LLM-Driven, Datasheet-Aware Automated Hardware Compatibility Verification for Early-Stage, Pre-Schematic Embedded System Design
This paper presents an LLM-driven framework that enables early-stage hardware compatibility verification by transforming datasheets and connectivity descriptions into a design graph to extract specific engineering properties and generate deterministic evaluation scripts, achieving 97.5% accuracy while significantly reducing context overhead compared to traditional query-based approaches.
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
Building a new electronic device, from a simple weather station to a complex medical monitor, begins with a moment of high-stakes selection. An engineer must choose a microcontroller, a sensor, a power source, and a communication module, then imagine how they will fit together. In the real world, these components do not just plug in; they must speak the same electrical language, tolerate the same temperatures, and share power without frying each other. If the chosen parts are incompatible, the entire project can fail, sometimes not until the circuit board is already manufactured. Fixing such a mistake is expensive and time-consuming, often requiring the entire design to be scrapped and restarted. For decades, the only way to catch these mismatches was for a human to manually read through hundreds of pages of technical manuals, searching for tiny details like voltage limits or communication speeds, a process that is slow, tedious, and prone to human error.
A team of researchers at the University of Michigan has developed a new way to handle this problem using advanced artificial intelligence, but not in the way one might expect. They created a system called DEVICES, which stands for Datasheet Extraction and Verification of Interface Compatibility for Embedded Systems. This tool is designed to work at the very beginning of a design project, before any detailed wiring diagrams are drawn. It takes a list of hardware parts and a simple description of how they connect, then automatically checks if they will work together. The researchers found that simply asking a powerful computer program to read all the manuals and answer a question is not enough; the program often misses critical details buried in the middle of long documents. Instead, their system breaks the massive task into small, manageable pieces, retrieves only the specific numbers needed for each check, and writes a computer script to perform the math. This approach allows the system to verify compatibility with 97.5% accuracy, a significant improvement over previous attempts, while using far less information to do the job.
The core challenge the researchers faced was that modern artificial intelligence models, while incredibly smart, struggle when asked to read a massive document and find a few specific facts. Imagine trying to find a single specific rule in a library of thousands of books all stacked in one pile; the more books you add, the harder it becomes to find the right page. When engineers tried to upload entire datasheets—documents that can be hundreds of pages long—into an AI and ask it to check for compatibility, the system often overlooked important details or confused specifications from different versions of the same part. The AI would sometimes mix up the voltage requirements for one type of sensor with the power limits of another, leading to incorrect conclusions. The researchers realized that the problem was not the intelligence of the AI, but the way the information was presented to it.
To solve this, the DEVICES system acts more like a careful librarian than a simple reader. First, it takes the engineer's description of how the parts connect and builds a map of the system. This map separates the connections into three distinct categories: how power flows, how data moves, and what environmental conditions the parts share. Instead of asking the AI to look at every page of every manual at once, the system identifies exactly which specific numbers are needed for each connection. If it needs to check if two parts can talk to each other, it only retrieves the pages discussing communication speeds and voltage levels. If it needs to check power, it only looks at battery and regulator specifications. By isolating these small, relevant chunks of information, the system prevents the AI from getting overwhelmed or distracted by irrelevant text.
Once the system has gathered the precise numbers it needs, it does not ask the AI to do the math itself. Large language models are excellent at understanding language and finding information, but they are not always reliable at performing calculations. Instead, the system uses the AI to write a small computer program that performs the necessary comparisons. This program takes the retrieved numbers and checks them against strict rules, such as ensuring the voltage coming out of one device is high enough to be understood by the device receiving it. Because the computer program executes the math deterministically, the result is precise and repeatable. The researchers tested this method on seven different embedded system designs, involving 34 different hardware manuals and over a thousand pages of text. The system successfully identified compatibility issues with 97.5% accuracy, whereas a standard approach that simply uploaded all the documents and asked a question got it right only about 15% of the time.
The study also revealed that even when the AI was given a list of rules to follow, it still made mistakes if it was forced to read through the entire, unorganized documents. This confirmed that the key to success was not just giving the AI better instructions, but also giving it a cleaner, more focused set of information to work with. By breaking the problem down into small, independent checks and only showing the AI the specific data required for each check, the researchers created a system that scales efficiently. As a design grows larger and more complex, the system does not need to read more text; it simply performs more small checks. This approach ensures that the verification process remains fast and accurate, regardless of how many components are involved.
The researchers are careful to note that their system is designed for the early stages of design, before the detailed wiring is finalized. It cannot yet interpret complex graphs or charts found in some manuals, which limits its ability to check certain types of power converters that rely on visual data. However, for the vast majority of standard components, the system provides a powerful safety net. It allows engineers to catch potential failures before they spend money on manufacturing, shifting the focus from tedious manual checking to creative system design. The work demonstrates that the future of engineering automation lies not in making the AI do everything at once, but in building structured systems that guide the AI to the right information at the right time, ensuring that the final result is both reliable and trustworthy.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.