A Generalizable Framework for Building Executable Domain-Specific LLMs under Data Scarcity: Demonstration on Semiconductor TCAD Simulation
This paper presents a schema-first alignment framework that synthesizes large-scale domain knowledge and leverages an IR-driven Direct Preference Optimization workflow to build compact, executable domain-specific LLMs that outperform general models in data-scarce scientific fields like semiconductor TCAD and FEM simulation.
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
In the invisible world of modern electronics, the chips that power our phones and computers are no longer just designed; they are simulated. Before a single silicon wafer is cut, engineers use powerful software to build digital twins of these microscopic devices. These simulations, known as Technology Computer-Aided Design, or TCAD, act as a virtual laboratory. They allow scientists to predict how electricity will flow through layers of material only a few atoms thick, testing millions of variations without the cost or time of physical manufacturing. However, running these simulations is notoriously difficult. It requires writing complex, computer-readable scripts that tell the software exactly how to build a virtual device, where to place materials, and how to measure the results. These scripts must be perfect; a single misplaced word or a wrong number can cause the entire simulation to fail, leaving the engineer to start over. For years, the industry has relied on human experts to write these scripts, a process that is slow, prone to error, and difficult to automate.
A team of researchers has now developed a new way to teach artificial intelligence to write these scripts correctly, even when there is very little data to learn from. They created a system called TcadGPT, a specialized computer model designed to understand the language of semiconductor engineering. Unlike general-purpose AI models that can write poetry or summarize news but often fail at precise technical tasks, this new model was trained to generate code that the simulation software can actually run. The researchers found that by combining a massive amount of synthetic questions and answers with a unique method of correcting the model's mistakes, they could teach the AI to produce valid scripts with high reliability. Their work suggests that in highly specialized fields where data is scarce and errors are costly, a carefully tuned, smaller AI model can outperform even the most advanced general-purpose systems, provided it is taught the right way.
The challenge the researchers faced was unique. In many fields, AI learns by reading vast libraries of existing text and code. But in semiconductor design, the real-world scripts used in factories are often secret, and the public manuals available for learning are not written for computers to read. This created a data shortage. To solve this, the team did not wait for more data to appear; they created their own. They took existing user guides and textbooks and used a powerful AI to generate over 1.5 million new question-and-answer pairs. They designed two different methods for this generation. One method read through the documents broadly to capture general concepts, while the other focused on specific keywords like material names or physical laws to ensure deep precision. This process gave the model a strong foundation of knowledge, allowing it to answer questions about how the simulations work with an accuracy of 85.6 percent, significantly beating general AI models that scored below 50 percent on the same tests.
However, knowing the facts is not the same as writing the code. The researchers discovered that simply teaching the model to answer questions was not enough to make it write scripts that the software could execute. The model would often get the physics right but fail the syntax, placing commands in the wrong order or missing a crucial step. To fix this, they introduced a second, more rigorous training phase. They took verified, working scripts and broke them down into a structured, logical format that stripped away the specific wording but kept the core meaning. They then created thousands of variations of these scripts, some perfect and some with deliberate, small errors. The model was then shown these pairs and asked to choose the correct one, learning to recognize and avoid even the smallest mistakes. This process, which the researchers call IR-to-DPO, taught the model to be strict about following instructions.
The results of this two-step training were striking. When tested on a set of twenty new instructions that the model had never seen before, the final version of TcadGPT successfully generated scripts that the simulation software accepted without error 80 percent of the time. In contrast, a top-tier general-purpose AI model failed every single one of these twenty attempts, producing code that looked correct to a human but was rejected by the machine. The study also revealed a surprising insight about how AI learns in these specialized fields. The researchers tested whether giving the model access to a library of documents during the test would help, a technique known as retrieval. While this helped general AI models, it actually hurt the performance of their specialized model. The researchers found that for a model already deeply trained on a specific subject, adding outside information during the test confused it, causing it to lose focus on the precise rules it had learned.
To prove that their approach was not just a lucky accident specific to one type of software, the team applied the exact same training recipe to a different kind of simulation tool called Elmer, which is used for solving complex physics problems in engineering. Even though this tool was completely different from the semiconductor software, the model trained with their method still outperformed general AI models in both answering questions and writing working code. This suggests that the framework they built is a robust method that can be adapted to other difficult, data-poor scientific fields. The work demonstrates that in the high-stakes world of engineering, where a single error can cost millions of dollars, the path forward is not necessarily bigger, more general models, but rather smaller, highly specialized ones that are taught with extreme precision and discipline.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.