Machine-Learned Interatomic Potentials for Structural and Defect Properties of YBa$_2$Cu$_3$O$_{7-δ}$

This paper develops and benchmarks four machine-learned interatomic potentials (ACE, MACE, GAP, and tabGAP) trained on DFT data to enable accurate, large-scale molecular dynamics simulations of radiation-induced defects and oxygen stoichiometry in YBCO high-temperature superconductors.

The Gist

The "Digital Twin" for Super-Materials: A Simple Guide

Imagine you are trying to design the perfect, indestructible superhero suit. To do this, you need to know exactly how every single thread, molecule, and atom in the fabric reacts when it gets hit by a speeding bullet, scorched by fire, or frozen in deep space.

In the real world, testing this is incredibly expensive and slow. You’d have to build the suit, shoot it, burn it, and then look at it under a microscope. In the world of science, we use Density Functional Theory (DFT)—which is like a super-powerful, ultra-precise microscope that can simulate atoms. The problem? It’s so slow and "heavy" that it’s like trying to simulate a whole superhero battle one single atom at a time. It would take a thousand years to finish a single second of action.

This paper is about building a "Cheat Code" for science.

---

The Problem: The Fragile Super-Material

The scientists are looking at a material called YBCO. This is a "High-Temperature Superconductor." Think of it as a "magic wire" that can carry electricity with zero wasted energy. This technology is the key to making compact fusion reactors—the "miniature suns" that could provide limitless clean energy.

However, there is a catch: YBCO is a bit of a diva. If you hit it with radiation (like the particles flying around inside a fusion reactor), it gets "damaged." Tiny atoms get knocked out of place, and suddenly, the magic wire stops working. To fix this, scientists need to see exactly how that damage happens at the atomic level.

The Solution: The Four "Digital Translators"

Since the "ultra-precise microscope" (DFT) is too slow, the researchers decided to train Machine-Learned Interatomic Potentials (MLPs).

Think of these MLPs as Digital Translators.

• The DFT (The Professor): He is incredibly smart and knows every tiny detail about how atoms interact, but he talks very slowly and takes forever to answer a question.
• The MLPs (The Fast-Talking Students): These are AI models. We show the Professor a few thousand examples of how atoms move, and the students "learn" the patterns. Once they learn, they can answer questions about how atoms move almost instantly, with nearly the same accuracy as the Professor.

The researchers tested four different "students" (different AI architectures) to see who was the best:

1. MACE: The "Genius Scholar." He is the most accurate and understands the most complex patterns, but he’s a bit slow and "expensive" to hire.
2. ACE: The "Reliable All-Rounder." Very smart, very fast, and great at most things.
3. GAP & tabGAP: The "Speedsters." They aren't quite as deep thinkers as the others, but they are incredibly fast. They are like the "quick-math" experts who can handle massive amounts of data in a heartbeat.

What did they find?

The researchers put these four students through a series of grueling "final exams":

• The Stress Test: How does the material squeeze and stretch? (The students passed!)
• The Heat Test: How does it change when it gets hot? (They accurately predicted a major structural change called the "orthorhombic-to-tetragonal transition.")
• The Damage Test: If we knock an atom out of place (a defect), how much energy does it take? (The students were much better at this than any previous AI models!)

Why does this matter to you?

By creating these "Digital Twins" of YBCO, scientists can now run massive, high-speed simulations of radiation damage.

Instead of waiting decades to build and test a fusion reactor, they can run a "virtual battle" in a computer. They can see exactly how the "magic wire" breaks and, more importantly, how to reinforce it so it doesn't.

In short: They’ve built a high-speed flight simulator for atoms, helping us pave the way toward a future of clean, limitless fusion energy.

Technical Summary

Technical Summary: Machine-Learned Interatomic Potentials for Structural and Defect Properties of $\text{YBa}_2\text{Cu}_3\text{O}_{7-\delta}$

1. Problem Statement

High-Temperature Superconductors (HTS), specifically Rare-Earth Barium Copper Oxides (REBCO) like $\text{YBa}_2\text{Cu}_3\text{O}_{7-\delta}$ (YBCO), are critical for the development of compact Tokamak fusion reactors. The superconducting performance of YBCO is highly sensitive to its oxygen stoichiometry ($\delta$) and the presence of structural defects.

While Density Functional Theory (DFT) provides high-fidelity insights into these materials, it is computationally restricted to very small length and time scales. This makes it impossible to directly simulate high-energy, non-equilibrium phenomena such as radiation-induced collision cascades and long-term defect evolution. Existing classical interatomic potentials often fail to capture the complex chemical environment, the orthorhombic-to-tetragonal phase transition, or the sub-stoichiometric variations essential for modeling real-world radiation damage.

2. Methodology

The authors developed and benchmarked four distinct Machine-Learned Potential (MLP) architectures to bridge the gap between DFT accuracy and large-scale Molecular Dynamics (MD) efficiency:

• MACE (Message-Passing Atomic Cluster Expansion): An equivariant Graph Neural Network (GNN) that captures high-order many-body correlations.
• ACE (Atomic Cluster Expansion): A numerically efficient framework representing local environments through a many-body series of basis functions.
• GAP (Gaussian Approximation Potential): A kernel-based method using Gaussian Process Regression (GPR) with many-body descriptors.
• tabGAP (Tabulated GAP): An accelerated version of GAP that uses spline interpolation on descriptor grids to significantly increase computational speed.

Data Generation:
The models were trained on a bespoke, extensive DFT database (generated via CP2K) comprising $\sim$2,712 structures and over 7 million force components. The dataset was specifically designed to include:

• Stoichiometric diversity: Ranging from $\text{YBCO}_7$ to $\text{YBCO}_6$.
• Non-equilibrium configurations: Including volume expansions/compressions, strained/distorted lattices, thermal displacements, and amorphous cells to represent collision cascades.
• Defect sampling: Extensive coverage of vacancies, antisites, and interstitials across all constituent elements.

3. Key Contributions

• Bespoke Database: Creation of one of the most chemically and structurally complex training sets for a single material system to date.
• Phase Transition Modeling: Successful capture of the temperature-driven orthorhombic–tetragonal phase transition, a feat previous potentials struggled to achieve.
• Defect Fidelity: Development of potentials capable of predicting complex defect formation energies and migration pathways (NEB) with DFT-level accuracy.
• Performance Benchmarking: A comprehensive comparison of accuracy versus computational cost, providing a roadmap for researchers to choose the right tool for specific simulation scales.

4. Results and Discussion

• Accuracy: MACE emerged as the most accurate model, particularly for interatomic forces and complex defect configurations (e.g., oxygen interstitials and antisite defects). ACE provided a strong balance, showing high consistency in vacancy energies.
• Thermodynamics: All MLPs successfully reproduced the lattice parameter changes associated with oxygen depletion and the orthorhombic-to-tetragonal transition. The models accurately tracked the equation of state (energy-volume curves) and thermal expansion.
• Defects: The MLPs significantly outperformed previous models (e.g., Gray et al., Chaplot) by correctly predicting the energetic ordering of oxygen vacancies and cation antisites. While GAP/tabGAP struggled slightly with oxygen chemical potentials, they remained reliable for migration barrier (NEB) calculations.
• Efficiency: There is a clear trade-off between speed and precision. tabGAP is the most efficient, being $\sim$5 times faster than ACE on GPUs and significantly faster on CPUs. MACE is the most computationally expensive and scales less efficiently with system size.

5. Significance

This work provides a robust computational foundation for the next generation of fusion energy research. By enabling large-scale, high-fidelity MD simulations of radiation damage, these MLPs allow scientists to:

1. Predict how neutron irradiation affects the superconducting properties of HTS magnets.
2. Study the evolution of defect morphologies in complex, non-stoichiometric environments.
3. Inform the engineering design of shielding and material compositions to extend the operational lifespan of fusion reactor components.