Machine-learned Interatomic Potential for Ti$_{n+1}$C$_n$ MXenes: Application to Ion Irradiation Simulations

This paper presents a robust, machine-learned interatomic potential for Ti$_{n+1}$C$_n$ MXenes that enables efficient molecular dynamics simulations of ion irradiation, providing critical insights into defect statistics and guidelines for defect engineering.

The Gist

Imagine you have a super-thin, super-strong sheet of material called a MXene. It's like a microscopic piece of graphene, but made of titanium and carbon layers stacked like a sandwich. Scientists love these sheets because they could revolutionize everything from batteries to flexible electronics.

But here's the problem: To understand how these sheets behave when they get hit by things (like radiation in space or during manufacturing), scientists need to simulate them on a computer.

Usually, there are two ways to do this:

1. The "Super-Accurate" Way: You calculate every single electron interaction. It's like trying to count every grain of sand on a beach to understand the tide. It's incredibly accurate, but it takes so long that you can only simulate a tiny speck for a split second.
2. The "Fast but Rough" Way: You use old, simplified rules (like classical physics). It's fast, but the rules are often wrong for these fancy new materials, like using a map of the 19th century to navigate a modern city.

The Solution: A "Smart" Computer Model
This paper introduces a new tool: a Machine-Learned Interatomic Potential.

Think of this as teaching a computer to be a "crystal ball" for atoms.

• The Training: The researchers fed the computer thousands of examples of how titanium and carbon atoms behave, calculated using the "Super-Accurate" method. They showed the computer every possible scenario: stretching the sheet, squishing it, melting it, and even breaking it apart.
• The Result: The computer learned the "personality" of these atoms. It now knows exactly how they push and pull on each other, but it does it 20 to 40 times faster than the old rough methods, while staying almost as accurate as the super-accurate method.

The Experiment: The "Bullet" Test
Once they had this smart model, they put it to the test. They simulated shooting tiny "bullets" (ions) at these MXene sheets. They used two types of bullets:

• Light bullets: Helium atoms (like a ping-pong ball).
• Heavy bullets: Titanium atoms (like a bowling ball).

They shot these bullets at the sheets with varying amounts of energy, from a gentle tap to a massive slam.

What Did They Find?

1. The "Self-Healing" Sheet: When the heavy bullets hit, they caused chaos. Atoms flew everywhere, and the sheet looked like a shattered window. But here's the magic: The sheet healed itself. Within a fraction of a second, the atoms rearranged themselves, and the sheet became whole again, leaving only tiny, invisible scars (defects). It's like a superhero skin that knits itself back together instantly.
2. The "Preferential Ejection": When the bullets hit, they didn't knock out the atoms evenly. The heavier Titanium atoms were much easier to knock off the sheet than the Carbon atoms (unless the bullet was very light and slow). It's like hitting a stack of bricks and marbles; the heavy bricks fly off easier than the tiny marbles in certain conditions.
3. The "Bounce vs. Pass-Through":
   • Light bullets (Helium): If they hit slowly, they bounced off. If they hit fast, they zipped right through the sheet like a ghost.
   • Heavy bullets (Titanium): If they hit slowly, they got stuck inside the sheet (implanted). If they hit fast, they punched right through.

Why Does This Matter?
This research is a game-changer for Defect Engineering.

Imagine you want to build a better battery or a faster computer chip using MXenes. You might want to intentionally poke holes in the material or stick specific atoms inside it to change how it conducts electricity.

Before this paper, scientists were guessing how to do this because they didn't have a reliable map. Now, they have a high-speed, accurate simulator. They can say, "If I shoot Helium at 50 eV, I'll get this specific pattern of holes."

The Big Picture
This paper didn't just solve a problem for one specific material; it provided a recipe. It showed other scientists exactly how to build these "smart" computer models for any new 2D material they discover. It's like giving everyone a master key to unlock the secrets of the next generation of super-materials, allowing us to design them atom-by-atom before we even build them in the lab.

Technical Summary

Here is a detailed technical summary of the paper "Machine-learned Interatomic Potential for Tin+1Cn MXenes: Application to Ion Irradiation Simulations" by Jesper Byggmästär.

1. Problem Statement

MXenes (specifically $Ti_{n+1}C_n$) are a promising class of 2D materials with applications in energy storage, electronics, and catalysis. However, their exploration via Molecular Dynamics (MD) simulations is severely limited by the lack of accurate and computationally efficient interatomic potentials.

• Current Limitations: Existing classical potentials (e.g., Tersoff) lack accuracy in reproducing complex bonding environments and defect behaviors. Ab initio MD (AIMD) is too computationally expensive for the large length and time scales required to study phenomena like ion irradiation, defect engineering, and sputtering.
• Specific Gap: There was no robust machine-learning (ML) potential available for $Ti_{n+1}C_n$ MXenes that could handle the diverse chemical environments (bulk, surface, defects, irradiation) necessary for simulating ion beam interactions.

2. Methodology

A. Machine-Learning Potential Development (tabGAP)

The author developed a tabulated Gaussian Approximation Potential (tabGAP).

• Framework: Based on the Gaussian Approximation Potential (GAP) but optimized for speed by using 1D and 3D cubic spline interpolations on pre-calculated tables.
• Descriptors: The potential utilizes a combination of:
  • Two-body interatomic distances ($r_{ij}$).
  • Three-body permutation-invariant vectors.
  • Scalar Embedded Atom Method (EAM) density.
• Training Strategy (Iterative Learning):
  1. Initial Database: Constructed from strained/sheared unit cells (ABC and AB stacking), AIMD frames at 300K/1000K, and reference structures (bulk Ti, C, TiC, graphene).
  2. Iterative Expansion: The potential was trained, then used to generate new structures (defects, disordered sheets, molten states) via short MD runs. These were relaxed using Density Functional Theory (DFT) and added back to the training set.
  3. Defect Handling: Included vacancies, self-interstitials, adatoms, and antisites.
  4. Artifact Correction: An initial version produced unphysical dense Carbon clusters during melting. Iterative training corrected this, allowing the potential to form physically reasonable fullerene-like cages.
• Helium Interactions: He interactions were treated separately using purely repulsive screened Coulomb potentials (ZBL and Beck potentials) added to the tabGAP energy, as He does not form chemical bonds in this context.

B. Validation

The potential was rigorously validated against DFT (VASP) and experimental data:

• Thermodynamics: Lattice constants and vacancy formation energies for $Ti_2C$, $Ti_3C_2$, and $Ti_4C_3$.
• Mechanical Response: Elastic constants, plastic shear response (stacking fault energies), and theoretical shear strength.
• Dynamics: Phonon dispersion relations and high-temperature melt-quench simulations (up to 5000 K).
• Static Drag: Energy landscapes of dragging Ti and C atoms from the surface to test sputtering limits.

C. Irradiation Simulations

Using the trained potential, the author performed MD simulations of He (light ion) and Ti (heavy ion) irradiation on $Ti_2C$ and $Ti_3C_2$ sheets.

• Parameters: Ion energies ranging from 15 eV to 100 keV.
• Statistics: ~1 million individual impact simulations (10,000 per energy point) to gather robust statistics on reflection, implantation, transmission, and sputtering yields.

3. Key Contributions

1. First Robust ML Potential for $Ti_{n+1}C_n$: Developed a highly accurate, transferable ML potential capable of simulating a wide range of $Ti_{n+1}C_n$ compositions (up to $n=4$) under mechanical load and irradiation.
2. Iterative Training Recipe: Established a repeatable "self-learning" workflow for constructing ML databases that includes defect generation, disorder, and artifact correction, providing a blueprint for other MXenes.
3. Computational Efficiency: The tabGAP is 20–40% slower than the classical Tersoff potential but orders of magnitude faster than DFT, enabling large-scale statistics (1M+ events) that were previously impossible.
4. Comprehensive Irradiation Database: Generated the first extensive dataset quantifying ion-MXene interactions, including sputtering yields, defect creation probabilities, and implantation depths.

4. Key Results

Potential Performance

• Accuracy: The tabGAP reproduces DFT lattice constants and vacancy formation energies with high fidelity (e.g., $VTi$ error < 0.2 eV).
• Shear Strength: Accurately captures the theoretical shear strength and stacking fault profiles, outperforming Tersoff which showed unphysical local maxima due to cutoff artifacts.
• Phonons: Correctly reproduces phonon dispersion relations, indicating accurate near-equilibrium thermoelastic properties.
• Spin Polarization: The potential correctly handles the energy limits of isolated atoms by training on spin-polarized DFT data for isolated atoms, correcting a known deficiency in non-spin-polarized DFT drag calculations.

Ion Irradiation Insights

• Ion Fate:
  • He Ions: Mostly reflected at low energies (<100 eV) and transmitted at high energies. Implantation probability is low but non-zero.
  • Ti Ions: Implantation is the dominant outcome at low energies (<300 eV for $Ti_3C_2$); transmission dominates at high energies (>1 keV).
• Sputtering Yields:
  • Preferential Sputtering: Confirmed preferential sputtering of Ti over C at higher energies, consistent with experiments.
  • Low-Energy Anomaly: At very low He energies, C sputtering is higher than Ti due to more efficient kinetic energy transfer (mass similarity).
  • Double Peak: A unique double-peak in C sputtering yield for He irradiation on $Ti_3C_2$ was observed, attributed to a threshold where a single ion can sputter both a Ti and a C atom.
• Damage and Self-Healing:
  • Heavy ion irradiation creates significant disorder and vacancies.
  • Resilience: Despite heavy damage, MXene sheets exhibit remarkable self-healing capabilities. After 100 ps of relaxation at 300 K, the lattice largely recovers into an intact sheet, leaving primarily point defects (vacancies, adatoms) rather than open pores.
  • Implantation: He implantation can displace C atoms, creating adatoms, though the probability is low (~0.0001–0.0005 for low energies).

5. Significance

• Defect Engineering: The study provides quantitative guidelines for tuning MXene properties via ion irradiation. It demonstrates that MXenes are robust enough to withstand heavy ion bombardment while allowing for controlled defect introduction (e.g., for electronic tuning or functionalization).
• Methodological Advancement: The work proves that ML potentials can bridge the gap between the accuracy of DFT and the scale of classical MD for complex 2D materials.
• Future Applications: The developed potential and training strategy serve as a foundation for simulating surface chemistry, high-entropy MXenes, and other transition metal carbides/nitrides, accelerating the discovery of new 2D material applications in radiation environments.