The next phase: Using neural networks to identify gas-phase molecules

September 13, 2018, Argonne National Laboratory
This schematic of a neural network shows the assignment of rotational spectra (red bars at left) by an algorithm (center) to identify the structure of a molecule in the gas phase (right). Credit: Argonne National Laboratory

Scientists at the U.S. Department of Energy's (DOE) Argonne National Laboratory have begun to use neural networks to identify the structural signatures of molecular gases, potentially providing new and more accurate sensing techniques for researchers, the defense industry and drug manufacturers.

This breakthrough work has been recognized as a finalist for a 2018 R&D 100 award.

"This means that when you're at an airport running a security test on an unidentified chemical or if you're a scanning your sample for impurities, you can run so many more of these tests accurately in a much smaller period of time."—Daniel Zaleski, Argonne postdoctoral researcher

Neural networks—so named because they operate in an interconnected fashion similar to our brains—offer chemists a major opportunity for faster and more rigorous science because they provide one way in which machines are able to learn and even make determinations about data. To be effective, though, they have to be carefully taught. That is why this area of research is called machine learning.

"Say you wanted to teach a computer to recognize a cat," said Argonne chemist Kirill Prozument. "You can try to explain to a computer what a cat is by using an algorithm, or you can show it five thousand different photos of cats."

But instead of looking at cats, Prozument and former Argonne postdoctoral researcher Daniel Zaleski wanted to identify the structure of gas-phase . To do so, they used the molecules' rotational spectra.

Scientists determine a molecule's rotational spectra by observing how the molecule interacts with electromagnetic waves. In classical physics, when a wave of a particular frequency hits a molecule in the gas phase, it causes the molecule to rotate.

Because molecules are quantum objects, they have characteristic frequencies at which they absorb and emit energy that are unique to that type of molecule. This fingerprint gives researchers an excellent idea of the pattern of quantum energy levels of gas-phase molecules.

"We're particularly interested in looking at the products that result from chemical reactions," Prozument said. "Suppose we don't know what chemical products we've generated, and we don't know what molecules there are. We sweep with a millimeter-wave pulse through all possible frequencies, but only frequencies that 'ring the bell' for the molecules will be absorbed and only those will be re-emitted."

Zaleski coded thousands of these rotational spectra, labeling each different spectrum for the neural network. The advantage of using a neural network is that it only had to "learn" these spectra once, as opposed to each time a sample was tested.

"This means that when you're at an airport running a security test on an unidentified chemical or if you're a drug manufacturer scanning your sample for impurities, you can run so many more of these tests accurately in a much smaller period of time," Zaleski said. Even though these resonances act as a filter, the amount of spectroscopic data produced is still daunting. "Going from raw spectroscopic data to actual chemical information is the challenge," Zaleski said. "The data consist of thousands if not tens of thousands of elements—it's messy."

Zaleski, now an assistant professor at Colgate University, compared the search for specific molecular signatures to the children's picture book "Where's Waldo?", in which the reader has to scan a crowded scene to find the titular character. "Waldo has a very specific dress and a specific pattern, so you'll know him if you see him," Zaleski said. "Our challenge is that each molecule is like a different version of Waldo."

According to Zaleski, there are fewer than 100 scientists in the world trained in assigning rotational spectra. And while it could take up to a day to determine the molecular signatures using previous methods, reduce the processing time to less than a millisecond.

The neural network runs on graphics processing unit (GPU) cards typically used by the video gaming community. "Until a couple of years ago, the GPU cards we're using just didn't really exist," Zaleski said. "We are in an amazing time right now in terms of the computing technology available to us."

Ultimately, Prozument and Zaleski hope to make their spectroscopic technique as fully automated as possible. "Our goal is to offer the tools of rotational spectroscopic analysis to non-experts," Prozument said. "If you can have spectra accurately assigned by a machine that can learn, you can make the whole process much more portable and accessible since you no longer need as much technical expertise."

An article based on the study, "Automated assignment of rotational spectra using artificial neural networks," appeared in the September 13 issue of the Journal of Chemical Physics.

Explore further: Test tube artificial neural network recognizes 'molecular handwriting'

Related Stories

Artificial intelligence for obtaining chemical fingerprints

September 26, 2017

Researchers at the Universities of Vienna and Göttingen have succeeded in developing a method for predicting molecular infrared spectra based on artificial intelligence. These chemical "fingerprints" could only be simulated ...

The molecular structure of a forest aroma deconstructed

December 7, 2017

The fresh, unmistakable scent of a pine forest comes from a medley of chemicals produced by its trees. Researchers have now, for the first time, accurately determined the chemical structure of one of these compounds in its ...

SMART: Facial recognition for molecular structures

October 31, 2017

An interdisciplinary team of researchers at the University of California San Diego has developed a method to identify the molecular structures of natural products that is significantly faster and more accurate than existing ...

A common electronic language for magnetic resonance

June 7, 2018

Scientists working in the field of organic chemistry create and study new molecules using magnetic resonance. The standards used to re-transcribe the collected data is, however, specific to each laboratory or publication, ...

Recommended for you

New targets in the battle against antibiotic resistance

November 16, 2018

Bacteria are increasingly resistant to available antibiotics. A team of chemists from the Technical University of Munich (TUM) have now identified important enzymes in the metabolism of staphylococci. Blocking these enzymes ...

AI heralds new frontiers for predicting enzyme activity

November 16, 2018

Researchers from the Departments of Chemistry and Engineering Science at the University of Oxford have found a general way of predicting enzyme activity. Enzymes are the protein catalysts that perform most of the key functions ...

6 comments

Adjust slider to filter visible comments by rank

Display comments: newest first

Hyperfuzzy
not rated yet Sep 13, 2018
Really? Ya got more work to do. You can add energy Lambda by Lambda+Lambda; incrementally. Choose a source; stay in the lines; but I think ya need a better model! For instance, reality shows only two centers! So your whole model is only a single point, well, ya might need two, keep going until it makes sense; juz N pairs; watch out for those pairs hiding with each other. remember the all respond and trace the response as equal potential surfaces about a point. They all look alike, so i think u git tha point! So I thank ya lazy!
Hyperfuzzy
not rated yet Sep 13, 2018
Don't 4git ta look at stuff from every angle, literally; boundary conditions. Science is not sloppy; base you models always on reality! Any simplification, i.e. superimposed field at each point! Nothing is unknown; only the arrangement, blah, blah
Hyperfuzzy
not rated yet Sep 13, 2018
Try a 4D unit-less space; c=1; T = Lambda then [Set] => Measured&Predicted+Anaolmy!
Hyperfuzzy
not rated yet Sep 13, 2018
Gee how many atoms do ya need to see it in your head, orbiters(-) avoid each other; but, ya got every point n space and time in ya model So ya see exactly what is happening, right or wrong. ie. tune ya sw. pretty simple, +/-1/Si. for each ... & c= 1 Then S[1 ]=4pi Lambda^2; ... it ain't hard
Parsec
not rated yet Sep 18, 2018
this application is a perfect place where machine learning can really shine. Not only is every molecules spectra unique, but so its each portion. That is, the parts combine to make the whole.

For example, the spectra of H2O2 combines the spectra of H2O, HO, HO2, H, and O where the entire spectra is also modified by the combination (technically neighbors influence each other). Machine learning could thus be used to identify unknown spectra from simpler ones.
Hyperfuzzy
not rated yet Sep 18, 2018
OK, 4git what i said 'bout being able to explicitly design with logic and causal physics, i.e. drop the nonsense talk

Please sign in to add a comment. Registration is free, and takes less than a minute. Read more

Click here to reset your password.
Sign in to get notified via email when new comments are made.