The AI ​​system learns from many types of scientific information and conducts experiments to discover new materials Myth news

Machine learning models can speed up the discovery of new materials by making forecasts and suggesting experiments. But most models today consider only a few specific types of data or variables. Compare this with human scientists who work in the cooperation environment and consider experimental results, wider scientific literature, imaging and structural analysis, personal experience or intuition, as well as the contribution of colleagues and reviewers.

Now scientists MIT has developed a method of optimizing recipes for materials and planning experiments that contain information from various sources, such as literature, chemical insights, microstructural images and many others. This approach is part of the new platform, called Copilot for real experimental scientists (Crest), which also uses robotic equipment to test high bandwidth materials, the results of which are transferred to large multimodal models to further optimize the recipes for materials.

Human researchers can talk to the system in natural language, without the required coding, and the system makes their own observations and hypotheses along the way. Cameras and visual language models allow the system to monitor experiments, detect problems and suggest corrections.

“In the field of artificial intelligence, the key is the design of new experiments,” says Ju li, School of Engineering Carl Richard Soderberg Professor of Energy Engineering. “We use multimodal feedback – for example, from previous literature on how Palladium behaved in fuel cells at this temperature and human opinion – to supplement experimental data and design new experiments. We also use robots for synthesis and characterizing the structure of the material and for testing.”

The system is described in Article published in Nature. Scientists used Crest to examine over 900 chemicals and carry out 3500 electrochemical tests, which leads to the discovery of a catalyst material that ensured a record power density in fuel cells, which works on format salt for electricity production.

Attaching to Li in the article as the first authors are a doctoral student Zhen Zhang, PhD Zhichu Ren '24, Phd Student Chia-Wei Hsu and Postdoc Weibin Chen. Their co -authors are assistant to Professor Iwnetim Abate; Extraordinary professor Pulkit Agrawal; JR East Professor of Engineering Yang Shao-Horn; Mit. Aubrey Penn researcher; Zhang-Wei Hong '25, Hongbin XU PHD '25; Dr Daniel Zheng '25; Students Mit Shuhan Miao and Hugh Smith; Myth Postdocs Yimeng Huang, Weiyin Chen, Yungsheng Tian, ​​Yifan Gao and Yaoshen Niu; Was the myth of Postdoc Sipei Li; and colleagues, including Chi-FENG Lee, Yu-Cheng Shao, Hsiao-Tsu Wang and Ying-Rui Lu.

Wiser system

Experiments from material materials can be time consuming and expensive. They require scientists to carefully design work flows, create new materials and carry out a series of tests and analyzes to understand what happened. These results are then used to decide on the improvement of the material.

To improve this process, some researchers turned to the machine learning strategy known as active learning to effectively use previous experimental data points and examine or use this data. In combination with a statistical technique called Bayesian optimization (because), active learning helped researchers identify new materials for batteries and advanced semiconductors.

“Bayesian optimization is like Netflix recommends the next film to watch based on the history of watching, except that it recommends the next experiment,” explains Li. “But the basic Bayesian optimization is too simplified. It uses the design space in a box, so if I say that I will use platinum, palladium and iron, it only changes the ratio of these elements in this small space. But real materials have a lot more dependence, and because it often gets lost.”

Most of the active approaches to learning are also based on individual data streams that do not capture everything that happens in the experiment. To equip computing systems with greater knowledge similar to man, while using the speed and control of automated systems, and his colleagues built Crest.

Robotic Crest equipment includes a fluid service, a car shock system for a quick synthesis of materials, an automated electrochemical workstation for testing, characterizing equipment, including automated electron microscopy and optical microscopy, as well as auxiliary devices, such as gas pumps and valves, which can also be controlled remotely. You can also tune many processing parameters.

Thanks to the user interface, scientists can talk to Crest and say to use active learning to find promising recipes for materials for various projects. Crest may include up to 20 precursor particles and substrates for its recipe. To conduct materials designs, Crest models search scientific documents in search of descriptions of precursor elements or molecules that may be useful. When human scientists tell Crest to implement new regulations, he begins a robotic symphony of preparation, characteristics and testing of the sample. The researcher may also ask Crest to analyze an image from imaging electron microscopy, X -ray diffraction and other sources.

Information from these processes is used to train active learning models that use both literature knowledge and current experimental results to suggest further experiments and accelerate the discovery of materials.

“For each recipe, we use the previous text or literature database and it creates these huge representations of each recipe based on the previous knowledge database before the experiment was carried out,” says Li. “We are analyzing the main ingredients in this space for setting knowledge to obtain a reduced search space that records most of the performance variability. Then we use Bayesian optimization in this reduced space to design a new experiment. After a new experiment, we provide newly acquired experimental data and feedback from the big language.

Experiments with material materials can also face challenges of playback. To solve this problem, Crest monitors his experiments with cameras, looking for potential problems and suggesting solutions through text and voice for human researchers.

Scientists used Crest to develop a material for advanced high -density fuel cells known as a direct format fuel cell. After examining over 900 chemistry in three months, Crest discovered a catalyst material made of eight elements that achieved a 9.3x improvement in the power density to the dollar above the clean palladium, expensive precious metal. In further tests, the Crests material was used to provide record power density to a working direct fuel cell in format, despite the fact that the cell contained only a quarter of precious metals of previous devices.

The results show the potential of Crest in finding solutions to energy problems in the real world, which have been harassing the material and engineering community for decades.

“The use of precious metal is an important challenge for fuel cell catalysts,” says Zhang. “In the case of fuel cells, scientists used various precious metals, such as palladium and platinum. We used a multi -one catalyst, which also includes many other cheap elements to create an optimal coordination environment for catalytic activity and resistance to poisoned species, such as carbon monoxide and adsorbed hydrogen atom. People have been looking for cheap options for many years. cheap.

Helpful assistant

At the beginning, poor playback appeared as the main problem that limited researchers' ability to perform a new active learning technique in experimental data sets. The material properties can be influenced by the way the precursors are mixed and processed, and any number of problems can subtly change the experimental conditions that require careful control.

To partially automatize this process, scientists combined models of computer vision and vision with domain knowledge from scientific literature, which allowed the system to hypothesize the sources of non -production and proposing solutions. For example, models can see when a sample shape is a deviation of the size of a millimeter or when the pipette moves something out of place. Scientists have enabled some model suggestions, which leads to better consistency, suggesting that the models are already good experimental assistants.

Scientists have noticed that people still performed most of the debug in their experiments.

“Crest is an assistant, not a deputy for human researchers,” says Li. “Human researchers are still necessary. In fact, we use natural language so that the system can explain what to do and present observations and hypotheses. But this is a step towards more flexible, independent laboratories.”