Machine learning potential for interacting dislocations in the presence of free surfaces

Area of Science:

• Materials Science
• Computational Physics
• Solid State Physics

Background:

• Calculating the total energy of interacting dislocations near free surfaces is computationally intensive, typically requiring high-resolution Finite Element (FE) methods.
• FE calculations demand very fine meshes around dislocation cores, significantly increasing computational cost and complexity.

Purpose of the Study:

• To introduce a Machine Learning (ML) approach as an efficient alternative to FE calculations for determining the energy of interacting dislocation systems.
• To develop a novel ML methodology for accurately predicting dislocation energies and forces, facilitating large-scale simulations.

Main Methods:

• Formulated the elastic problem using one- and two-body interaction terms.
• Employed Sobolev training with a feed-forward neural network (NN) architecture to learn energy and force interactions.
• Applied the ML model to corrugated, heteroepitaxial semiconductor films, using Monte Carlo methods to find minimum-energy dislocation distributions.

Main Results:

• Demonstrated that ML can effectively replace computationally expensive FE calculations for dislocation energy computations.
• Successfully applied the ML approach to analyze dislocation distributions in semiconductor films.
• Showcased the ability to perform millions of energy evaluations, infeasible with traditional FE methods.
• Validated the use of ML-derived forces for 2D dislocation dynamics simulations.

Conclusions:

• The proposed ML methodology offers a computationally efficient and accurate alternative for studying dislocation systems.
• The interaction cutoff in the ML model allows for scalable simulations across different system sizes without retraining.
• This ML approach significantly accelerates materials simulations, enabling the study of complex dislocation behaviors and dynamics.