[Paper Review] Deep-learning interatomic potential for irradiation damage simulations in MoS2 with ab initial accuracy

Hao Wang, Xun Guo+1 more|arXiv (Cornell University)|Oct 19, 2020

Ion-surface interactions and analysis16 references4 citations

TL;DR

This paper presents a deep-learning interatomic potential (DP-Tab) for monolayer MoS₂ that achieves ab initio accuracy in simulating irradiation damage. By combining all-electron calculations, active learning, and a hybrid deep-learning model, the potential accurately predicts near-equilibrium properties and reproduces irradiation-induced nanopore formation—validating with 500 keV Au⁺ experiments, including single-ion generation of nanopores >2 nm in diameter.

ABSTRACT

Potentials that could accurately describe the irradiation damage processes are highly desired to figure out the atomic-level response of various newly-discovered materials under irradiation environments. In this work, we introduce a deep-learning interatomic potential for monolayer MoS2 by combining all-electron calculations, an active-learning sampling method and a hybrid deep-learning model. This potential could not only give an overall good performance on the predictions of near-equilibrium material properties including lattice constants, elastic coefficients, energy stress curves, phonon spectra, defect formation energy and displacement threshold, but also reproduce the ab initial irradiation damage processes with high quality. Further irradiation simulations indicate that one single highenergy ion could generate a large nanopore with a diameter of more than 2 nm, or a series of multiple nanopores, which is qualitatively verified by the subsequent 500 keV Au+ ion irradiation experiments. This work provides a promising and feasible approach to simulate irradiation effects in enormous newly-discovered materials with unprecedented accuracy.

Motivation & Objective

• To develop a highly accurate interatomic potential for simulating irradiation damage in 2D materials like monolayer MoS₂.
• To overcome the limitations of traditional analytical potentials, which often fail to capture complex irradiation processes due to inaccurate many-body interactions.
• To achieve ab initio-level accuracy in predicting both equilibrium properties and dynamic irradiation cascades.
• To enable reliable simulation of defect formation, including nanopore generation, under high-energy ion irradiation.
• To validate the model against experimental results, particularly the formation of large or multiple nanopores via single-ion irradiation.

Proposed method

• Employed all-electron density functional theory (DFT) calculations to generate high-accuracy reference data for training.
• Applied an active-learning sampling strategy to iteratively select the most informative atomic configurations, improving data efficiency.
• Developed a hybrid deep-learning model combining a repulsive table potential with a deep neural network to capture both short-range repulsion and long-range many-body effects.
• Trained the potential using a loss function that minimizes errors in energy, forces, and stress across diverse configurations, including defective and strained systems.
• Validated the potential against DFT-calculated properties such as lattice constants, elastic coefficients, phonon spectra, defect formation energy, and displacement threshold energy.
• Conducted large-scale molecular dynamics simulations using the trained potential to simulate irradiation cascades with 500 keV Au⁺ ions.

Experimental results

Research questions

• RQ1 Can a deep-learning interatomic potential achieve ab initio-level accuracy in predicting both equilibrium and non-equilibrium properties of monolayer MoS₂?
• RQ2 To what extent can the potential reproduce the complex dynamics of irradiation damage, including collision cascades and defect formation?
• RQ3 Can the simulated nanopore formation via single high-energy ion irradiation match experimental observations in terms of size and morphology?
• RQ4 How does the performance of the deep-learning potential compare to traditional analytical potentials in simulating extreme conditions like localized melting and rapid recrystallization?
• RQ5 What is the computational feasibility of using such a high-accuracy potential for large-scale, long-timescale irradiation simulations?

Key findings

• The DP-Tab potential achieves ab initio-level accuracy in predicting lattice constants, elastic coefficients, energy-stress curves, phonon spectra, defect formation energy, and displacement threshold energy.
• The potential successfully reproduces irradiation damage processes, including collision cascades and defect evolution, with high fidelity.
• Single 500 keV Au⁺ ion irradiation simulations predict the formation of a single nanopore with a diameter exceeding 2 nm, or multiple nanopores, which is qualitatively consistent with experimental observations.
• The model's predictions of defect formation energy and displacement threshold are in good agreement with DFT benchmarks, avoiding the under/overestimation common in classical potentials.
• The DP-Tab potential runs 30–50 times faster on GPU compared to CPU, with performance comparable to SNAP potentials, suggesting feasibility for large-scale simulations.
• With further optimization using TPUs or NPUs, the DP-Tab potential could potentially outperform classical analytical potentials in speed while maintaining ab initio accuracy.