FPNNK: Achieving DFT-Level Accuracy in Materials Dynamics Simulations Using Neural Networks

The First-principle Neural Network Kinetics (FPNNK) framework, open-sourced by the Cao Lab at the University of California, Irvine, combines deep neural networks (DNN) with kinetic Monte Carlo (kMC) methods to achieve DFT-level prediction accuracy for vacancy diffusion simulations while significantly improving computational efficiency, providing a solution to the long-standing trade-off between simulation accuracy and efficiency in materials science.

In materials science, researchers have long faced a dilemma: Density Functional Theory (DFT) can provide high-precision energy calculations, but its cost is extremely high, making it difficult to handle large-scale and long-time simulations; traditional empirical potential methods are fast but sacrifice accuracy, failing to reliably predict complex material behaviors. Vacancy diffusion is a core issue in materials science, directly affecting key properties such as alloy creep, phase transitions, and radiation damage recovery. However, its accurate simulation requires calculating the energy barriers for atomic jumps—this is exactly the task that DFT excels at but is expensive.

The core innovation of FPNNK lies in combining deep neural networks with kinetic Monte Carlo methods: 1. Neural Network Energy Prediction: Using a DNN trained on DFT diffusion barrier datasets, with the local atomic environment encoded in a "grid representation" as input, it predicts energy barriers, combining DFT-level accuracy, inference speed several times faster than DFT, and scalability; 2. Kinetic Monte Carlo Sampling: Based on the barriers predicted by the neural network, it samples diffusion jump directions and time scales, achieving kMC's computational efficiency while maintaining DFT accuracy.

FPNNK is built based on modern deep learning frameworks, with main dependencies including Python 3.12.4, PyTorch 2.4.0, CUDA 12.4.0 (GPU acceleration), NumPy, pandas, scipy, etc. VASP 5.x/6.x is only used for DFT/NEB calculations during the training data generation phase; after training is completed, simulations can run without VASP, lowering the threshold for use.

The project provides a complete case of vacancy diffusion simulation for Mo-Ta-W ternary alloy at 1600K. The usage workflow includes: 1. Prepare a .dump atomic model file in LAMMPS format; 2. Edit the user_inp file to specify parameters such as model path, temperature, and kMC steps; 3. Run python nnk_simu.py user_inp to start the simulation; 4. Use postprocess.py to extract trajectories and calculate properties, and plot_vacancy_trajectory.py to visualize diffusion paths.

FPNNK opens up new research paths for materials science: 1. Accelerate Materials Design: Reduce the cost of accurate dynamics simulations and quickly explore material behaviors under different conditions; 2. Complex Alloy Systems: Neural networks can learn complex interactions in multi-component alloys, providing reliable guidance for high-performance alloy design; 3. Radiation Damage Research: High-precision simulation of vacancy diffusion mechanisms helps understand the behavior of nuclear materials in radiation environments, which is of great significance for nuclear energy safety.

FPNNK is released under an open-source license, embodying the spirit of academic sharing. The project documentation emphasizes citation norms and encourages users to cite relevant manuscripts when publishing papers. This open and transparent attitude promotes progress in the field.

FPNNK is an important advancement in the interdisciplinary field of machine learning and materials science, proving that neural networks can be core components of physical simulations, achieving practical computational efficiency while maintaining first-principles accuracy. For researchers in materials dynamics, alloy design, or radiation damage, FPNNK is a powerful and easy-to-use tool. With the development of deep learning, such "machine learning-enhanced physical simulation" methods are expected to show value in more fields.