Prediction of Time–Temperature–Transformation Diagrams of NiAl Alloy: An Evaluation of Intelligent Algorithms

• Published in: Advanced engineering materials 2024-11, Vol.26 (21), p.n/a
• Main Authors: Hernández‐Flores, Leonardo, Talavera‐Rivera, Luis Fernando, García‐Moreno, Angel‐Iván
• Format: Article
• Language: English
• Subjects: deep learning; machine learning; NiAl alloys; predictions; TTT diagrams
• ISSN: 1438-1656; 1527-2648
• DOI: 10.1002/adem.202400757

Time‐temperature‐transformation diagrams are essential in the field of metallurgy. However, constructing these diagrams through empirical and simulating methods can be time‐consuming and expensive. Therefore, there is a need to develop accurate and affordable alternatives that can rapidly predict TTT diagrams. The present study offers a novel analysis of multiple algorithms for predicting TTT diagrams, as well as an examination of various data preprocessing techniques (data augmentation ‐ DA and exploratory data analysis ‐ EDA). A database is created by integrating multiple data sources and increased using DA by ensuring the synthetic data remains physically realistic and meaningful. Subsequently, an EDA is performed to prepare the data before its use in training. The influence of each hyperparameter on the prediction was studied and optimal hyperparameter configuration was defined for each algorithm. The performance of the algorithms is evaluated using RMSLE, RMSE, R2, Mean IoU, and SMAPE. The multilayer perceptron demonstrated the greatest robustness in predicting TTT curves. Additionally, the Ft‐Transformer is a viable alternative if an appropriately sized dataset is available. These results provide valuable insights into the use of Machine Learning techniques as a new alternative in predicting isothermal transformation curves. This article evaluates intelligent algorithms for the prediction of isothermal transformation diagrams. A database is created by integrating multiple sources and the data augmentation and exploratory data analysis are analyzed to ensure that the synthetic data remains physically realistic. In addition, a strategy for the optimization of hyperparameters is defined. The methodology allows to achieve accuracy scores above 98%.