New Machine Learning Model for Combating Corrosion

Machine Learning Model for Corrosion Prevention Scientists at Max-Planck-Institut für Eisenforschung have developed a machine learning model that increases the prediction accuracy of corrosion behavior in alloys by up to 15% compared to existing methods. The model, developed by the Max Planck Eisenforschung Institute team, reveals novel yet realistic corrosion-resistant alloy compositions. The model's distinguishing feature is that it combines both numerical and textual data. Although initially designed to combat pitting corrosion in high-strength alloys, the model can be applied to various other alloy properties. The researchers published their findings in Science Advances journal. Kasturi Narasimha Sasidhar, lead author of the study and former postdoctoral researcher at Max-Planck-Institut für Eisenforschung, stated: "Every alloy has unique properties related to corrosion resistance. These properties depend not only on the alloy composition but also on the alloy's production process. Current machine learning models can only benefit from numerical data. However, processing methodologies and experimental test protocols, which are mostly documented with textual descriptors, are highly important in explaining corrosion." The research team used natural language processing methods similar to ChatGPT combined with machine learning (ML) techniques for numerical data, developing a fully automated natural language processing framework. Additionally, the incorporation of textual data into the machine learning framework enables the identification of advanced alloy compositions resistant to pitting corrosion. Michael Rohwerder, co-author of the publication and Head of the Corrosion Group at Max-Planck-Institut für Eisenforschung, stated: "We trained the deep learning model with proprietary data containing information about corrosion properties and composition. The model can now identify alloy compositions critical for corrosion resistance even when individual elements were not initially included in the model." In the recently developed model, Sasidhar and his team used manually collected data as textual descriptors. Their current objective is to automate the data mining process and seamlessly integrate it into the existing framework. The inclusion of microscopy images marks another milestone that will enable next-generation artificial intelligence models combining textual, numerical, and image-based data. Source