Machine learning algorithm identifies salts from their drying patterns

Scientists have trained a machine learning algorithm to identify the chemical composition of different salt solutions from the patterns formed when they dry.

The technique could become a rapid and cheap way to analyse mystery substances, including suspected drugs.

The work appears in the journal Proceedings of the National Academy of Sciences.

“We are taking chemical fingerprints of different salts,” says Oliver Steinbock, a professor of chemistry at Florida State University in the US.

An image of an iris-shaped salt stain against a black background.
A microscope image of sodium sulphite. Credit: Oliver Steinbock

“Thinking of sodium chloride, or table salt, for example — among all samples of this type, they always look similar. There are differences from sample to sample, but all examples are distinct enough from other types that we can tell what kind of salt it is.”

The researchers recorded 7,500 photos of 42 different types of salt stains and translated each image into 16 parameters – such as deposit area, compactness and texture.

The resulting dataset was used to train the machine learning algorithm. Once trained, the algorithm was able to correctly identify 90% of salt images that were not part of the initial dataset.

An image of a salt crystal under the microscope. It is made up of many cuboid shapes against a black background.
A microscope image of potassium chloride. Credit: Oliver Steinbock

The ability to quickly provide insight into the chemical composition of a sample from a photograph has many potential applications, such as rapid screening for suspected drugs, low-cost blood analysis in places without access to hospital, and on a rover exploring the chemistry of another planet.

“If you want to have a rough idea of what that stain or spill is on a lab bench, you might use this as a cursory, first-step analysis,” adds Bruno Batista, a senior researcher in Steinbock’s lab and the paper’s lead author.

Sign up to our weekly newsletter

Please login to favourite this article.