Enhancing Corrosion Resistance Predictions of Alloys Using Process-Aware Neural Networks

The advancement of machine learning techniques has ushered in a new era of accelerated materials design and discovery. The rapid progress witnessed in this domain can be attributed to the concerted efforts of the materials research community, which has systematically organized protocols for collecting and cataloging materials data, making it amenable to machine learning analysis. In  new study led by Professor Dierk Raabe   from the Max-Planck-Gesellschaft in Germany, which was published in Science Advances. The study focuses on employing a unique combination of automated natural language processing and deep neural networks to predict the pitting resistance of corrosion-resistant alloys. This groundbreaking research holds immense promise in revolutionizing our ability to design alloys with superior corrosion resistance, thus addressing a critical challenge faced by industries worldwide.

Corrosion of metallic materials results in staggering economic losses amounting to approximately 2.5 trillion USD annually. In response to this challenge, machine learning techniques have been harnessed to predict corrosion rates and identify corrosion-resistant alloys. Previous efforts have explored the prediction of various forms of corrosion, including high-temperature oxidation kinetics, atmospheric corrosion rates, and even image-based analysis for identifying corrosion patterns. However, the complexity of corrosion-resistant alloys presents unique challenges in terms of the vast array of factors influencing their behavior, such as alloy composition, processing history, and environmental conditions. This study aims to enhance the accuracy of corrosion resistance predictions by integrating textual information about alloy processing history and test conditions into machine learning models.

The heart of this research lies in the integration of textual information with traditional numerical data in the context of deep neural networks. Conventionally, deep learning models have been limited to numerical input data. However, corrosion resistance in alloys is profoundly influenced by alloying elements, processing history, and testing procedures—information typically conveyed through textual descriptions. The researchers employ a novel approach, combining natural language processing with deep neural networks, to transform textual data into a structured numerical format that the model can understand. This process extracts valuable information from the text, such as heat treatment details, test solution conditions, and more.

The study presents a comprehensive methodology that combines multiple layers of data processing to achieve its objectives. The process begins with tokenization, wherein words from textual data are transformed into unique integer tokens. Subsequently, word embedding is employed to convert these integer tokens into vectors with semantic meaning, capturing relationships between words. To process this data, recurrent neural network (RNN) layers, specifically long-short term memory (LSTM) layers, are employed. LSTM layers excel in handling sequential data, enabling the model to comprehend the contextual nuances of the text.

The process-aware neural network model demonstrates its prowess by achieving significantly improved accuracy in predicting pitting resistance compared to traditional models. The model showcases its ability to unravel the complexities of alloy behavior by discerning the influence of various alloying elements on pitting resistance. For instance, the model uncovers the contributions of elements like Mo, C, and N in stainless steels and Ni-based alloys. These findings align with existing literature, confirming the model’s efficacy in accurately identifying mechanisms that enhance corrosion resistance.

One of the standout features of this study is its capacity to provide mechanistic insights into corrosion behavior. By analyzing keyword trends within the textual data, the researchers uncover trends that correlate with improved corrosion resistance. Elements such as atomic packing efficiency, electronegativity differences, and configurational entropy are identified as crucial factors in determining pitting resistance. This discovery opens doors for further research, shedding light on the intricate relationships between alloy composition and corrosion performance.

The utility of this approach extends beyond the confines of training data. The study demonstrates the model’s ability to predict the behavior of alloys containing elements absent from the training dataset. This remarkable capability stems from the model’s feature-transformed architecture, which allows it to extrapolate insights from known elemental properties. This advancement holds immense promise in accelerating materials discovery, enabling predictions for unexplored alloy compositions.

In conclusion, the study conducted by Professor Dierk Raabe and colleagues showcases a paradigm shift in predicting corrosion resistance of alloys. By seamlessly integrating textual information with deep neural networks, the researchers have harnessed the power of explainable AI to unravel the intricate relationships between alloy composition, processing history, and corrosion behavior. The results highlight the potential of this methodology to revolutionize materials design, paving the way for the creation of corrosion-resistant alloys with unprecedented precision and efficiency. As industries strive to mitigate the economic burden of corrosion, this research offers a beacon of hope through the fusion of advanced data analytics and materials engineering.