Hydride formation and phase separation in palladium nanoparticles from a transferable atomic cluster expansion potential

This paper introduces a transferable atomic cluster expansion potential for the palladium-hydrogen system that achieves near-DFT accuracy and enables efficient, large-scale molecular dynamics simulations of PdH$_x$ nanoparticles, successfully resolving their nanoscale phase separation, size-dependent lattice expansion, and hydrogen-induced melting point depression.

The Gist

Imagine you have a tiny, invisible sponge made of metal (Palladium) that loves to drink hydrogen gas. When it drinks, it swells, changes shape, and sometimes splits into two different "personalities" inside itself. Scientists have known about this for a long time, but trying to simulate exactly how this happens on a computer is like trying to predict the weather by watching a single raindrop. It's too small, too fast, and too complicated for our usual computer tools.

This paper introduces a new, super-smart computer "rulebook" (called an Atomic Cluster Expansion or ACE potential) that acts like a crystal ball for these tiny metal sponges. Here is how the authors explain their work using simple concepts:

1. The Problem: The "Goldilocks" Difficulty

To understand how Palladium and Hydrogen interact, scientists usually use two types of computer models:

• The "Microscope" (DFT): This is incredibly accurate, like looking at every single atom with a high-powered microscope. But it's so slow that you can only watch a tiny speck of metal for a split second. It's like trying to film a whole movie by taking one photo every hour.
• The "Sketch Artist" (Old Potentials): These are fast and can watch big chunks of metal for a long time. But they are often wrong about the details. They might think the metal sponge is too stiff or that the hydrogen drinks too easily.

The authors needed a tool that was both fast enough to watch a whole nanoparticle for a long time and accurate enough to get the physics right.

2. The Solution: A New "Rulebook" (ACE)

The team created a new set of rules (the ACE potential) trained on thousands of high-accuracy "microscope" snapshots. Think of it as teaching a robot to play chess by showing it millions of grandmaster games. Once trained, the robot can play just as well as the grandmasters but much faster.

• What it does: It predicts how atoms move, how much energy it takes to move them, and how the metal surface reacts to hydrogen.
• The Result: It is nearly as accurate as the slow "microscope" method but runs thousands of times faster. This allows the scientists to simulate a nanoparticle with 28,000 atoms (about 12 nanometers wide) for several billionths of a second.

3. The Discovery: The "Core-Shell" Sandwich

Using this new rulebook, the scientists watched what happened when they filled these tiny metal sponges with hydrogen. They saw something very specific happen, which they call phase separation:

• The Setup: Imagine a ball of metal. You start pumping hydrogen into it.
• The Split: Instead of the hydrogen spreading out evenly like sugar in tea, the system gets messy. The hydrogen rushes to the outside (the shell) and packs itself tightly there, turning that outer layer into a "hard" hydride. Meanwhile, the inside (the core) stays mostly empty and soft.
• The Analogy: It's like a chocolate truffle where the outside is a hard, crunchy shell, and the inside is a soft, liquid center. The hydrogen prefers to live on the "skin" of the nanoparticle, leaving the "heart" alone.

4. The Melting Point Surprise

The scientists also heated these hydrogen-filled nanoparticles to see when they would melt (turn from solid to liquid).

• The Finding: The more hydrogen the nanoparticle drank, the lower its melting temperature became.
• The Metaphor: It's like adding salt to ice; the hydrogen acts like a "melting agent" that makes the metal structure unstable and easier to melt at lower temperatures.

5. Why This Matters (According to the Paper)

The authors state that this new tool bridges the gap between the "microscope" (too slow) and the "sketch artist" (too inaccurate).

• It allows them to see kinetic separation (how the phases split over time) in real-time.
• It reproduces experimental results that were previously hard to explain, like why the size of the nanoparticle changes the distance between atoms.
• It works even under extreme conditions, like heating the metal to 2000 Kelvin (hotter than lava) and cooling it back down, proving the rules are robust.

In summary: The paper presents a new, super-efficient computer model that finally lets scientists watch how tiny metal particles drink hydrogen, split into layers, and melt, all with a level of detail that matches real-world experiments. This helps us understand the fundamental physics of hydrogen storage and catalysis without needing to guess or rely on inaccurate shortcuts.

Technical Summary

Technical Summary: Hydride Formation and Phase Separation in Palladium Nanoparticles from a Transferable Atomic Cluster Expansion Potential

Problem Statement
The palladium–hydrogen (Pd–H) system serves as a prototype for hydrogen–metal interactions, underpinning technologies in hydrogen storage, catalysis, and purification. While the macroscopic behavior of bulk Pd–H is well-characterized, its nanoscale behavior—governed by surface energetics, elastic coherency strain, and size-dependent thermodynamics—has remained difficult to simulate accurately. Existing empirical potentials (e.g., EAM, ReaxFF) fail to quantitatively capture the energetics of interstitial hydrogen, often overestimating solution energies or failing to describe defect phases. Conversely, while machine-learning interatomic potentials (MLIPs) offer high accuracy, previous models for Pd–H (such as specific artificial neural networks) have been restricted to bulk configurations with low hydrogen content, rendering them unsuitable for studying surfaces, interfaces, or high-concentration hydride phases. Furthermore, standard Density Functional Theory (DFT) calculations are limited by computational cost, preventing the simulation of large nanoparticles (tens of thousands of atoms) over the nanosecond timescales required to observe phase separation and melting dynamics.

Methodology
To address these limitations, the authors developed a general-purpose Atomic Cluster Expansion (ACE) potential for the Pd–H system. The methodology involved:

1. Training Data Generation: A comprehensive dataset of 13,127 structures was generated using DFT (VASP with the GGA-PBE functional). This dataset included bulk Pd phases (fcc, hcp, bcc, etc.), Pd–H binary phases ($\alpha$ and $\beta$), defect structures (vacancies, grain boundaries), surface configurations, and active-learning selected structures from molecular dynamics (MD) runs at temperatures up to 2000 K.
2. Potential Parameterization: Two ACE potentials were developed: one trained on PBE+D3 data (ACED3) and a primary model trained on PBE data without dispersion corrections (ACE). The ACE model utilizes a large number of basis functions (938 for Pd–Pd interactions) to ensure high accuracy and transferability.
3. Validation: The potentials were rigorously benchmarked against reference DFT calculations and experimental data (neutron scattering, high-pressure EOS, lattice expansion) for key properties including cohesive energies, elastic constants, phonon spectra, hydrogen migration barriers, and surface adsorption energies.
4. Large-Scale Simulation: The validated ACE potential was employed in MD simulations using LAMMPS to study PdH$_x$ nanoparticles. Simulations included systems with up to 28,383 atoms (approx. 12 nm diameter) across various geometries (octahedral, truncated octahedral, capped cubic) and hydrogen concentrations ($x = 0.03, 0.3, 0.6$). These simulations were performed over 3 ns timescales to reach equilibrium and included rapid melt-quench cycles to test stability under extreme non-equilibrium conditions.

Key Contributions and Results

• Near-DFT Accuracy and Efficiency: The ACE potential reproduces formation energies, phonon spectra, elastic constants, and hydrogen migration barriers with near-DFT accuracy. It achieves this at a computational cost orders of magnitude lower than DFT, with near-linear scaling on multi-core CPUs. The authors report that the ACE potential is at least three times faster than previous neural network potentials (NNP) on CPU architectures.
• Resolution of Phase Separation: Simulations of PdH$_x$ nanoparticles revealed the kinetic separation of $\alpha$ and $\beta$ phases into a core–shell architecture. In the mixed phase ($x=0.3$), hydrogen segregates to the surface to form a saturated $\beta$-PdH shell, while the core remains a dilute $\alpha$-PdH region. This resolves the atomic-scale mechanism of phase separation previously elusive to simulation.
• Size-Dependent Thermodynamics: The simulations successfully reproduced the experimentally observed size dependence of the lattice parameter. The model captured the differential expansion between the unconstrained near-surface regions and the constrained core, explaining the variance in lattice constants across different particle sizes.
• Hydrogen-Induced Melting Point Depression: The study uncovered a pronounced lowering of the nanoparticle melting temperature as hydrogen concentration increases. MD simulations of melt-quench cycles demonstrated that hydrogen content significantly affects the thermal stability of PdH$_x$ nanoparticles.
• Surface and Defect Interactions: The potential accurately describes hydrogen adsorption on low-index Pd surfaces ((111), (100), (110)), migration barriers between surface and subsurface sites, and the segregation of hydrogen into vacancies (up to six H atoms in octahedral sites surrounding a vacancy).

Significance
The paper claims that this work brings experimentally relevant scales of metal-hydride dynamics within quantitative reach. By providing a transferable, computationally efficient potential that maintains DFT-level accuracy, the authors bridge the gap between electronic structure methods and large-scale atomistic simulations. This enables the study of complex phenomena such as phase separation, surface roughening, and melting in Pd–H systems at realistic length and time scales. The potential is presented as a robust tool for future investigations into hydrogen-metal interactions, offering an optimal balance of accuracy, efficiency, and transferability without the need for GPU acceleration, making it accessible for extensive simulations of nanoparticles exceeding 28,000 atoms.