This article first appeared here on https://cen.acs.org/ by Chemical & Engineering News on March 9, 2026.
The large language model automates literature search, synthesis, and structural analysis to speed up materials discovery and development
The artificial intelligence agent created a crystallization protocol that successfully produced COF-2000, a material that had never been made before. (Credit: Yaghi Group/UC Berkeley)
A team led by 2025 chemistry Nobel laureate Omar Yaghi has developed an artificial intelligence agent to speed the discovery of covalent organic frameworks (COFs). The AI agent acts as a lab assistant—automating the research process from literature search and synthesis to structural analysis and optimization—with extraordinary results, including a 350% boost in crystallinity for a benchmark material (J. Am. Chem. Soc. 2026, DOI: 10.1021/jacs.5c23233).
The paper demonstrates a platform to “significantly accelerate an otherwise really laborious process,” says Aurelio Mateo-Alonso , a supramolecular materials researcher based at the Basque Center for Macromolecular Design and Engineering who wasn’t involved in the work. The algorithm independently automated experiments to make a previously unreported COF. The researchers aim to use the AI-enabled system to predict and perfect protocols for different classes of materials, including other polymers and perovskites.
COFs are a family of materials that look like mesh on the nano level. They feature molecule-size pores that can be tuned according to their building blocks. This structure means they can host different sorts of compounds, which makes them useful in applications like filtration, catalysis, and gas storage.
Daehyun Daniel Ahn (left) and Kaiyu Wang, first authors of the paper, hold molecular models of covalent organic frameworks. (Credit: Yaghi Group/UC Berkeley)
But to be useful, a COF needs to crystallize in an orderly fashion. Crystallinity is crucial because the uniform structure of the material’s pores is directly linked to its stability, surface area, and other properties that affect the COF’s performance, says Safiya Khalil Alhashmi , an engineer developing AI and high-throughput methodologies for materials discovery at New York University Abu Dhabi.
Chemists need to find crystallization conditions that check several boxes, including the right level of solubility and the ability to form covalent bonds that are strong but not so strong that the growing crystal can’t break and correct malformed linkages. Some of the variables are solvent selection, temperature, time, additives, and concentration. “The number of possible combinations is enormous,” says Yaghi, who is a materials scientist at the University of California, Berkeley.
The new AI agent, which is built in the GPT-4o large language model (LLM), mimics a member of the lab . It learns by screening reaction conditions and then refines them after every experiment round, Yaghi explains. Instead of a tedious trial-and-error process, this approach is “systematic, data-driven, and iterative,” he says.
The algorithm mines the literature, recommends chemical conditions, and devises a 96-well matrix to run high-throughput experiments on. “The LLM agent recommends experiments in a structured, iterative loop . . . based on prior results and readouts of crystallinity,” Yaghi says.
LLMs and automations accelerate the process and can arrive at protocols that produce completely new structures, such as a material the researchers named COF-2000. Though Yaghi’s team is not investigating applications for COF-2000, he says its hydrophilic structure suggests “a strong water uptake” capability that seems promising for capturing water from air— similar to previously reported metal-organic frameworks (MOFs) .
“It is a paradigm shift for COF discovery,” Khalil says. The LLM agent could “free chemists to focus on applications, which could finally propel the field [of COFs] from promise to practice” and into the market, she says. Until now, she adds, “fragmented fundamental discovery kept COFs stuck in the valley of death.”
The framework is “agnostic” to the materials, Khalil says. That means the AI agent approach could potentially “extend to other hard-to-crystallize materials beyond COFs,” including MOFs, perovskites, and pharmaceuticals, she says.
Yaghi agrees that, provided the robot can measure a “reliable quantitative objective to optimize upon,” it will be easy to streamline new materials’ synthesis, characterization, and crystallization.
This article first appeared here on https://cen.acs.org/ by Chemical & Engineering News on March 9, 2026.