When salt solutions evaporate, deposits of intricate and often beautiful crystallisation patterns are formed which may seem random and unpredictable. But now a machine learning technique has shown how images of such patterns from aqueous inorganic salt deposits can be used to predict their chemical composition. The approach could find various uses from space exploration to deposit-identifying smartphone apps.


Source: © 2024 Bruno C Batista et al

When inorganic salts crystallise they form a range of different patterns depending on their composition. This ‘fingerprint’ has been used for the first time to distinguish one salt from another

‘The processes that form deposit patterns during drop drying are very complicated and computing the diverse stain patterns from just the crystallising salt type is challenging, perhaps impossible,’ says Oliver Steinbock, whose lab conducted the work at Florida State University, US. ‘We looked at this problem in the opposite direction and asked whether it is possible to find the composition solely from a photo of the drop stain.’

To find out, the researchers first took 7500 photos of 42 different types of inorganic salt stains. They then used software they developed that translated each image into a set of numbers, which corresponded to 16 parameters, including deposit area, texture and other geometric features. This meant that each photo was represented as a point in a 16-dimensional space. A machine learning algorithm was then able to quickly compare their symmetry in this space and build a fingerprint profile for each one. This revealed a salt-based family tree that could be used to accurately identify a salt deposit 90% of the time.

‘Crystal structures are complicated and our salt stains are kind of messy so we were amazed at how well our method was able to cut through this and extract compositions from photos alone,’ says Steinbock. ‘Who would think that from a photo you can tell the difference between sodium chloride and potassium chloride? They look very similar in the pictures. But the method is very good.’


Source: © 2024 Bruno C Batista et al

A wide range of inorganic salts were painstakingly dried and photographed to provide the necessary dataset for the machine learning program

Part funded by Nasa, the work could have uses in space exploration. ‘This approach would benefit weight-restricted space missions aiming to identify brines on Mars or ocean waters on Europa and other moons,’ explains Steinbock. ‘We also envision smartphone-based applications in fields ranging from environmental science and forensics to home and lab safety.’

Preparing 7500 drops, letting them dry on glass slides and then taking photos was monotonous and time-consuming, Steinbock notes. But the researchers have since built a robotic system capable of doing this faster, allowing them to collect about 1000 photos daily. ‘Having a large database of reference photos will surely provide new insights. Our study already showed unexpected similarities between different salts – a hidden family tree so to speak,’ says Steinbock.

‘In the field of crystallisation, we often say morphology does not equal mechanism, meaning that you cannot a priori determine the pathways of crystallisation based on visual appearance of the final product,’ says Jeffrey Rimer, a crystallisation expert at the University of Houston, US. ‘This fascinating study shows there is a caveat to this axiom pertaining to the use of morphology as a “fingerprinting” method to predict the composition of inorganic salt crystals. This unexpected outcome using a simple, yet highly elegant approach is exciting and has significant potential to accelerate materials diagnostics.’

‘We also think that the salt stains have their own intricate beauty,’ says Steinbock. His lab has since created an online image collection called Saltscapes. ‘They remind me of the 19th-century German chemist Runge, who published the book Bilder, die sich selber malen (Pictures that paint themselves) based on patterns formed by chemical reactions and diffusion on filter paper,’ Steinbock adds.