“Democratic Chemistry Analysis”: Using machine learning and robotics to identify chemical compositions from images

A chemist at Florida State University has created a machine learning tool that can identify the chemical composition of a dry salt solution from an image with 99% accuracy.

By preparing thousands of samples and artificial intelligence using robotics to analyze the data, we created a simple and inexpensive tool that expands the possibilities of performing chemical analysis. This work is featured in Digital Discovery.

“We live in an age of artificial intelligence and big data,” says Oliver Steinbock, a professor at the FSU Department of Chemistry and Biochemistry. “We thought that if we had a large enough database with sufficient photographs of different compounds and stains, we could use AI to determine the composition.”

This research could enable cheaper and faster chemical analysis, which can be used in space exploration, law enforcement, and home testing.

This paper builds on previous research from Steinbock’s lab, where researchers used machine learning to identify the chemical composition of salt staining from photographs. In that study, the researchers analyzed approximately 7,500 samples, which were prepared manually.

In this paper, the behavior was amplified by processing the samples using a robot and later analyzed by an improved machine learning program. Instead of hand pipe samples, researchers created a robotic drop imager, or something called Rodi, which can prepare more than 2,000 samples per day. This allowed us to build a library of over 23,000 images. This is more than three times the size of the original study.

After preparing samples and taking photos, the researchers simplified each image by converting them to grayscale, then extracted 47 features such as patterned area, brightness and other attributes used in the analysis.

Additional images increased the accuracy of machine learning programs from about 90% to almost 99%. Researchers also analyzed initial concentrations of salt solutions at five different levels and trained machine learning programs to distinguish them. This program reached 92% accuracy in identifying solution concentration and salt identity.

“The accuracy required for different analyses varies from situation to situation,” says postdoctoral researcher Amrutha SV. “From my experience, I know that some types of spectroscopy and other analytical methods are expensive and require special technical expertise to operate. So I’m excited by the possibility of a simple method. I decide the chemical composition to take a photograph. It’s very useful.”

Most chemical analytical methods focus on the molecular level, examining atoms, molecules or crystal structures.

“If you have hundreds of thousands of dollars for a good sample, instrument, it works well and there is no weight limit,” Steinbock said. “But if you go on a space mission and want to ship things to the moon of Saturn, where every gram is important. If you can do chemical analysis with the camera, that’s a game changer.”

The project was developed for NASA, who was looking for a cheap, low-cost, low-weight method for determining chemical concentrations. Instead of transporting samples to Earth, an extraterrestrial rover with a simple chemical lab and camera can analyze the chemical composition of the materials in the field.

In addition to space exploration, methods developed in Steinbock’s lab can be used to provide chemical analysis for other applications. This test relies on fine sample volumes (a few milligrams) that are valuable in scenarios where it is difficult to obtain large samples in just a few milligrams. Law enforcement agencies will conduct preliminary tests on suspicious drugs, allowing the lab to test spilled materials for safety, and it can be used by hospitals to assist patients in diagnosing without access to a complete chemical analysis lab.

“This is important because it can democratize chemical analysis,” Steinbock said.

AI: A new tool for research

Artificial intelligence promises to change what is possible in research. Facility at Florida State University is engaged in innovative projects that push the boundaries of this rapidly developing tool.

FSU’s AI initiatives provide tools and insights to faculty in education and research.

“I think being in a place where you get this kind of support is very helpful. It doesn’t necessarily have to be money, but I’m just grateful to try something new,” Steinbock said. “AI is changing the way we approach scientific discovery. What once required expensive equipment and specialized expertise can be done with simple cameras and appropriate algorithms. This opens up new possibilities not only for space missions, but for medicine, forensics, etc.”