[Paper Review] A general-purpose neural network potential for Ti-Al-Nb alloys towards large-scale molecular dynamics with ab initio accuracy

Zhiqiang Zhao, Wanlin Guo+1 more|arXiv (Cornell University)|Mar 14, 2024

Machine Learning in Materials ScienceMaterials Science3 citations

TL;DR

This paper presents a general-purpose machine-learned potential (MLP) for Ti-Al-Nb alloys using a neural evolution framework with active learning, trained on first-principles data. The resulting MLP achieves ab initio accuracy in large-scale molecular dynamics simulations of up to tens of millions of atoms, accurately capturing Nb doping effects on stacking fault and formation energies, and enabling high-fidelity study of micro-mechanical behaviors in TiAl lamellar structures.

ABSTRACT

High Nb-containing TiAl alloys exhibit exceptional high-temperature strength and room-temperature ductility, making them widely used in hot-section components of automotive and aerospace engines. However, the lack of accurate interatomic interaction potentials for large-scale modeling severely hampers a comprehensive understanding of the failure mechanism of Ti-Al-Nb alloys and the development of strategies to enhance the mechanical properties. Here, we develop a general-purpose machine-learned potential (MLP) for the Ti-Al-Nb ternary system by combining the neural evolution potentials framework with an active learning scheme. The developed MLP, trained on extensive first-principles datasets, demonstrates remarkable accuracy in predicting various lattice and defect properties, as well as high-temperature characteristics such as thermal expansion and melting point for TiAl systems. Notably, this potential can effectively describe the key effect of Nb doping on stacking fault energies and formation energies. Of practical importance is that our MLP enables large-scale molecular dynamics simulations involving tens of millions of atoms with ab initio accuracy, achieving an outstanding balance between computational speed and accuracy. These results pave the way for studying micro-mechanical behaviors in TiAl lamellar structures and developing high-performance TiAl alloys towards applications at elevated temperatures.

Motivation & Objective

• To address the lack of accurate interatomic potentials for large-scale modeling of high Nb-containing TiAl alloys.
• To develop a machine-learned potential that maintains ab initio accuracy while enabling simulations of tens of millions of atoms.
• To capture the critical role of Nb doping in modifying stacking fault energies and formation energies in Ti-Al-Nb systems.
• To support the study of micro-mechanical behaviors in TiAl lamellar structures under thermomechanical loading.
• To provide a general-purpose, transferable potential for future alloy development in aerospace and automotive applications.

Proposed method

• Employing a neural evolution potentials framework to optimize the architecture and hyperparameters of the neural network potential.
• Integrating an active learning scheme to iteratively select the most informative atomic configurations for first-principles training data.
• Training the MLP on a comprehensive dataset of first-principles calculations covering diverse crystal structures, defects, and high-temperature phases.
• Using symmetry functions and environment descriptors to encode atomic neighborhood information for invariance under rotation and translation.
• Validating the MLP against first-principles data for lattice parameters, defect formation energies, thermal expansion, and melting points.
• Scaling the potential to large-scale molecular dynamics simulations using many-body interactions and efficient parallelization.

Experimental results

Research questions

• RQ1 How accurately can a machine-learned potential describe the thermodynamic and mechanical properties of Ti-Al-Nb alloys across diverse phases and defects?
• RQ2 To what extent does the MLP capture the influence of Nb doping on stacking fault energy and formation energy in TiAl-based systems?
• RQ3 Can the MLP maintain ab initio accuracy while enabling large-scale molecular dynamics simulations involving tens of millions of atoms?
• RQ4 How transferable is the potential across different crystal structures, dislocations, and high-temperature conditions in Ti-Al-Nb alloys?
• RQ5 What is the performance-accuracy trade-off of the MLP in simulating micro-mechanical behaviors in TiAl lamellar microstructures?

Key findings

• The MLP achieves ab initio-level accuracy in predicting lattice parameters, defect formation energies, and elastic properties across multiple phases of Ti-Al-Nb alloys.
• The potential accurately reproduces the reduction in stacking fault energy induced by Nb doping, a key factor in deformation mechanisms.
• Formation energies of point defects and extended defects are predicted with high fidelity, consistent with first-principles benchmarks.
• Thermal expansion coefficients and melting points are predicted with excellent agreement to reference first-principles data.
• The MLP enables large-scale molecular dynamics simulations of up to tens of millions of atoms with computational efficiency comparable to classical potentials and accuracy approaching that of DFT.
• The model demonstrates strong transferability across diverse atomic environments, including complex dislocation configurations and high-temperature phases.