Enabling large language model agents for high-fidelity many-body perturbation theory calculations

To understand the material world, scientists often look at how atoms behave when they are excited, such as when light hits a solar cell or when an electric current flows through a chip. While standard computer models can describe how atoms sit still, they struggle to predict what happens when electrons jump to higher energy levels or interact with vibrations in the crystal lattice. To get these answers right, researchers use a set of advanced mathematical techniques known as many-body perturbation theory. These methods are incredibly powerful and accurate, but they are also notoriously difficult to use. They require a researcher to act as a master conductor, manually linking together several different software programs, each with its own language and rules. If the settings in one program do not perfectly match the next, the entire calculation fails, often in ways that are hard to spot until hours of computing time have been wasted. This complexity has kept these powerful tools out of reach for all but the most specialized experts.

A team of researchers has now developed a new system designed to bridge this gap, allowing artificial intelligence to handle the heavy lifting of these complex calculations. The scientists tested whether modern large language models—advanced AI systems capable of reasoning and following instructions—could automatically set up these difficult workflows on their own. They found that without help, even the most sophisticated AI models failed to construct a working calculation more often than they succeeded. The models frequently made subtle errors, such as misaligning the grid of points used to map the material's properties, which caused the entire process to collapse later on. However, when the researchers introduced a new framework called AI-Donkey, the results changed dramatically. This system acts as a knowledgeable guide, providing the AI with expert-curated instructions and a set of rigid tools that ensure every step connects correctly to the next. With this assistance, the AI agents, ranging from powerful commercial models to open-source versions, were able to successfully run complete, error-free calculations for a variety of materials, including one-dimensional tubes, two-dimensional sheets, and solid blocks.

The core of the problem the researchers addressed lies in the nature of these calculations. Unlike simpler computer simulations that stay within a single program, many-body perturbation theory requires a chain of events where the output of one software package becomes the input for another. Each step depends on precise physical details, such as the symmetry of the material and the specific arrangement of atoms. If a researcher makes a small mistake in the first step, that error travels down the line, corrupting the final result. The team discovered that current AI models, despite their ability to write code and answer questions, lack the specific, deep domain knowledge required to manage these chains reliably. When asked to set up a calculation for a material like silicon, the AI would often guess the necessary settings. In one test, the AI successfully set up a basic calculation about half the time, but for more complex tasks involving optical properties or electron vibrations, the success rate dropped to as low as 19 percent. The models would often invent settings that sounded plausible but were physically incorrect, leading to calculations that could not be run or produced useless data.

To solve this, the researchers built AI-Donkey, a framework that does not rely on the AI to guess the rules. Instead, the system provides the AI with a structured library of expert knowledge and a set of automated tools that enforce consistency. Think of the AI as a skilled architect who can design a building but needs a foreman to ensure the bricks are laid according to the blueprint; AI-Donkey acts as that foreman. The system breaks down the complex calculation into manageable steps, each guided by a specific "skill" document written by human experts. These documents explain exactly what information is needed, what the computer should do, and what common mistakes to avoid. The AI then uses these documents to direct a set of computer programs that generate the necessary files and link them together. Because the computer tools handle the technical details and check for errors automatically, the AI no longer needs to memorize every rule or worry about making a typo that breaks the chain.

The results of this approach were striking. When the researchers tested the same AI models with the AI-Donkey framework, the success rate jumped to nearly 100 percent. Every single attempt, whether using a top-tier commercial model or a smaller, open-source version, produced a complete, runnable workflow that passed all checks. The system worked consistently across repeated trials, eliminating the random failures that plagued the unassisted models. The researchers demonstrated this capability by having the AI autonomously calculate the properties of several different materials, including a single layer of a compound made of molybdenum and selenium, a sheet of carbon atoms known as graphene, and a tube of carbon atoms. In each case, the AI correctly identified the specific settings needed for that material's unique structure, ran the calculations on powerful supercomputers, and even analyzed the results to produce visualizations of how electrons behave.

The study also pushed the boundaries of what AI can do in this field by tackling a particularly difficult calculation known as GW perturbation theory, which is used to study how electrons interact with vibrations in the material. This specific task had never been automated before and involves coordinating four different software packages. Without the AI-Donkey framework, the AI models failed completely, unable to figure out the strict requirements for linking the programs. With the framework, however, the AI successfully completed the entire process, correctly handling the complex dependencies between the different stages. This included calculating how a polar semiconductor like silicon carbide responds to electric fields, a task that requires precise handling of long-range forces that are difficult to model. The researchers found that the AI could even perform advanced tasks, such as testing different settings to ensure the results were accurate, a process known as convergence testing, which is usually a tedious and time-consuming job for human scientists.

The implications of this work extend beyond just making calculations faster. By creating a system that can reliably guide AI through complex scientific workflows, the researchers have opened the door for a wider range of scientists to use these powerful tools. The framework is designed to be flexible, meaning it can be updated to include new methods and software as they are developed. The team showed that the system could integrate advanced features, such as techniques to speed up calculations for large systems, without needing to be completely rewritten. This suggests that in the future, scientists might be able to ask an AI to explore new materials or test new theories without needing to be an expert in the underlying computer codes. The key finding is that while AI is becoming better at reasoning, it still needs a structured, expert-guided environment to handle the rigid, multi-step processes required in high-level science. By combining the reasoning power of AI with the reliability of human-curated knowledge, the researchers have created a path toward fully autonomous scientific discovery in the field of materials science.