Phase prediction in high-entropy alloys with multi-label artificial neural network

Data-driven analysis and prediction of stable phases for high-entropy alloy design

“…It is noteworthy that the BCC structure is the most frequently mentioned in the literature, closely followed by the FCC structure. Additionally, the dual-phase structure combining both FCC and BCC is also commonly discussed 55 , 57 , 89 , 95 . Although the dataset includes alloys with varying numbers of components, ranging from three to ten, the majority (93%) of the data falls within the range of four to six components.…”

Microstructural Characterization and Phase Prediction of High Entropy Alloys using Empirical and Machine Learning Based Methodologies

“…At the time of writing, the vast majority of neural network-based ML models used for HEA phase prediction consist of FFNN utilizing either empirical parameters [14] or the concentrations of selected alloying elements [124], [125] or combination of the two [126]. With regards to empirical parameters, commonly used ones tend to be similar to those utilized in the parametric approach and thus are limited by the scope of the parameters.…”

“…Generally, machine learning based prediction methods for HEAs, a sample of which is shown in Table 2-1, achieve accuracy within the 80-90% range, with the best performing models achieving accuracies in the low to mid 90's; these models utilize more advanced/complex methodologies such as generative adversarial networks (GAN) and machine learning optimization of input parameters. Most model predictions are classification based and are therefore limited to 2-4 classes; however, the labelling strategy has seen some recent success in expanding the scope of possible predicted phases without overly diluting the number of datapoints per label [118], [125].…”

“…To date, much of the research surrounding machine learning, specifically neural networks, for phase prediction of HEAs relies entirely on parameters calculated from the elemental composition of HEAs as input parameters, without directly considering the elemental concentrations [133], or only specific elements and their concentrations [124], [125] and make use of the feed forward neural networks (FFNN). In the case of only using calculated parameters, this input structure could limit the model's ability to account for specific interactions that are not already accounted for by the chosen parameters.…”

“…Even if element concentrations of alloying elements are considered, the nature of the input structure of FFNNs restricts the possible inputs to only elements hard coded into the model [134], and each additional element added to the model would inevitably increase the dimensionality of the input and necessitate changing the input structure. For example, the work presented in [125] found that utilizing an input structure based on the concentrations of 16 alloying elements found in the training dataset outperformed an input structure consisting of various empirical parameters, and that a hybrid approach combining the two inputs performed the best. However, this approach of a FFNN limits the model to only HEAs containing those from the 16 elements.…”