A Robust Machine Learned Interatomic Potential for Nb: Collision Cascade Simulations with accurate Defect Configurations
Utkarsh Bhardwaj, Vinayak Mishra, Suman Mondal, Manoj Warrier, [0.5ex] Homi Bhabha National Institute, Anushaktinagar, Mumbai, Maharashtra, India – 400094 [0.5ex] · 2025
Read A Robust Machine Learned Interatomic Potential for Nb: Collision Cascade Simulations with accurate Defect Configurations by Utkarsh Bhardwaj, Vinayak Mishra, Suman Mondal, Manoj Warrier, [0.5ex] Homi Bhabha National Institute, Anushaktinagar, Mumbai, Maharashtra, India – 400094 [0.5ex] (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-03126.txt.
Introduction
Bhardwaj and co-authors build a machine-learned model of how niobium atoms push on each other. The metal and its alloys serve in demanding technologies, and predicting radiation damage in them depends on getting one detail right: the shape a squeezed-in extra atom takes. Older potentials get that shape wrong, so the team trains a spectral neighbor analysis potential on quantum calculations.
The new potential matches the quantum reference closely while staying fast enough for large simulations. In tests of elasticity, heat behavior, and defect properties, including simulated collision cascades, it fixes the known failures of standard models. Notably, it captures the correct ground-state shape of the interstitial atom in fresh cascade damage, making it a sturdier tool for irradiation studies.
Catalog record
| CREATOR | Utkarsh Bhardwaj, Vinayak Mishra, Suman Mondal, Manoj Warrier, [0.5ex] Homi Bhabha National Institute, Anushaktinagar, Mumbai, Maharashtra, India – 400094 [0.5ex] |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 19 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-03126
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A Robust Machine Learned Interatomic Potential for Nb: Collision Cascade Simulations with accurate Defect Configurations Utkarsh Bhardwaj, Vinayak Mishra, Suman Mondal, Manoj Warrier, [0.5ex] Homi Bhabha National Institute, Anushaktinagar, Mumbai, Maharashtra, India – 400094 [0.5ex] Abstract Niobium (Nb) and its alloys are extensively used in various technological applications owing to their favorable mechanical, thermal and irradiation properties. Accurately modeling Nb under irradiation is essential for predicting microstructural changes, defect evolution, and overall material performance. Traditional interatomic potentials for Nb fail to predict the correct self-interstitial atom (SIA) configuration, a critical factor in radiation damage simulations. We develop a machine learning interatomic potential (MLIP) using the Spectral Neighbor Analysis Potential (SNAP) framework, trained on ab-initio Density Functional Theory (DFT) calculations, which accurately captures the relative stability of different SIA dumbbell configurations. The resulting potential reproduces DFT-level accuracy while maintaining computational efficiency for large-scale Molecular Dynamics (MD) simulations. Through a series of validation tests involving elastic, thermal, and defect properties—including collision cascade simulations—we show that our SNAP potential resolves persistent limitations in existing Embedded Atom Method (EAM) and Finnis–Sinclair (FS) potentials and is effective for MD simulations of collision cascades. Notably, it accurately captures the ground-state SIA configuration of Nb in the primary damage of a collision cascade, offering a robust tool for predictive irradiation studies. 1 Introduction Niobium (Nb) and its alloys play critical roles in high-temperature and radiation-intensive environments, such as nuclear reactors and aerospace industries, due to their high melting point and desirable mechanical properties [ 1 , 2 ] . Molecular Dynamics (MD) is an effective simulation method for predicting radiation damage in materials [ 3 , 4 ] . However, the accuracy of an MD simulation is limited by the interatomic potential used to calculate energies and forces. Traditional interatomic potentials (e.g., EAM, FS) for Nb have struggled to reproduce correct defect properties, especially the ground-state configuration of self-interstitial atoms (SIAs) [ 5 , 6 , 7 , 8 ] . In body-centered cubic (BCC) non-magnetic metals, the most common ground state SIA configuration is a \langle 111\rangle dumbbell, while \langle 110\rangle is less stable with higher formation energy [ 9 ] . Some group-6 metals, such as Cr, Mo, and W, may exhibit symmetry-broken SIA configurations, but for Nb, \langle 111\rangle remains consistently the lowest-energy orientation [ 10 ] . The discrepancy between different interatomic potentials regarding the morphology of defects is a well studied topic in material damage studies [ 11 , 12 ] . Furthermore, SIA morphology strongly influences elasticity, stress fields, and defect mobility, ultimately altering microstructural evolution under high-dose irradiation [ 13 , 14 , 9 ] . Consequently, accurately predicting the stable SIA orientation is crucial for reliable large-scale simulations of Nb under defect-inducing environments. Machine Learning Interatomic Potentials (MLIPs) offer a promising alternative to classical potentials by approximating the potential energy surface with high fidelity, while remaining less computationally expensive than first-principles methods [ 15 ] . MLIPs have shown success in capturing mechanical, thermal, and defect properties of transition metals more accurately than classical approaches [ 16 , 17 ] . The robustness of the MLIP needs a validation on a long simulation that involves all the complexities of the desired application. Nevertheless, despite close matches with DFT data, MLIPs can exhibit instabilities during extended MD simulations [ 18 ] . It is thus vital to validate MLIPs under application-specific scenarios, particularly those involving complex phenomena such as high-energy irradiation. In this work, we develop and extensively validate a SNAP-based MLIP [ 19 ] for Nb using a robust DFT dataset generated via VASP [ 20 ] . The subset used to train the potential is curated using diversity scoring and optimization methods [ 21 , 22 ] . Our primary goal is to accurately capture Nb’s mechanical, thermal, and defect characteristics, particularly under irradiation-induced damage. By benchmarking against well-known EAM [ 23 ] and FS [ 24 ] potentials, we demonstrate that the developed SNAP potential resolves the persistent discrepancy in existing potentials (e.g., incorrect \langle 110\rangle SIA ground state [ 5 ] ) by reproducing the DFT-observed \langle 111\rangle SIA dumbbell orientation, but also remains stable and predictive during 5 keV collision cascade simulations. These simulations reveal that differences in defect formation energies lead to significantly distinct primary damage morphologies, underscoring the MLIP’s suitability for large-scale irradiation studies of Nb. 2 Methodology 2.1 DFT Dataset Generation We generated an extensive DFT dataset to ensure that the relevant atomic configurations were adequately sampled: • Strained Configurations: Covering a range of elastic deformations around the equilibrium lattice parameter. • Defect Configurations: Including single vacancies up to the size four and self-interstitial atoms (SIAs) in different configurations. • Liquid Configurations: Sampling disordered states, which are relevant for exploring formation of complex defect morphologies in a collision cascade [ 25 ] . All DFT calculations were performed using VASP [ 20 ] , employing projector augmented-wave (PAW) pseudopotentials [ 26 ] and a Perdew–Burke–Ernzerhof (PBE) exchange-correlation functional. The computed energies, forces, and stresses formed the training targets for the potential. We preform ab-initio MD with the defect and liquid configurations and extract multiple frames from each of the simulations. Figure 1: The Cumulative diversity scores based on Vendi method [ 21 ] for DFT dataset of category SIA defects. We chose the saturation point of the score to decide the maximum number of datasets to be chosen for a category. To enhance the effectiveness of our training dataset, we employed the Vendi Score [ 21 ] , a diversity evaluation metric defined as the exponential of the Shannon entropy of the eigenvalues of similarity matrix of the features of the dataset. The features for dataset selection algorithm included energy per atom, maximum force and mean force in addition to the Bispectral features representing atomic configuration. This ensures that the diversity is measured in terms of energies, forces as well as the configuration space. Fig 1 shows the cumulative diversity score as we keep on adding the samples to the dataset. The absolute value of the diversity score shows the effective number of samples in the dataset while we chose the maximum saturation point as the measure for maximum number of samples that we would use from a category of configurations. This approach allowed us to assess the diversity within each category of configurations and determine the appropriate number of samples accordingly. These scores were then combined with a D-optimal selection method [ 22 , 27 ] to select the most representative configurations. The D-optimal selection method seeks to maximize the determinant of the information matrix, thereby minimizing the variance of parameter estimates. We train the potential on 550 total configurations out of which 90% are selected in the training set and 10% are kept for testing and validation. 2.2 Machine Learning Potential Fitting We used the SNAP framework [ 28 , 29 ] , which leverages bispectral feature descriptors derived from atomic neighbor environments via spherical harmonics. These feature descriptors capture the necessary radial and angular correlations of the atomic configuration. Because the bispectral features have linear relationship with energy, they can be efficiently trained with simple linear models. The continuous closed form of bispectral features ensures that the forces can