The economic damage caused by corrosion comes to $2.5 trillion a year on a global scale. Researchers and industrial companies have always been on the lookout for new, corrosion-resistant alloys and coatings that protect alloys from corrosion. Artificial intelligence (AI) is increasingly being used in this search effort to predict the corrosion characteristics of different materials with a view to determine the optimum alloy composition.
However, the predictive abilities of existing AI models have been limited, as it has not been possible to factor in all the relevant data. Researchers from the Düsseldorf-based Max Planck Institute for Iron Research (MPIE) have developed a new machine learning model that can predict corrosion-based defects 15% more accurately than existing models and identify new resistant alloys. Originally developed for the critical area of pitting corrosion in high-strength alloys, the model can be applied to all alloy properties. The researchers have now published their results in the journal Science Advances.
Text and numbers in an AI model
“Each alloy’s resistance to corrosion depends on its composition as well as its production and processing. However, to date, AI models have only been able to evaluate the composition based on numerical data. But as the production and processing of the alloy are documented in writing, this data was not fed into AI models. As a result, existing AI models have been of limited informative value,” explains Dr Kasturi Narasimha Sasidhar, who was first author of the publication and former post-doctoral researcher at MPIE.
The team of researchers uses language processing methods similar to ChatGPT and combines them with machine learning (ML). In doing so, the MPIE researchers were able to develop a machine learning model that processes both numerical data and natural language in a fully automatic process. It can now better predict how alloys behave during corrosion and which alloys are corrosion-resistant.
“At the outset, we trained the learning model with data on corrosion properties and alloy composition. Now the model is able to recognise corrosion-resistant alloys on its own, even if the individual factors were not originally input into the model,” says Dr Michael Rohwerder, a co-author of the publication and head of the "Corrosion" group at the Max Planck Institute for Iron Research.
Looking ahead: automated data mining and image processing
Until now, the AI model has been based on data manually collected by the researchers. Their goal now is to automate the process of data mining and incorporate it seamlessly into their model. In addition, they are also aiming to use the model on microscopic images so that all relevant sources of information, text, figures and images are fed into the AI model, further improving its informative value.
