Palladium plays a crucial role in advancing the hydrogen economy, serving in applications such as hydrogen purification, sensing, catalysis, and conversion. While machine learning approaches have shown great promise to accelerate first-principles calculations, it is currently unclear if they are accurate enough to contribute to the design of Pd-based membranes and catalyst materials. The core challenge lies in navigating the vast compositional space in non-stoichiometric Pd-hydride phases. To address this challenge, we develop a machine-learning interatomic potential (MLIP) for palladium hydride using an on-the-fly active learning approach. This method efficiently samples diverse atomic configurations across a wide range of hydrogen concentrations. The trained MLIP is validated against experimental data, demonstrating prediction of thermo-mechanical and chemical properties with density functional theory (DFT)-level accuracy. Additionally, its transferability and performance are compared against existing semi-empirical methods. The heat of mixing for Pd-H is found to be negative with the trained MLIP, whereas it is positive valued for the EAM potential available in literature. Consequently, our MLIP yields a more realistic phase diagram with the -phase hydrogen concentration being closer to the experimental one.
Ultra-precision machining of silicon underpins semiconductor, optical and MEMS manufacturing, yet atomistic predictions remain highly sensitive to the interatomic potential employed. Here, seven popular interatomic potentials of silicon were benchmarked for thermal, cohesive, stacking-fault and tensile properties, with five carried through detailed dislocation and nanometric-cutting analyses and compared against TEM characterisation of ground silicon. The predicted responses vary markedly: apparent one-phase thermal-instability temperatures span 2309–3427 K, the unstable stacking-fault energies associated with the first {111} glide-set Shockley-partial displacement range from 1.918 to 2.948 J m−2, and Young's moduli vary from 65.06 to 99.93 GPa. Fitted dislocation-core energies, i.e., Ec(110) differ by almost a factor of six, from 0.1109 to 0.6398 eV Å−1. These differences translate into distinct stress-accommodation pathways during cutting. ABOP (Erhart and Albe) and classical Tersoff revealed localised deformation within narrow transformed/disordered subsurface regions, Stillinger–Weber activates transient dislocation-mediated plasticity, Screened ABOP promotes extensive disorder-mediated accommodation, and Kumagai exhibits an intermediate response. TEM observations of ground silicon revealed coexisting amorphisation, {111} stacking faults and partial-dislocation activity, supporting a mixed-mechanism picture of machining-induced damage. The results show that predictive atomistic simulations of silicon machining require mechanism-specific potential selection rather than validation against a single bulk property.
Phase-field surrogates reduce microstructure simulation cost, yet transfer to unseen material conditions requires target-specific capacity. Standard parameter-efficient fine-tuning (PEFT) methods apply static parameter capacities, failing to adapt to time-varying spectral distributions and physical constraints. In this work, we propose Physical Time-Adaptive Nested Low-Rank Transfer (ST-NLR), a novel framework that utilizes physical time as a dynamic coordinate to modulate model capacity during fine-tuning. ST-NLR introduces a unified, nested factor bank and adaptively activates prefix ranks (e.g., ranks 4, 8, and 16) governed by thermodynamic and physical constraints, including free-energy dissipation, phase-boundary dynamics, and mass conservation. Evaluated across three canonical phase-field benchmarks — Allen–Cahn, Cahn–Hilliard, and anisotropic solidification (PFHub-3) — ST-NLR consistently outperforms state-of-the-art PEFT baselines. Notably, ST-NLR reduces free-energy prediction errors by up to 16.92% while decreasing adapter arithmetic operations (GMACs) by 59.09%. A 30-round V100 benchmark reflects implementation overhead in eager PyTorch, clearly distinguishing theoretical arithmetic reduction from current deployment latency. ST-NLR provides an efficient, physically grounded transfer learning strategy for scalable phase-field foundation models.
This study aims to quantify the contributions of atomic motion in jump, gradual-motion, and stay sections to the self-diffusion coefficient in liquid metals. A computational framework is proposed to reconstruct the self-diffusion coefficient from the average squared displacement and frequency of each section type. Molecular dynamics simulations of liquid Pb at 773K were performed using a polynomial machine-learning potential. The trajectory of each atom was classified into the three section types. A correction derived from localized oscillatory motion was introduced to relate the calculated section displacements to the effective translational displacements. The self-diffusion coefficient reconstructed from the classified atomic motions differed by less than 1% from that independently obtained from the long-time mean-squared displacement. At 773K, gradual motion contributed approximately 65% of the reconstructed self-diffusion coefficient, while jump and stay motions contributed approximately 25% and 10%, respectively. These results demonstrate that the framework can quantitatively relate classified microscopic atomic motions to macroscopic self-diffusion. An exact decomposition of the calculated displacement vectors further revealed a net negative contribution from correlations among these vectors. Analysis of temporally adjacent displacement vectors belonging to the same atom showed that shared section boundaries generate negative correlations through local oscillatory fluctuations. An important part of this correlation effect is therefore approximately incorporated into the oscillation correction rather than being simply neglected in the motion-resolved formulation. At 773K, the scalar oscillation correction was close to the magnitude of the negative calculated-vector cross contribution. The framework thus provides a quantitative connection between macroscopic self-diffusion and classified microscopic atomic motions while allowing the assumptions and applicability of the motion-resolved representation to be examined from the calculated trajectories.
Diamond-like phases of boron nitride (BNDLPs) are fully sp3-bonded crystal structures analogous to diamond and cubic BN, promising candidates for super-hard coatings and electronic applications. However, the elastic properties of most BNDLPs remain unstudied, which hinders their practical application. The aim of this work is to investigate the elastic properties of 14 diamond-like BN phases based on BN nanotubes and hexagonal BN layers by calculating the complete set of elastic constants (stiffness constants, Young’s modulus, shear modulus, Poisson’s ratio, and bulk modulus) using molecular dynamics combined with analytical calculations. The considered structures are mechanically stable except for TA3 and exhibit high stiffness. The stiffness constants of BNDLPs are comparable to or exceed those of cubic and wurtzite BN. Phase TA6 achieves a maximum Young’s modulus of 1695 GPa, exceeding that for diamond. Some phases (LA6, LA7, TA4, and TA6) exhibit high anisotropy (anisotropy ratio ≥ 2), partial auxeticity (minimum Poisson’s ratio of −0.12 for LA1), or a Poisson’s ratio close to zero, opening up new possibilities for property control through atomic rearrangement. The obtained results show that BNDLPs combine high stiffness with a broad range of values of Poisson’s ratio, highlighting their potential as next-generation materials with super-hardness and strength.
