Molecular Dynamics and VASP Machine Learning Force Fields

Introduction to Speaker

• Speaker: Gear Cresser, Head of VASP software company, Professor at the University of Vienna
• Focus: Presentation on new feature in VASP related to machine learning force fields

Overview of Machine Learning Force Fields

Why Machine Learning?

• Motivation: Speed up calculations, handle more complex systems beyond traditional density functional theory (DFT) methods
• Performance: Machine learning force fields can accelerate calculations by factors of 1000-10,000

How It Works

Three-Step Process

1. Database Construction: Calculate energies, forces, stress tensors for ~1000 structures via ab initio calculations
2. Representation of Local Environment: Use descriptors to capture atom surroundings up to a cutoff distance
3. Fitting the Force Field: Fit a finite range force field using regression or neural networks

Local Environment Descriptors

• Pair Correlation Functions: Likelihood of atoms at specific distances
• Angular Correlation Functions: Distribution of angles between atoms and their neighbors
• Descriptors: Calculate using spherical harmonics and Bessel functions

Model Training and Evaluation

Necessary Assumptions

• Energy and forces as functions of local environments
• Representation involves up to 1000 coefficients per atom

Kernel Methods

• Select reference atoms and evaluate similarity using kernels
• Fit weights to kernels to create surrogate energy models

Practical Implementation in VASP

Key Parameters

• Cutoff Radius: Typically 5 Angstroms
• Broadening Parameter: Fixed at 0.5
• Number of Radial Basis Functions: 8 recommended
• Maximum Angular Quantum Number: Lmax typically 4
• Threshold for Forces: Determines when first principle calculations are necessary, adjustable via mlcti4

Application Examples

Zirconia Study

• Phase Transitions: Monoclinic to tetragonal to cubic
• Training: 592 first-principle calculations, retrained with Singular Value Decomposition
• Thermodynamic Integration: Used to fine-tune phase transition temperatures

Other Applications

• Thermal Conductivity: Calculating using Green-Kubo equations
• Elastic Constants: Machine learning force fields agree well with DFT results
• Melting Properties: Predicted accurately using interface pinning method combined with ML force fields
• Solvation Energies: Calculated correctly using ML force fields based on thermodynamic integration

Final Notes

• Importance of Testing: Ensuring reliability of machine learning models
• Machine Learning Necessity: Significant for handling complex systems efficiently
• Future Developments: Continued refinements in parameters and applications

Q&A Highlights

• Ground Truth: Grounded in density functional theory
• Ensemble Simulations: Handling specific temperature and pressure setups
• Starting ML Force Fields: No initial guess needed, starts from scratch if necessary

Conclusion

• Finite Temperature Materials Modeling: Now accessible via VASP and machine learning
• Efficiency of Machine Learning: 500-1000 DFT calculations typically sufficient for each phase studied
• Machine Learning in Material Science: It’s essential, not just a hype, providing accurate and efficient solutions
• Warnings and Caveats: Always validate the machine learning models, particularly for untrained structures

Emoji: 🧬