRC errors were reportedly zero. The article also notes that human finishing was required and that there are limits in high-frequency regions for a scale of approximately 10,000 pins. While this example is a significant achievement, the application rate for overall PCB design cannot be derived from this report. Quilter Case Study in Lab

In an evaluation by Circuit Mind based on a report from the Los Alamos National Laboratory, results showed that manual work for a medium-difficulty task took 60 to 80 hours. Circuit Mind Evaluation in Lab

The AI side is said to have completed the work in 4 hours and 13 minutes. Circuit Mind Evaluation in Lab

The completion rate for that task was 90%. Since the evaluation targets and platforms are limited, this figure must be read as a result of an individual evaluation. Circuit Mind Evaluation in Lab

For a difficult task, the completion rate was 50% compared to approximately 80 hours of manual work, with unsupported components and manual routing still remaining. Since the evaluation targets and platforms are limited, this figure must be read as a result of an individual evaluation. Circuit Mind Evaluation in Lab

Flux's funding announcement is business-side material stating that they raised $3.7 million for the development of AI Hardware Engineer. In Flux's auto-layout explanation, it is stated that convergence is increased by up to 4 times and routing and vias are improved. High-speed interfaces, length matching, power nets, layer reservations, and critical differential pairs remain as areas to be handled by the user. Flux Funding Announcement Flux Auto-layout Explanation

When comparing time figures, it is necessary to align the number of components, pin count, frequency, input data, pre-defined placement, manual finishing, and the depth of verification. While speed is important, unless the same task is measured under the same conditions, these are records of application range rather than superiority or inferiority.

What humans prepare before handing over to AI

Quilter's Quick Start explains that to handle layout, a completed schematic, board outline, footprints, and a netlist are required. It supports Altium, KiCad, Allegro, and Xpedition, and requests that components with critical positions be placed in advance. This is an implementation where the AI searches based on inputs and constraints defined by humans. Quilter Setup Procedure

Circuit Mind promotes AI electronic design automation that creates schematics and bills of materials from architecture, optimizing cost, size, power, and availability. While it mentions reports for power, FMEA, derating, and ICD, these are functional descriptions from the service provider. Circuit Mind

On the research side, SchGen, which handles schematic generation using semantic-rooted code representations, and PCBnet, a dataset and method for automatically generating SPICE netlists from schematic images, have been proposed. Once the layer for reading and structuring inputs is established, the entry point for design support expands; however, other problems such as placement, routing, and verification remain before a read schematic becomes a manufacturable board. SchGen PCBnet

What defines "successfully designed"

PCB-Bench is a benchmark that evaluates PCB design not as a language-only task, but using approximately 3.700 text tasks annotated by experts, about 500 multimodal tasks, and over 170 completed board projects. The report states that there are still significant differences in the handling of spatial relationships, constraints, and professional deliverables. This points out the need to separately measure the ability to provide correct explanations in text and the ability to create deliverables that satisfy placement, constraints, and file requirements. PCB-Bench

A survey overviewing PCB design via generative AI also organizes data shortage, physical constraints, and integration with existing tools as major challenges. The pass/fail of a board is determined not by appearance, but by whether electrical, thermal, and mechanical conditions can be satisfied simultaneously. Survey on PCB Design Automation

My perspective

I believe that measuring the current state of AI-driven PCB design solely by the "high performance of auto-routing" is too narrow. In reality, processes such as creating circuit candidates, selecting components, proposing placement, iterating routing, running physical analysis, and cross-checking bills of materials with procurement conditions are gradually merging into a single loop.

Different evidence is required to judge the replacement of experts. This is because there remain people who organize inputs, determine constraints, judge critical nets, revert designs based on simulation results, and approve deliverables before manufacturing. While the scope AI undertakes is expanding, the boundary of responsibility remains on the human side.