Machine learning has emerged as a powerful tool in materials research, aiding in the discovery of composition-processing-properties relationships. However, challenges remain, primarily the limited availability of large datasets for specific materials tasks. Traditional machine learning methods often rely on manually engineered features, a process that is time-consuming and requires significant domain expertise. Finding the optimal set of descriptors is often done through trial and error, hindering efficient materials development. This paper highlights the need for general and transferable machine learning frameworks that can effectively utilize limited data and existing models, incorporating domain expertise. Transfer learning, a technique that leverages knowledge sharing between related domains, combined with the automatic feature extraction capabilities of deep learning, offers a promising solution. Predicting the phases of materials, including crystalline structures (BCC, FCC, HCP), intermetallics, amorphous phases, and phase mixtures, is a fundamental challenge in materials science. This research focuses on developing a deep learning framework to tackle these challenges for materials with limited data and unclear transformation mechanisms.

Existing literature demonstrates the application of machine learning in various materials science tasks, such as predicting compound formation energies, superconductor critical temperatures, and alloy phases. However, most studies suffer from limitations like reliance on small datasets and manual feature engineering. Previous work on predicting glass-forming ability (GFA) and high-entropy alloy (HEA) phases often employs empirical thermo-physical parameters, the CALPHAD method, and first-principles calculations, along with conventional machine learning techniques. These approaches often struggle with limited data and the complexity of the underlying physical mechanisms. The authors reference several key papers that highlight both the successes and limitations of prior machine learning approaches in materials science, underscoring the need for a more general and transferable framework.

The proposed GTDL framework consists of three main stages: data representation, feature extraction, and knowledge transfer. First, raw data (composition and processing parameters) are mapped to pseudo-images using a 2D structure, specifically a periodic table representation (PTR). This representation embeds domain knowledge directly into the input data. A VGG-like convolutional neural network (CNN) is then employed for automatic feature extraction. The CNN architecture is designed to be compact, minimizing the risk of overfitting to limited datasets. The well-trained feature extractors (convolutional layers) are then reused (transfer learning) for new, related tasks with smaller datasets. Different data representations were explored, including PTR, randomized PTR, and atom table representation, to evaluate the impact of domain knowledge on model performance. The authors compared the GTDL framework with conventional machine learning models using manual feature engineering, demonstrating the advantages of automatic feature extraction and knowledge transfer. For glass-forming ability prediction, a dataset of over 10,000 entries was used, including data from various sources. A binary classification approach (amorphous/crystalline) was adopted to address data imbalance issues. For high-entropy alloy phase prediction, a smaller dataset of 355 HEAs was utilized, employing transfer learning from the GFA model. The performance of various models was evaluated using metrics such as accuracy, precision, recall, and F1-score, with cross-validation techniques employed to ensure robustness.

The GTDL framework, particularly the model using PTR (CNN3), significantly outperformed conventional machine learning models in predicting glass-forming ability (GFA). CNN3 achieved a testing accuracy of 96.3% in 10-fold cross-validation, surpassing other models that relied on manual feature engineering. Leave-one-system-out cross-validation on 160 ternary systems further demonstrated the superior generalization ability of CNN3, showing a 7% accuracy improvement over other models. The framework also accurately predicted the GFA of newly reported alloy systems not included in the training data, highlighting its predictive power. In the HEA phase prediction task, transfer learning from the well-trained GFA model (CNN3) achieved remarkable results, distinguishing five types of phases (BCC, FCC, HCP, amorphous, and mixtures) with accuracy, precision, and recall above 94% in fivefold cross-validation. This demonstrates the effectiveness of transferring knowledge learned from a larger dataset (GFA) to a smaller dataset (HEA). Feature analysis using principal component analysis (PCA) revealed that the PTR model effectively captured periodic trends of elements' properties, suggesting that embedding domain knowledge (periodic table structure) significantly enhanced model performance, especially with limited data. The randomized PTR and atom table representations did not show the same level of performance, underscoring the importance of incorporating the periodic table structure into the data representation.

The results demonstrate the significant advantages of the GTDL framework. The use of PTR and transfer learning proves particularly beneficial when dealing with small datasets. The superior performance of CNN3, which incorporated periodic table knowledge, highlights the importance of integrating domain expertise into machine learning models. The successful transfer learning from GFA prediction to HEA phase prediction showcases the framework's generalizability and potential for application across various materials problems. The automatic feature extraction capabilities of the CNN reduced the reliance on manual feature engineering, simplifying the process and improving efficiency. The findings suggest that integrating domain knowledge into data representation is crucial for achieving high accuracy and generalizability, especially when working with limited data. The framework could be adapted to other materials domains by modifying the 2D representation (e.g., crystal structure instead of periodic table).

This research successfully developed a general and transferable deep learning framework for predicting phase formation in materials. The GTDL framework, using a periodic table representation and transfer learning, significantly improved the prediction accuracy and generalizability compared to conventional methods, especially when dealing with small datasets. The framework's ability to leverage domain knowledge and transfer learned features between related materials problems positions it as a powerful tool for accelerating materials discovery and design. Future research could explore the application of this framework to other materials systems and expand its capabilities by incorporating additional data sources and improving the data representation schemes.

While the GTDL framework demonstrated significant improvements, some limitations exist. The reliance on specific data representations (e.g., PTR) might limit its applicability to materials systems where such representations are not readily available or suitable. The generalization ability of the model might be affected by the distribution of data in the training dataset, requiring careful data curation. Furthermore, the interpretability of the learned features could be enhanced through further analysis and visualization techniques. Future work could address these limitations by exploring more flexible data representation methods and developing more robust feature interpretation strategies.