An LLM agent for end-to-end computational materials discovery
The paper introduces MAESTRO, an LLM-based agent system that automates the end-to-end computational discovery of metal-organic frameworks by integrating literature mining, database curation, and multi-scale screening to identify high-performance materials for wet flue gas separation that conventional methods would likely overlook.
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
Finding new materials to solve global problems is often a slow, fragmented process. Scientists must first locate a promising crystal structure in a vast library of scientific papers, then clean and verify that structure, calculate how it behaves under specific conditions, and finally screen thousands of candidates to find the best one. Each of these steps usually requires different software tools and specialized human expertise, making it difficult to connect them into a single, seamless search. This is particularly true for metal-organic frameworks, a class of porous materials made from metal ions linked by organic molecules. These materials act like microscopic sponges with tunable holes, capable of trapping specific gases while letting others pass. They hold immense promise for capturing carbon dioxide from industrial smoke, a critical step in reducing greenhouse gas emissions. However, the sheer volume of existing research and the complexity of the calculations needed to test these materials have made it hard to find the perfect candidate for humid, real-world conditions.
A team of researchers has now demonstrated a new way to bridge these gaps using an artificial intelligence system called MAESTRO. Instead of relying on a human to manually coordinate the different stages of discovery, this system uses a large language model to act as a central brain that directs a team of specialized digital agents. These agents work together to read scientific literature, extract crystal structures, verify their quality, and run complex simulations to test their performance. The system does not just answer questions; it executes an entire research campaign from start to finish, moving from a broad scientific question to a list of validated, high-performing materials.
The researchers put this system to the test with a specific challenge: finding metal-organic frameworks that can capture carbon dioxide from wet flue gas, such as the exhaust from power plants. In these conditions, water vapor is present and often interferes with the capture process, clogging the pores of the material or competing with the carbon dioxide. The MAESTRO system began by scanning over 400,000 scientific papers to link published research with the actual crystal structures of the materials described. It then cleaned and organized these structures into a database, removing duplicates and fixing errors. From a starting pool of more than 64,000 verified structures, the system applied a series of increasingly detailed filters. First, it removed materials that were too small to hold carbon dioxide or contained metals that would react poorly with water. Next, it used rapid computer simulations to identify materials that were naturally resistant to water while still attracting carbon dioxide. This initial screening narrowed the field down to just over 1,000 candidates.
The system then performed more expensive and accurate simulations on these remaining candidates. It tested how much carbon dioxide the materials could hold in a mixture of gases and checked whether the pores were large enough to let the gas molecules move through freely. After these rigorous tests, the pool was reduced to 205 promising structures. The researchers then ran detailed process simulations to see how these materials would perform in a real-world carbon capture setup, measuring both how pure the captured gas would be and how much energy the process would require. The results were striking. The system identified eleven top-performing candidates that outperformed a well-known commercial material used as a benchmark. While the commercial benchmark lost most of its ability to capture carbon when exposed to high humidity, the eleven new candidates retained over 93 percent of their dry-state capacity even when fully saturated with water.
Perhaps the most surprising discovery was where these successful materials came from. By tracking the original research papers, the team found that none of the eleven top candidates had ever been studied for carbon capture. They originated from completely unrelated fields of science, including research on luminescent sensors, magnetic properties, and chemical separation of other gases. This suggests that the best materials for a specific job are often hiding in plain sight within unrelated scientific literature, overlooked because researchers were not looking for them in those specific contexts. The system's ability to cross these boundaries allowed it to find high-performance solutions that a traditional, application-focused search would have missed.
To ensure these findings were robust, the researchers subjected the top candidates to further testing. They simulated the materials under humid conditions with a mixture of carbon dioxide, nitrogen, and water vapor. The results confirmed that the new candidates maintained their performance, whereas the commercial benchmark collapsed under the same conditions. The team also checked whether the rigid structures assumed in their simulations would hold up in reality. Using advanced molecular dynamics, they found that most of the materials kept their shape, though a few showed slight changes when exposed to gas, indicating that real-world testing will be necessary to confirm their behavior.
This work demonstrates that an artificial intelligence agent can successfully coordinate the complex, multi-stage workflow required for modern materials discovery. By connecting the dots between disparate fields of research and automating the tedious process of validation and screening, the system uncovered a new generation of materials that are both effective and resistant to the challenges of humidity. The findings do not just offer a list of new candidates; they provide a blueprint for how future scientific discovery can be conducted, moving beyond isolated experiments to a connected, automated search for solutions that can tackle some of the world's most pressing environmental challenges.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.