Only seemingly random patterns: A study shows the possibility of identifying salt substances based on the stains they leave behind after drying. Based on the subtle pattern similarities of the crystal structures, the artificial intelligence system has so far been able to assign 42 different types of salt stains with high accuracy. This “fingerprint” approach could develop into a streamlined analysis method that can provide basic information about materials, for example when exploring celestial bodies, say the researchers.
We all know it from the kitchen: a drop of salted pasta cooking water leaves a whitish stain on the counter after it dries. If you look at it with a magnifying glass, you can see something that looks like an abstract work of art: delicate structures that have been created by the crystallization of sodium chloride. Each drop forms its own individual pattern. Complex fluid movements and a range of environmental factors influence the respective deposit pattern. However, crystal growth is also influenced by the specific characteristics of the respective substance. However, this does not make it possible to easily distinguish between the many types of salt: at least not at first glance, the relic of a sodium chloride solution cannot be distinguished from a potassium chloride stain.
Can AI develop an “expert eye”?
That is why researchers led by Bruno Batista from Florida State University in Tallahassee have now explored the extent to which artificial intelligence could recognize the subtle differences: Is it possible to determine the identity of a salt deposit based on its appearance alone? To capture the basic potential, the researchers took 7,500 detailed photos of 42 different types of salt spots. They were “grown” from tiny drops under standardized conditions on glass slides. Using a special software concept, each image was then characterized using 16 parameters. These included features such as the deposit area, compactness and texture. This in turn reflects the subtle peculiarities in the arrangement of the tiny crystals in rings, needles and leaf-like structures.
The researchers then fed and trained an AI system with this image information: Machine learning methods make it possible to uncover hidden patterns in the data. Specifically, the system should learn to recognize typical characteristics of salt stains of a certain type based on the subtle structural similarities. As the team reports, the approach was successful. This was demonstrated by their system’s ability to recognize the identity of a salt based on new stains. Despite the comparatively modest training dataset so far, the AI was able to determine the respective salt with high accuracy based on the stain photo. “We were surprised at how well it worked,” says senior author Oliver Steinbock from Florida State University. “Who would have thought that you could distinguish sodium chloride from potassium chloride from a photo? They look very similar in the pictures – but the method recognizes the difference,” says the researcher.
Promising approach
As he and his colleagues emphasize, this is a proof of concept so far: at least for pure aqueous salt solutions, it has already been shown that it is possible to identify the substance based on the appearance of microgram-sized deposits. The concept is now to be expanded significantly: the researchers plan to expand the training data set by analyzing hundreds of thousands of additional images. In addition, more compounds and substance mixtures are to be included, making the tool even more precise and versatile. However, this large number requires automation. The researchers are currently testing the use of a robotic drop imager.
They believe that the concept has considerable potential for application. Salts are very important in chemistry and nature, which makes the ability to identify them quickly and easily from a photo interesting. The method could also be used in space travel, for example. It is difficult and expensive to equip a rover that is researching an alien celestial body with a complete chemistry laboratory. For the new concept, a high-resolution camera could be enough to obtain at least basic information about the composition of sample material. Another advantage of the approach, according to the team, is the low material requirements: just a few milligrams of a salt deposit could reveal what it is.
Source: Florida State University, professional article: Proceedings of the National Academy of Sciences, doi: 10.1073/pnas.2405963121