LAKE WALES Fla.—Instead of using trial-and-error, researchers at the University of Connecticut are discovering new organic electronic materials using machine learning, in the spirit of President Obama's Materials Genome Initiative and now offers their results to the public domain. The online program enables engineers to predict the electronic qualities of a particular polymer, or specify their desired electronic qualities and automatically derive a polymer that satisfied their criteria in seconds.
President Obama inspects some new materials used to make more energy efficient fluorescent-like tubular light bulbs for his Materials Genome Initiative (MGI).
Electronic materials do not actually have a genome—the complete set of genes for a particular organism—but as a metaphor the Materials Genome Initiative seeks to predict the properties of materials using quantum mechanics to simulate them on a computer. Using this method the University of Connecticut has now discovered the properties of organic electronics (polymers) with a machine learning algorithm that discovered the specific laws governing the dielectric breakdown of polymers thus accurately predicting their bandgap and dielectric constant from their atomic structure.
[ Understanding the IoT Platform Wars: Technical Trade Offs at ESC Boston]
"We use a similarity-based learning algorithm—kind of like regression. We compare similar cases and to come up with accurate estimates of a new material's electrical properties," University of Connecticut materials scientist Ramamurthy ‘Rampi’ Ramprasad told EE Times in an exclusive interview. "And at the end of the day we can predict the electronic properties of a new organic polymer from its composition.'
Machinwbe learning is uncovering an organic electronics genome that can be used for designer materials.
(Source: Kim, Ramprasad Lab/UConnUniversity of Connecticut)
The learning algorithm— implemented in their online tool Khazana which means the discovery of hidden treasure—as been handed over to the public domain, and is available for any engineer to use for free.
How's it work
Ramprasad achieved this marvel with a load of computer time running quantum level simulations of polymers to determine their theoretical bandgap and dielectric constants, picking the most commonly used building blocks of electronic polymers, and his own machine learning algorithm to combine them into 283 representative candidates.
As you may already know, a polymer is composed of a set of repeating building blocks, somewhat like the genome of a person. By narrowing them down to the most common sequences of the organic polymers already doing duty as electronic components, Ramprasad was able to cover a gigantic portion of the polymer-space with just seven building blocks: CH2, C6H4, CO, O, NH, CS, and C4H2S.
Machine learning discovered which atomic building blocks gave which properties, here accurately predicting the band gap and dielectric constant for any polymer.
(Source: Kim, Ramprasad Lab/University of University of Conn)
First they analyzed 283 polymers using these building blocks using those time-consuming quantum-mechanical calculations, then they quantified the atomic-level configurations of these polymers using machine learning to arrange them into groups with like properties. Then he constructed the free online program that uses Khazana to predict the electronic properties of polymers from their formula, or the formula of a polymer from the desired electronic characteristics.
"And at the end of the day we can predict the electronic qualities of a polymer, or the polymer that has a desired set of electronic properties," Ramprasad told us.
Polyurea represented here where N is nitrogen, H is hydrogen, O is oxygen and R stands in for any number of chemicals that could slightly alter the polymer with the repeating NH-O-NH-O is the basic structure. Most polymers look like that, made of carbon (C), H, N and O, with a few other elements thrown in occasionally.
(Source: Yikrazuul, public domain)
There are many more details about how he divided the polymers into groups by their formula, first by component parts, then by pairs and finally by triplet combinations contained in each formula. He also has performed similar prediction engines for the band gap of perovskites—inorganic compounds often found in solar cells, lasers, and light-emitting diodes.
Read about all the details in Machine Learning Strategy for Accelerated Design of Polymer Dielectrics, From Organized High-Throughput Data to Phenomenological Theory using Machine Learning: The Example of Dielectric Breakdown and A polymer dataset for accelerated property prediction and design.
Next Ramprasad plans to expand beyond purely organic polymers to organo-metallic (polymers with inorganic blocks) which represent a much larger space of materials, as well as adding another electronic property—the breakdown voltage at which insulator begin tunneling—aiming to fulfill his part of Obama's dream of decoding the genome of materials in general.
Contributing to the work by validating many of the predictions of his home-grown learning algorithm were University of Connecticut professors Greg Sotzing, Yang Cao. Then Sotzing made several of the most novel polymers, which Cao verified Ramprasad's predictions.
Funding was provided by a variety of organizations including the Office of Naval Research and the U.S. Department of Energy.
— R. Colin Johnson, Advanced Technology Editor, EE Times