Manufacturing Bits: June 29

Speeding up ALD with AI; discovering molecules; natural drugs.


Speeding up ALD with AI
The U.S. Department of Energy’s (DOE) Argonne National Laboratory has developed various ways to make atomic layer deposition (ALD) more efficient by using artificial intelligence (AI).

ALD is a deposition technique that deposits materials one layer at a time on chips. For years, ALD has been used for the production of DRAMs, logic devices and other products.

In operation, wafers are inserted in a chamber within an ALD system. A chemistry or precursor is pumped into the chamber. The wafers are processed. Then, the chemistries are purged from the system. The process is repeated, sometimes using a different chemistry.

ALD, however, is a slow process. On top of that, each process requires different materials and films. And adjusting the process for each new material can take time, according to researchers from Argonne National Laboratory.

There are other factors involved. Each chemistry is complex with different variables. There are several suppliers of ALD systems in the market. Each one may have a different reactor design with different settings and conditions. In some cases, vendors must use a time-consuming trial and error process to identify the optimal conditions.

In response, Argonne hopes to make ALD more efficient. Researchers evaluated three “optimization strategies” for ALD–random, expert system and Bayesian optimization. The latter two makes use of different machine learning approaches. A form of AI, machine learning is a neural network that crunches data and identifies patterns. It then matches certain patterns and learns which of those attributes are important.

Researchers evaluated the three strategies by comparing the dosage and purge times of the two precursors used in ALD. “Dosage time refers to the time period when a precursor is added to the reactor, while purge time refers to the time needed to remove excess precursor and gaseous chemical products,” according to researchers from Argonne.

All three “optimization approaches” accelerated the ALD process. But the two AI approaches effectively determined the optimal dose and purge timings for different simulated ALD processes using a closed-loop system.

“All of these algorithms provide a much faster way of converging to optimum combinations because you’re not spending time putting a sample in the reactor, taking it out, doing measurements, etc., as you would, typically. Instead, you have a real-time loop that connects with the reactor,” said Angel Yanguas-Gil, a principal materials scientist at Argonne.

“In a closed-loop system, the simulation performs an experiment, gets the results, and feeds it to the AI tool. The AI tool then learns from it or interprets it in some way, and then suggests the next experiment. And this all happens without human input,” said Noah Paulson, a computational scientist at Argonne.

“This is exciting because it opens up the possibility of using these types of approaches to rapidly optimize real ALD processes, a step that could potentially save manufacturers precious time and money when developing new applications in the future,” concluded Jeff Elam, a senior chemist at Argonne.

Discovering molecules
Researchers from Carnegie Mellon University and the St. Petersburg State University created an algorithm called MolDiscovery that would help scientists categorize unknown molecules using mass spectrometry data.

MolDiscovery would help save time and money as scientists wouldn’t spend resources studying already identified molecules.

Mass spectrometry is a technique that measures the mass-to-charge ratio of a molecule, giving information on the atomic mass of the molecule. The same element may have different atomic masses, due to varying numbers of neutrons in the nucleus, so the atomic mass of the molecule is displayed in a spectrum.

However, within one environment, there may be hundreds of thousands of molecules, and identifying individual unknown molecules is a challenge.

MolDiscovery is an algorithm that searches through millions of molecules’ data using a mass spectral database search method. Using the mass spectra data of a molecule, MolDiscovery uses a probabilistic model to predict what makes up the molecule. By looking at the molecule’s structure, MolDiscovery constructs metabolite graphs and breaks a molecule down to fragmentation graphs. Knowing the mass spectra and graphs of the molecule, the algorithm can predict the molecule to see if it’s been previously discovered.

MolDiscovery is more efficient and accurate compared to previous molecule identification processes. “The existing approaches are based on chemistry domain knowledge, and they fail to explain many of the peaks in mass spectra of small molecules. A search of over 8 million spectra from the Global Natural Product Social molecular networking infrastructure shows that MolDiscovery correctly identify six times more unique small molecules than previous methods,” writes Hosein Mohimani, an assistant professor and research contributor from Carnegie Mellon University, in a paper published in Nature Communications. Others contributed to the paper.

MolDiscovery can help scientists and researchers in the medical and pharmaceutical industries, as well as in the discovery of other novel natural products. By identifying new molecules early on, scientists can save money in developing new drugs, characterizing microbes, and help with disease diagnosis.

Natural drugs
Using a machine learning technology, Carnegie Mellon has found a way to accelerate the development of natural medications to treat cancer, viral infections and other ailments.

Others contributed to the work, including the University of California at San Diego, Saint Petersburg University, Max-Planck Institute, Goethe University, the University of Wisconsin at Madison, and the Jackson Laboratory.

Many antibiotics, antifungal and antitumor medications have come from natural products. They are considered safe, so it’s imperative to accelerate the development of these natural drugs.

In response, researchers have developed a machine learning algorithm, called NRPminer, which is a platform that helps scientists isolate natural products. More specifically, NRPminer speeds of the discovery of non-ribosomal peptides (NRPs), an important type of natural product used to make many antibiotics, anticancer drugs and other medications.

NRPs are difficult to detect and identify. “Natural products are still one of the most successful paths for drug discovery,” said Bahar Behsaz, a project scientist at Carnegie Mellon. “And we think we’re able to take it further with an algorithm like ours. Our computational model is orders of magnitude faster and more sensitive.”

“What is unique about our approach is that our technology is very sensitive. It can detect molecules with nanograms of abundance,” said Carnegie Mellon’s Mohimani said. “Our hope is that we can push this forward and discover other natural drug candidates and then develop those into a phase that would be attractive to pharmaceutical companies.”

Leave a Reply

(Note: This name will be displayed publicly)