Machine Learning Pipeline for Novel High Entropy Carbides

In this work, we present a comprehensive pipeline for the discovery and evaluation of high-entropy carbides (HECs) with desired properties, leveraging a combination of machine learning (ML) techniques, density functional theory (DFT) calculations, and material fingerprinting. The pipeline comprises three key tasks: material discovery, synthesizability prediction, and property identification. The initial step in the pipeline involves evaluating the entropy-forming ability (EFA) descriptor, which determines whether a candidate composition can form a stable, single-phase high-entropy carbide. This descriptor is assessed using a threshold value established from prior experimental studies. If the candidate surpasses the threshold, the next step evaluates its synthesizability, determined by a disordered enthalpy-entropy descriptor that accounts for the ratio between enthalpy and entropy in disordered systems. A descriptor called compensation temperature, which is positively correlated with the critical temperature of the miscibility gap, can be derived from this. Candidates that pass both stages are classified as potential high-entropy carbides, marking the completion of the material discovery process. For predicting the entropy forming ability, we utilized a Random Forest model to analyse high-entropy carbide compositions, employing the permutation importance method to evaluate feature significance, with composition features obtained from the matminer library. This approach achieved over 95% accuracy (coefficient of determination value) in matching DFT-calculated and ML-predicted EFA values. Similarly, for synthesizability, we applied the same methodology to predict the disordered enthalpy-entropy descriptor (compensation temperature), also attaining over 95% accuracy in the match between DFT-calculated and ML-predicted values. Using the trained models for predicting EFA and synthesizability, we screened more than 19,000 new candidates for 5-element high-entropy carbides that are not included in the original dataset composed of some transition metals. More than 50 candidates meet the established thresholds, showing potential for the development of new high-entropy carbides. Together, these two metrics—EFA and synthesizability—form the basis for identifying promising high-entropy carbide candidates. In the final stage of the pipeline, we assess specific material properties by generating material fingerprints. These fingerprints are vector representations of the materials’ structural and electronic properties, extracted using a graph neural network (GNN). The GNN is trained on the features of constituent carbides. The network architecture includes graph attention layers followed by a fully connected layer, with mean pooling applied to condense the feature space into a low-dimensional vector. These vectors serve as the materials’ fingerprints, which can be used to predict and analyse a wide range of material properties. Principal component analysis (PCA) is applied to reduce the dimensionality of the fingerprints for visualization and further clustering. Fingerprints of similar materials are expected to form distinct clusters, which can be correlated with desired properties. This allows for a more targeted discovery of materials with specific functional attributes. This pipeline offers a scalable and efficient approach for the discovery of high-entropy carbides with tailored properties. By combining machine learning predictions with DFT validation and advanced fingerprinting techniques, we can rapidly screen and evaluate thousands of potential candidates, significantly accelerating the material discovery process in high-entropy systems.