Text to PCB: How a Prompt Actually Becomes a Manufacturable Board

Turning a simple request into a functional printed circuit board (PCB) is a process that often leaves engineers skeptical. The reality is that many AI-generated PCB designs result in pictures that cannot be manufactured by a fabricator. In order to clarify this process and define the boundary between a workable board and an unachievable design, it is crucial to break down the two distinct aspects of PCB design: the judgement half and the measurement half.

The judgement half involves selecting the appropriate components and determining how they are interconnected. This aspect relies heavily on the large language model's ability to reason based on a vast amount of prior knowledge. Such language models are excellent at making informed choices about regulators, microcontrollers, sensors, and communication protocols.

On the other hand, the measurement half encompasses the physical aspects of the board. This includes determining the exact placement of each component, verifying trace clearances, and ensuring that multiple nets can be routed across multiple layers without crossing. Unlike the judgement half, the measurement half is based on hard constraints and requires precise measurements — something that language models are not equipped to handle.

The conventional approach to AI PCB design mistakenly attempts to combine both halves within the same tool. However, this approach fails because language models lack the capability to accurately measure trace clearances or determine if a design can be manufactured. Consequently, an estimate produced by the model does not constitute a valid board for fabrication.

To clarify the design process, we can split it into two distinct stages. Begin by describing the desired board in plain English. For instance, "an ESP32 sensor board with USB-C power, a LiPo charger, and two Qwiic connectors." Subsequently, an LLM will generate a suggested netlist, addressing the judgement half of the design. The next step involves employing a deterministic engine to create the actual board layout.

This engine handles the positioning of components and the routing of copper traces based on a series of hard constraints, similar to traditional autorouters and design-rule checkers.

Following this, a Design-Rule Check is performed to verify the design's compliance with the specified constraints. Unrouted nets are identified and reported, and the output is exported in various formats, including Gerber files, drill files, pick-and-place files, and Bill of Materials (BOM). The deterministic nature of the routing process is essential for reproducibility.

Since the output remains consistent regardless of model updates, temperature variations, or any other changing factors, a reliable hardware design can be achieved.

If a deterministic router is unable to complete the entire design, the honest approach to failure is to explicitly list the nets that could not be routed. This information guides the user towards making the necessary adjustments, most commonly related to component placement. This approach offers a more reliable and transparent solution compared to presenting a fabricated completion percentage, as it directly points out the specific components that require attention.

In conclusion, the process of converting a textual description of a board into a genuine, fabrication-ready PCB is real. However, it is important to recognize that it is not magic. The honest version of this process, which separates the judgement and measurement halves, provides a more reliable and practical outcome compared to the misleading promises made by AI-driven PCB design tools.