Machine Learning Potential-Driven Molecular Dynamics Simulations of Dehydrogenation in Pristine and Doped MgH$_2$

Technical Summary: Machine Learning Potential-Driven Molecular Dynamics Simulations of Dehydrogenation in Pristine and Doped MgH2

Problem Statement
Magnesium hydride (MgH2) is a leading candidate for solid-state hydrogen storage due to its high gravimetric capacity (7.6 wt%), abundance, and low cost. However, its practical application is hindered by high dehydrogenation enthalpy (~76 kJ/mol H2) and sluggish kinetics, primarily caused by strong ionic Mg-H bonds and high activation energy barriers for hydrogen diffusion. While nanostructuring and catalyst doping are known to improve performance, the atomic-scale mechanisms governing surface-mediated dehydrogenation and the specific role of dopants remain difficult to elucidate through experimental characterization alone. Furthermore, traditional computational methods face a bottleneck: classical molecular dynamics (MD) lacks the accuracy for complex bond-breaking scenarios in multi-component systems, while ab initio MD (AIMD) is computationally prohibitive for the large systems and timescales required to observe rare desorption events.

Methodology
To overcome these limitations, the authors employed a universal machine learning potential (MLP) driven molecular dynamics (ML-MD) approach. Specifically, they utilized the 'PreFerred Potential' (PFP), a neural network potential trained on 59 million DFT structures, which offers quantum-mechanical accuracy with the efficiency of classical force fields.

• Simulation Setup: Large-scale MD simulations were conducted on MgH2 systems containing approximately 2,700 atoms over 50 ps. Four low-index surface orientations—(110), (100), (101), and (001)—were investigated.
• Doping Strategy: A comprehensive screening of 22 distinct doping elements was performed by substituting Mg atoms in the surface layers of the MgH2(100) model.
• Conditions: Simulations were run at 1000 K to accelerate kinetics for statistical sampling, with a supplementary run at 800 K to capture mechanistic details with higher fidelity. H2 molecules were removed upon diffusing 30 Å from the surface to maintain low partial pressure.
• Analysis Framework:
  • Kinetic Analysis: Cumulative H2 release was quantified across different surfaces and dopants.
  • Machine Learning Descriptor Analysis: An XGBoost model coupled with SHAP (SHapley Additive exPlanations) analysis was used to correlate 22 time-coupled elemental descriptors with dehydrogenation kinetics.
  • Validation: First-principles DFT calculations were performed to validate desorption energies of subsurface H2 formation. Experimental validation was conducted via Simultaneous Thermal Analysis (STA) and DSC on pure and Ni-doped MgH2 powders to measure activation energies.

Key Contributions and Results

1. Surface Orientation Dependence:
   The study identified the MgH2(100) surface as the most active low-index facet for hydrogen release, outperforming (001), (101), and (110). This activity is attributed to the (100) surface's open structure with larger interstitial voids, which minimizes steric hindrance, despite the (001) surface having a higher surface energy.

2. Novel Dehydrogenation Pathway:
   Contrary to the conventional view that H atoms diffuse to the outermost surface to recombine, the ML-MD simulations revealed a subsurface recombination mechanism. H2 molecules preferentially form in the subsurface region (specifically between the second and third H atomic layers) before diffusing to the surface for desorption. This pathway is thermodynamically favored because the subsurface offers a balance of sufficient lattice relaxation for H2 formation and high local H density, avoiding the severe strain energy penalties of bulk formation and the trapping effects of under-coordinated surface atoms.

3. Dopant Screening and the "Hydrogen Pump" Mechanism:
   Among 22 screened elements, Nickel (Ni) emerged as the most effective dopant, exhibiting the highest total hydrogen release and fastest initial desorption rate.

   • Mechanism: Ni dopants act as thermodynamic sinks, attracting H atoms from the bulk matrix to form a local reservoir. This facilitates H-H coupling and subsequent H2 release.
   • Structural Evolution: Radial distribution function (RDF) analysis indicates that Ni dynamically forms local atomic clusters resembling the catalytic phase Mg2NiH4. These clusters serve as active sites for a "hydrogen pump" effect, continuously absorbing and releasing hydrogen.

4. Quantitative Property-Activity Relationship:
   Machine learning analysis identified the time-coupled Miedema electron density ($n_{ws}^\tau$) as the critical descriptor governing catalytic activity. A volcano-shaped relationship was established between $n_{ws}$ and total hydrogen release.

   • Optimal Window: An optimal range of 4.0 < $n_{ws}$ < 5.4 × 10⁻² e/bohr³ was identified.
   • Sabatier Principle: Dopants within this range (e.g., Ni, Cr, Pd) provide the ideal balance: strong enough to attract and concentrate H atoms, yet weak enough to allow efficient H2 desorption. Elements with too low $n_{ws}$ (e.g., La, Zr) fail to attract H sufficiently, while those with too high $n_{ws}$ (e.g., W) bind H too strongly, trapping the formed H2.

5. Experimental Corroboration:
   DSC measurements on Ni-doped MgH2 confirmed the simulation predictions. The Ni-doped sample exhibited a reduced onset desorption temperature (349°C vs. 413°C for pure MgH2) and lower activation energies (101–135 kJ/mol vs. 171 kJ/mol for pure MgH2), validating the catalytic enhancement predicted by the ML-MD framework.

Significance
The paper claims that this study demonstrates the robust capability of universal machine learning potentials to navigate complex catalytic mechanisms and establish quantitative property-activity relationships that were previously inaccessible. By revealing the subsurface nucleation mechanism and identifying Miedema electron density as a universal descriptor for catalyst design, the work provides a theoretical framework for the rational design of high-performance catalysts. The findings not only explain the experimentally observed "hydrogen pump" effect in Ni-doped systems but also suggest that the identified electronic criteria can guide the design of other hydrogen storage alloys, such as LaNi5 and FeTi, by optimizing their aggregate electronic properties.