AI-driven systems blend literature, experiments and robotics to discover new materials

Machine learning models can speed up discovery of new materials by making predictions and proposing experiments. However, most models today only consider some specific types of data or variables. Compare it with human scientists who work in a collaborative environment and consider experimental results, broader scientific literature, imaging and structural analysis, personal experiences or intuition, and opinions from colleagues and peer reviewers.

Currently, MIT researchers have developed methods to optimize recipes for ingredients and design experiments that incorporate information from a variety of sources, including insights such as literature, chemical composition, and microstructural images. This approach is part of a new platform called Copilot by Real World Experimental Scientist (CREST), which also uses robotic equipment for high-throughput material testing.

Human researchers can talk to the system in natural language without the need for coding, and the system makes its own observations and hypotheses along the way. Cameras and visual language models also allow the system to monitor experiments, detect problems and propose fixes.

“In the field of AI for science, the key is to design new experiments,” says Ju Li, professor of engineering at Carl Richard Soderbergh's Department of Electrical Engineering. “We use multimodal feedback. For example, information from previous literature on how palladium behaves in fuel cells at this temperature, as well as human feedback, complement experimental data and design new experiments. We also use robots to synthesize and characterize the structure of materials and test performance.”

This system is explained in a paper published in Nature. Researchers used Crest to explore more than 900 chemicals, conducted 3,500 electrochemical tests, and led to the discovery of catalytic materials that provide record-breaking power density in fuel cells that run on Formate salts to generate electricity.

https://www.youtube.com/watch?v=0cv5zbilk4y

A smarter system

Materials science experiments can be time-consuming and expensive. Researchers need to carefully design their workflows, create new materials, perform a series of tests and analysis to understand what happened. These results are used to determine how to improve the material.

To improve the process, some researchers have resorted to machine learning strategies known as active learning to efficiently use previous experimental data points and to investigate or misuse those data. Combined with a statistical technique known as Bayesian Optimization (BO), active learning has helped researchers identify new materials such as batteries and advanced semiconductors.

“Bayesian optimization is like Netflix recommending the next movie to watch based on your viewing history, except that it recommends the next experiment instead,” explains Li. “But the basic Bayesian optimization is too simple. Because we use box-in design spaces, if we say we use platinum, palladium, or iron, we will change the ratio of these elements in this small space.

Most active learning approaches also rely on a single data stream that does not capture everything that is happening in the experiment. Li and his collaborators built the summit to equip the computing systems with more human-like knowledge while taking advantage of the speed and control of automated systems.

Crest's robotics also include liquid processing robots, carbohydrate impact systems for rapid synthesis of materials, automated electrochemical workstations for testing, characterization devices including automated electron and optical microscopes, and auxiliary devices such as pumps and gas valves. Many processing parameters can also be adjusted.

The user interface allows researchers to chat with Crest and tell them to use Active Learning to find recipes for promising ingredients for various projects. Crest can include substrates in up to 20 precursor molecules and their recipes.

To guide material design, Crest's model searches scientific papers for descriptions of useful elements or precursor molecules. When human researchers tell Crest to pursue new recipes, they begin a robotic symphony of sample preparation, characterization and testing. Researchers can also ask Crest to perform image analysis from scanning electron microscope imaging, X-ray diffraction, and other sources.

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

“For each recipe, we use previous literature text or databases, so before we do any experiments, we create these huge representations of all recipes based on our previous knowledge base,” says Li.

“Private component analysis of this knowledge is performed to obtain a reduced search space that captures most of the performance variation. Design new experiments using Bayesian optimization in this reduced space. After new experiments, newly acquired, human feedback is fed to a large scale language model to expand the knowledge base and redefine reduced searches.

Materials science experiments can also face reproducibility challenges. To address the problem, Crest monitors experiments with cameras, looks for potential problems, and proposes solutions to human researchers via text and speech.

Researchers used coats of arms to develop electrode materials for a sophisticated type of high density fuel cell known as direct formation fuel cells. After investigating over 900 chemicals over three months, Crest discovered a catalytic material made from eight factors with a 9.3x improvement in power density per dollar over the expensive precious metal Pure Palladium. Further testing used Crests material to provide record power density to a working direct-forming fuel cell, despite only a quarter of the precious metals of previous devices.

The results show the possibility that coat of arms of arms can find solutions to real-world energy problems that have plagued the materials science and engineering communities for decades.

“A key challenge for fuel cell catalysts is the use of precious metals,” says Zhang. “For fuel cells, researchers have used a variety of precious metals, such as palladium and platinum. They have incorporated many other inexpensive elements to use multi-element catalysts that create the optimal tuned environment for catalytic activity and poisoned species such as carbon-adsorbed hydrogen atoms.

Useful assistants

Early on, it was revealed that poor reproducibility was a major problem limiting the ability of researchers to implement new active learning techniques on experimental datasets. Material properties can be affected by the way the precursors are mixed and processed, and any number of problems can subtly alter the experimental conditions and must be carefully inspected to correct them.

To partially automate the process, researchers have linked computer vision and vision language models with domain knowledge in the scientific literature. This has given the system the source of non-prevalence and proposed a solution. For example, a model can notice if there is a millimeter-sized deviation in the shape of the sample, or when the pipette goes out something. Researchers have adopted some of the model proposals, leading to increased consistency, suggesting that the model will create already good experimental assistants.

Researchers noted that humans still perform most of the debugging in their experiments.

“Crest is not a replacement for human researchers, it's an assistant, not a replacement,” says Lee. “Human researchers are still essential. In fact, we can use natural language to explain what the system is doing and present observations and hypotheses. But this is a step towards a more flexible, self-driving lab.”

This story has been republished courtesy of MIT News (web.mit.edu/newsoffice/), a popular site that covers news about MIT research, innovation and education.