Realize the "needle in the haystack"!

Each such study can take hours to days of work
.

Instead, Kulik and her team took 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
.
They then applied this knowledge to generate recommendations for the next generation of material that might be used in the next round of neural network training
.

Through four successive iterations of this process, the neural network improves significantly each time, until it reaches a tipping point where further iterations do not yield any further improvements
.

This iterative optimization system greatly simplifies the process of obtaining potential solutions that satisfy two conflicting criteria
.

The process of finding the best solution in this situation is known as the Pareto front, where improving one factor tends to make the other worse
.
The Pareto frontier represents an ideal state where there is no room for more Pareto improvement
.

This represents the best possible compromise point, depending on the relative importance assigned to each factor
.

Training a typical neural network requires very large datasets, ranging from thousands to millions of examples, but Kulik and her team were able to simplify the process using this iterative process based on the Pareto front model and using only a few hundred samples can provide reliable results
.

In the case of screening flow battery materials, the desired properties are in conflict, as often happens: the optimal material will have high solubility and high energy density (the ability to store energy for a given weight)
.

But increasing solubility reduces energy density and vice versa
.

Not only is the neural network able to quickly propose promising candidates, but it is also able to assign confidence levels to its different predictions at each iteration, which helps refine sample selection at each step
.

"We've developed a way better than best-in-class uncertainty quantification techniques to really know when these models are going to fail," Kulik said
.

The challenge they chose in their proof-of-concept trials was materials for redox flow batteries, a type of battery that holds promise for large-scale grid-scale batteries that could play an important role in the development of clean, renewable energy
.

Transition metal complexes are the material of choice for making such batteries, but there are too many possibilities to evaluate with traditional methods, Kulik said
.

They started out with a list of 3 million such complexes, then eventually whittled it down to eight good candidates, and developed a set of design rules that allowed experimenters to explore the potential of these candidates and their variations
.

Beyond the specific transition metal complexes suggested for further studies using the system, the method itself could have broader applications, she said
.

"We think of it as a framework that can be applied to any material design challenge where you're really trying to solve multiple goals at the same time
.
You know, all the most interesting material design challenges go like this: you have one thing you want to improve, But improving it makes another thing worse
.

"For us, redox flow battery redox coupling is just a good demonstration that we think we can continue to use this machine learning technique to accelerate the discovery of new materials
.
"

For example, optimizing catalysts for various chemical and industrial processes is another complex material search, Kulik said
.

Currently used catalysts tend to contain rare and expensive elements, so finding similarly effective compounds based on abundant and inexpensive materials could be a major advantage
.

"It's a great combination of concepts from statistics, applied mathematics and physical sciences that will be very useful in engineering applications," said George Schatz, a professor of chemistry and chemical and biological engineering at Northwestern University who was not involved in the research
.

The research addresses "how to do machine learning when there are multiple targets," he said
.

This work was supported by the Office of Naval Research, the U.
S.
Defense Advanced Research Projects Agency (DARPA), the U.
S.
Department of Energy, the Burroughs Wellcome Fund, and the AAAS Mar ion Milligan Mason Award
.