Nanomaterials can be used to create more durable and precise sensors and digital devices, biocompatible implants, and even surfaces that can push cells to grow faster. However, a lot of factors need to be taken into consideration when designing nanomaterials: their chemical makeup, topology (i.e. surface texture), and synthesis parameters. That’s why researchers need to analyze dozens of papers, form a hypothesis, and only then proceed to experiments with new nanomaterials. As the number of papers on materials science keeps growing, this task becomes ever more time-consuming.
The new platform developed at ITMO solves this issue by automating the analysis of research papers on nanomaterials design. In mere minutes, SciNanoAI, the new smart assistant, extracts facts from papers corresponding to the user’s request and converts them into concrete recommendations for experiments. For instance, the system can calculate the optimal parameters for synthesis of nanostructures and their expected biological effects or pick the necessary type of photoresist, a photosensitive element of the substrate used in printed circuit boards.
In terms of interface, SciNanoAI is a web app where users type in their prompts into a dialog box and a few minutes later get a detailed response with brief summaries of and quotes from the cited papers. Interaction with the platform is also possible in a dialog format, allowing researchers to quickly get the information on the nanomaterials’ printing parameters, their composition, and the biological results of cell interaction. This makes it possible to significantly shorten the transition from a hypothesis to experiments.
SciNanoAI responds with 81% accuracy. To minimize its “hallucinations,” the researchers supplied the system with an algorithm that compares generated responses with the data from the cited articles. Responses are additionally validated by a controller agent that evaluates the accuracy of generated data and, if the response is faulty, automatically repeats the request with a modified prompt.
“We deliberately created not an all-purpose AI assistant, but a specialized tool for a specific lab task – the design of nanostructures and prediction of cell responses. The entire system is dedicated to providing a solution to this particular problem. In our database, the papers are classified based on printing parameters, materials, and cell effects. The multiagent architecture with search-augmented generation – which is when an LLM generates responses based on data from outside sources – allowed us to manifoldly increase response quality, which is critical when working with research data. At the first stage, the decomposer agent classifies the query as ‘relevant’ or ‘irrelevant’ and initiates a search in the database only for relevant queries. Then the query is processed by the generator agent, which forms a response, and the conductor agent, which controls the quality of the generated data. All this significantly reduces the ‘noise’ in the answers compared to publicly available systems,” says Nikita Krotkov, one of the paper’s authors and a PhD student at ITMO’s Infochemistry Scientific Center.
Credit: iunewind / photogenica.ru
Another benefit of the platform is the option to quickly update the database: the app uses the FAISS library, which can be easily scaled and supports the addition of new papers without the need to restructure the entire system. Thanks to that, SciNanoAI always has relevant data.
Currently, the database includes 300 scientific sources: papers, books, and patents; the platform automatically recognizes data from tables and figure descriptions. Next, the researchers plan to expand the database, improve data generation from figures and tables, and train the system even more on materials science lingo.
This study is supported by the Russian Ministry of Education grant No. 075-15-2024-679 “Autonomous Laboratory of Intelligent Neuroengineering.”
