Abstract
Computational modeling of high entropy alloys (HEA) is challenging given the scalability issues of Density functional theory (DFT) and the non-availability of Interatomic potentials (IP) for molecular dynamics simulations (MD). This work presents a computationally efficient IP for modeling complex elemental interactions present in HEAs. The proposed random features-based IP can accurately model melting behaviour along with various process-related defects. Predicted atomic forces are within 0.08 eV/$\unicode{xC5}$ of corresponding DFT forces. MD simulations predictions of mechanical and thermal properties are within 7% of the DFT values. A novel sparse model is also proposed which reduces the computational cost by 94% without compromising on the force prediction accuracy.
Introduction
High Entropy Alloys (HEAs), characterized by their multi-element composition in near-equal proportions, exhibit superior mechanical and thermal properties. However, their complexity makes them difficult to model. While Density Functional Theory (DFT) provides high accuracy, its computational cost scales poorly with system size. Conversely, classical interatomic potentials often lack the fidelity to capture nuanced multi-element interactions. This paper introduces a machine-learning-based IP using Random Fourier Features (RFF) and Sparse Bayesian Learning to achieve DFT-level accuracy at a fraction of the computational cost.
Methods
The authors propose a Generalized Linear Model based on Random Fourier Features (RFF). The approach involves:
- Kernel Approximation: RFF transforms non-linear atomic interactions into a linear regression problem in a high-dimensional feature space.
- Orthogonal Features: The use of orthogonal random numbers ensures accuracy with a reduced number of features.
- Sparse Bayesian Learning: Utilizing the Relevance Vector Machine (RVM), the model is pruned to its most essential parameters, significantly enhancing computational efficiency.
- Descriptors: Simple two-body and three-body descriptors are employed to represent atomic environments efficiently.
Datasets
The model was trained and validated using a comprehensive dataset generated via DFT (VASP). This dataset includes ground states, strained geometries, and melting snapshots for both individual elements (Fe, Ni, Cu, Co, Cr) and the FeNiCuCoCr (FNCCC) alloy system.
Results and Discussion
The RFF potential demonstrates exceptional performance in capturing the physics of HEAs. The model achieved a force prediction accuracy within 0.08 eV/$\unicode{xC5}$ of DFT benchmarks, as shown in the fitting and error distribution plots.
A major breakthrough of this work is the "SparseRFF" model. By applying RVM, the researchers reduced the model parameters from 15,000 to 906—a 94% reduction—without compromising accuracy. This sparsity leads to a 94% reduction in FLOPS, enabling large-scale MD simulations of millions of atoms. The potential also accurately predicted the melting range of the FNCCC alloy (1500–2000K) and reproduced critical defect properties such as Generalized Stacking Fault Energy (GSFE).
Conclusion
The integration of Random Fourier Features and Sparse Bayesian Learning provides a robust framework for developing interatomic potentials for complex alloys. This approach successfully balances the trade-off between accuracy and scalability. Future research will focus on automating hyperparameter tuning and optimizing the data generation process through active learning.
Summary updated