Shulou(Shulou.com)05/31 Report--

When searching a theoretical list of possible new materials for a particular application, such as batteries or other energy-related devices, there are often millions of potential materials to consider, as well as multiple criteria that need to be met and optimized simultaneously. Now, MIT researchers have found a way to use machine learning systems to dramatically simplify the discovery process.

As a demonstration, the team came up with a set of eight promising materials out of nearly 3 million candidates, an energy storage system called mobile batteries. They said it would take 50 years of routine analysis, but they did it in five weeks.

MIT chemical engineering professors Heather Kulik and Joan Paul Janet report the findings in their paper ACS Central Science. 19 years old, Chen Ru graduate student.

The study looked at a group of materials called transition metal complexes. These substances can exist in many different forms, Kulik said. They are "really attractive, functional materials that differ from many other material stages. The only way to understand why they work the way they do is to study them using quantum mechanics. "

To predict the properties of any of these materials, it requires time-consuming and resource-intensive spectroscopy and other laboratory work, or it requires time-consuming and highly complex physics-based computer modeling of every possible candidate material or combination of materials. Each such study may require hours to days of work.

Instead, Kulik and her team used a small number of different possible materials and used them to teach an advanced machine learning neural network to understand the relationship between a material's chemical composition and its physical properties. This knowledge is then applied to suggest possible materials for the next generation to use in the next training of the neural network. Through four successive iterations of this process, the neural network improves significantly each time until it reaches a point where further iterations do not yield any further improvement.

This iterative optimization system greatly simplifies the process of finding potential solutions that satisfy two conflicting criteria. When one factor is improved, the process of finding the best solution often worsens another factor, a process known as the Pareto frontier, which represents a graph of points, so any further improvement in one factor worsens the other. In other words, the chart represents the best compromise based on the relative importance assigned to each factor.

Training a typical neural network requires very large data sets, ranging from a few thousand to millions of examples, but Kulik and her team can use this iterative process, based on the Pareto frontier model, using only a few hundred samples to simplify the process and provide reliable results.

In the case of screening mobile battery materials, the desired properties are conflicting, usually as follows: The optimal material will have high solubility and high energy density (ability to store energy at a given weight). However, increasing solubility tends to decrease energy density and vice versa.

Neural networks can not only quickly find promising candidates, but also help improve sample selection at each step by assigning confidence to different predictions at each iteration. "We developed the best uncertainty quantification techniques in the class to really understand when these models are going to fail," Culic said. "

The challenge they chose for proof-of-concept testing was materials for redox flow batteries, which are expected to be used in large-scale grid batteries and can play an important role in achieving clean and renewable energy. Transition metal complexes are the preferred material category for such batteries, Kulik said, but the possibilities evaluated by traditional methods are too numerous. They first listed 3 million such complexes and then eventually reduced them to eight good candidates, along with a set of design rules that should allow experimenters to explore the potential of these candidates and their variations.

"Through this process, neural networks become smarter and more pessimistic in terms of (design) space," she said. Anything beyond what we describe can further improve what we already know. "

In addition to specific transition metal complexes that suggest systems for further study, the method itself can be widely applied, she said. "We really think it's a framework that can be applied to any material design challenge, and you're really trying to solve multiple goals at once. You know, the most interesting material design challenge of all is that you have one thing to improve, but improving makes another thing worse. For us, redox pairs in redox flow batteries are a good example. We think we can learn something from this machine and accelerate matter discovery. "

For example, optimizing catalysts for various chemical and industrial processes is another such complex material search, Kulik said. Catalysts currently in use often involve rare and expensive elements, so finding similarly effective compounds on the basis of abundant and inexpensive materials may be a significant advantage.

"I think this paper represents the first application of multidimensional directional improvement in chemical science," she said. "But the long-term significance of this work lies in the methodology itself, as it may otherwise be impossible. "You begin to realize that even with parallel computing, we wouldn't have proposed design principles in any other way in these situations. The clues that our work produces are not necessarily ideas already known in the literature, nor are they necessarily pointed out to you by experts. "

"It's a very useful combination of concepts in statistics, applied mathematics and physical science," said George Schatz, a professor of chemistry, chemistry and bioengineering at Northwestern University. He said the study involved "how to do machine learning when there are multiple goals." Culic's method uses cutting-edge methods to train artificial neural networks to predict which combinations of transition metal ions and organic ligands are best suited for redox flow.

Electrolyte for kinetic batteries. "

"This approach can be used in many different environments, so it has the potential to transform machine learning, which is a major activity around the world," Schatz said. "

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.