janK · Helmholtz-Zentrum Dresden-Rossendorf

Current Research

Machine-learning molecular dynamics applied to energy materials and structural ceramics — connecting atomistic simulations to electrochemical and mechanical observables.

In progress
GAP / ML-MD Electrochemistry Surface Science Fuel Cells

Platinum Corrosion in Fuel Cells

Atomistic simulation of the full Pt degradation sequence in proton-exchange membrane fuel cell (PEMFC) cathodes using the PtOH GAP potential and TurboGAP molecular dynamics.

Motivation

Platinum dissolution is the primary degradation mechanism limiting PEMFC cathode lifetime. Under start-stop cycling at potentials U > 0.8 V vs. RHE, the Pt surface undergoes surface oxidation, Pt↔O place exchange, ionic dissolution, and Ostwald ripening — all depleting the electrochemically active surface area (ECSA) over time. Classical force fields cannot describe this chemistry; the PtOH GAP potential provides near-DFT accuracy for Pt–Pt, Pt–O, Pt–H, O–O, O–H, and H–H interactions.

Approach

A five-step GAP-MD workflow progresses from clean surfaces to realistic nanoparticles, connecting to experimental cyclic voltammetry and dissolution measurements:

1
Surface Oxidation Onset — O* coverage scan on Pt(111) at 353 K; NVT trajectories at 9 coverages (0–1 ML). Complete: spontaneous Pt ejection events confirmed at θ ≥ 0.5 ML; onset coverage θ* identified for subsurface O incorporation.
2
Dissolution Free Energy — umbrella sampling along Pt dissolution coordinate ξ (Δξ = 0.3 Å, kspring = 5 eV/Ų) at terrace, step, and kink sites; 200 ps NVT per window; WHAM reconstruction of PMF G(ξ).
3
Pt↔O Place Exchange — metadynamics characterisation of the thermally activated exchange mechanism converting chemisorbed O into a surface oxide.
4
Nanoparticle Stability — long NVT (2–10 ns) for three cuboctahedral Pt particles: 1.0 nm (147 atoms), 1.7 nm (309 atoms), 2.5 nm (923 atoms); size-dependent dissolution rate, facet stability, and Ostwald ripening.
5
Electrochemical Observables (CHE) — ΔG(OH*), ΔG(O*), ΔG(OOH*) via the computational hydrogen electrode; potential-dependent free energy diagrams.

Key result so far

All 9 Step-1 NVT trajectories completed to 500 ps at ⟨T⟩ = 352.4–352.6 K (σT ≈ 27–29 K, Bussi thermostat). The differential adsorption energy Δεdiff plateaus at −4 to −6 eV/O above 0.25 ML, signalling a nearly constant adsorption environment dominated by O–O nearest-neighbour repulsion. Spontaneous Pt displacement events (Δz > 0.8 Å for ≥ 10 ps) begin at exactly 0.5 ML — one atom (Pt #82) ejected at t = 12 ps, vacated site remaining metallic. At 0.875 ML two atoms eject simultaneously within 2.5 ps, consistent with O-induced correlated stress near saturation. At 1.0 ML the surface is highly dynamic (25 excursion events, 10–310 ps), resembling a disordered oxide precursor.

PotentialPtOH GAP (Pt, O, H; Oct 2024)
MD engineTurboGAP (NVT, Bussi thermostat)
Analysisgap_suite
Conditions353 K, acidic electrolyte, U = 0.6–1.0 V vs. RHE
SystemPt(111) 4×4 slab; cuboctahedral NPs 1.0–2.5 nm
Step 1 status✓ Complete
Key finding

Pt ejection onset at θ* = 0.5 ML O* coverage — the critical threshold for irreversible surface vacancy formation and the starting point for the corrosion cascade.

Draft manuscript
GAP / ML-MD Fracture Mechanics Ceramics Aerospace

Phase Stability, Mixing Thermodynamics, and Fracture of Ta1−xHfxC

Machine-learning molecular dynamics study of the TaC–HfC solid solution — an ultra-high-temperature ceramic (UHTC) candidate for hypersonic vehicle leading edges and space thruster nozzles.

Motivation

Ultra-high-temperature ceramics (UHTCs) must withstand temperatures exceeding 3000 K, rapid thermal cycling (500–1500 K in seconds), and mechanical stress at re-entry or thruster conditions. Ta1−xHfxC occupies a special position: both TaC and HfC adopt the rock-salt (B1) structure, the solid solution is fully miscible across the entire composition range, and Ta4HfC5 holds the highest experimentally reported melting point (~4215 K) of any known material. Despite this, the atomistic-scale thermodynamics and fracture mechanics of the alloy remained largely unexplored.

What we did

Using a Gaussian Approximation Potential trained on DFT reference data for the full Ta–Hf–C ternary (trained November 2023), three interconnected questions were addressed across supercells up to ~5 × 105 atoms:

I
Mixing thermodynamics — composition scan of ΔHmix, lattice parameter a(x), and bulk modulus B0(x) along the TaC–HfC tie-line. All mixing enthalpies are negative, confirming thermodynamic stability of the solid solution. B0 peaks at x = 0.5 (246 GPa). Deepest convex hull vertex at x = 0.778 (ΔHmin = −92.6 meV/f.u.).
II
Finite-temperature elastic properties — elastic constants and surface energies at six temperatures (0–2000 K) for four compositions, including Griffith fracture toughness KIc as a function of composition and temperature.
III
Crack propagation under aerospace conditions — mode-I crack propagation simulations in large supercells mimicking the thermal shock environment of a space thruster nozzle; crack velocity and fracture surface energy extracted as a function of composition and temperature.
PotentialGAP Ta–Hf–C ternary (Nov 2023)
FrameworkTurboGAP + ASE + QUIP/quippy
Max system~5 × 10⁵ atoms (crack supercell)
T range0–3000 K
ApplicationHypersonic TPS, thruster nozzles
StatusDraft Apr 2026
Key finding

The solid solution is thermodynamically stable across the full composition range. Bulk modulus peaks at x = 0.5 (245.6 GPa), while the deepest convex hull vertex is at x = 0.778 (ΔHmin = −92.6 meV/f.u.) — Hf-rich, not the Ta4HfC5 stoichiometry (x = 0.2) assumed by experiment.

In review
GAP / ML-MD CO₂ Reduction Nanoparticles Surface Science

General-Purpose CuAu Alloy GAP Potential

A machine-learning interatomic potential for the full copper–gold binary alloy system, trained on a broad DFT database spanning bulk phases, liquid structures, surfaces, and nanoparticles — with direct application to CO₂ reduction catalysis on CuAu nanoparticles.

Motivation

Copper–gold nanoparticles are promising catalysts for atmospheric CO₂ reduction. Copper activates CO₂ while gold promotes hydrogen dissolution, and the geometric and electronic synergy of the alloy enables selective formation of multi-carbon products. Accurate large-scale exploration of nanoparticle thermodynamics and adsorption landscapes requires an interatomic potential with near-DFT accuracy across the full CuAu composition range — something neither classical empirical potentials (Gupta, KIM) nor single-element GAPs could provide.

What we did

Using the Gaussian Approximation Potential (GAP) framework with two-body and many-body SOAP descriptors, a potential was trained on a DFT (VASP/PBE) database covering bulk phases, liquids, surfaces of diverse crystallographic orientations, and nanoparticles across the full composition range:

1
Potential training & validation — Training data generated via cluster expansion (CASM) combined with TurboGAP MD rattling. Validated against held-out DFT data and against Gupta and KIM empirical potentials. Formation energy and surface energy reproduction substantially better than both empirical benchmarks.
2
Grand-canonical nanoparticle cluster atlas — Scanned thermodynamically most stable CuAu nanoparticles from 10 to 200 atoms using ~3 million MD/CG computations via the FIREWORKS workflow. Identifies closed icosahedral magic-number shells at 55 and 147 atoms, missed by both Gupta and KIM. The stable Cu:Au ratio varies systematically with cluster size.
3
Surface energies — GAP surface energies computed for 205 crystallographically distinct slab orientations spanning the full composition range. Mean absolute error vs. DFT: MAE = 0.060 J m⁻² (R = 0.9987), confirming DFT-level accuracy for surface physics across all facets.
4
CO and H adsorption scanning — Thermodynamically stable nanoparticles from the atlas were used to scan all possible CO and H adsorption positions on the low-energy nanoparticle surfaces, providing starting points for NEB transition-state searches toward CO₂ reduction pathways to methane and methanol.
PotentialCuAu GAP (2b + many-body SOAP)
DFT referenceVASP PBE, 520 eV cutoff
MD engineTurboGAP, LAMMPS
WorkflowFIREWORKS automation
NP size range10 – 200 atoms
Surfaces205 slab orientations
AuthorsKloppenburg, Jónsson, Caró
StatusIn review
Key finding

The GAP reproduces surface energies for 205 orientations with MAE = 0.060 J m⁻² and correctly captures icosahedral magic-number stability shells at 55 and 147 atoms — both missed by empirical Gupta and KIM potentials — enabling reliable large-scale CO₂ reduction catalyst screening.