Machine-learnt potential highlights melting and freezing of aluminium nanoparticles

Davide Alimonti, Francesca Baletto · 2024

Read Machine-learnt potential highlights melting and freezing of aluminium nanoparticles by Davide Alimonti, Francesca Baletto (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-16294.txt.

Introduction

Davide Alimonti and Francesca Baletto wrote this materials paper in 2024 on aluminium nanoparticles. They ran computer runs of the full cycle of melting and freezing for specks from 200 to 11,000 atoms. The forces came from a new machine-learnt field, part of the FLARE suite, trained with care so melted drops behave sanely at the change point.

Shape held in bands as size grew: ball-like twenty-faced forms ruled up to about 2,000 atoms, ten-faced forms took over to 25,000, and the common cube-like crystal won past that. Melting always ran hotter than freezing at the set rate, a loop called hysteresis. Cooling a drop slowly stretched the twenty-faced band to 5,000 atoms. No skin-first melting showed, though the skin hinted at early freezing. The map guides makers of metal specks for catalysts and prints.

Catalog record

CREATORDavide Alimonti, Francesca Baletto
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE51 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-16294

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Machine-learnt potential highlights melting and freezing of aluminium nanoparticles Preprint: AIP/123-QED Davide Alimonti, Francesca Baletto Abstract We investigated the complete thermodynamic cycle of aluminium nanoparticles through classical molecular dynamics simulations, spanning a wide size range from 200 atoms to 11000 atoms. The aluminium-aluminium interactions are modelled using a newly developed Bayesian Force Field (BFF) from the FLARE suite, a cutting-edge tool in our field. We discuss the database requirements to include melted nanodroplets to avoid unphysical behaviour at the phase transition. Our study provides a comprehensive understanding of structural stability up to sizes as large as 3 ​ × ​ 10 5 3×10^{5} atoms. The developed Al-BFF predicts an icosahedral stability range of up to 2000 atoms, approximately 2 nm, followed by a region of stability for decahedra, up to 25000 atoms. Beyond this size, the expected structure favours face-centred cubic (FCC) shapes. At a fixed heating/cooling rate of 100K/ns, we consistently observe a hysteresis loop, where the melting temperatures are higher than those associated with solidification. The annealing of a liquid droplet further stabilizes icosahedral structures, extending their stability range to 5000 atoms. Using a hierarchical k-means clustering, we find no evidence of surface melting but observe some mild indication of surface freezing. In any event, the liquid droplet’s surface shows local structural order at all sizes. I Introduction The melting and freezing of metallic nanoparticles (MNPs) deviate remarkably from their bulk counterparts because of the interplay between surface effects and solid-solid structural transitions, resulting in a complex free energy surface and different solid-to-solid and solid-to-liquid pathways. Baletto and Ferrando 2005 ; Li and Truhlar 2014a ; Li and Truhlar 2014b Understanding the phase transitions of MNPs is fundamental in designing devices for targeted applications. In particular, Aluminum nanoparticles (AlNPs) have attracted much interest due to their applications in strategic industries, including catalysis Das, Rej, and Panda 2022 and energy storage Zheng et al. 2008 . To this aim, Molecular Dynamics (MD) simulations are an invaluable tool revealing the atomistic mechanisms of phase transitions at the nanoscale, including surface melting Zeni et al. 2021a . Unfortunately, ab initio MD is unfeasible due to large timescales and configurational ensembles, while classical MD based on EAM Puri and Yang 2007 , Streitz-Mintmire Puri and Yang 2007 , or second-moment approximation of the tight-binding potentials (TBSMA) Zeni et al. 2021a fails to replicate the melting temperature of the bulk surfaces. A reason for such discrepancy could reside in the fitting over bulk crystal properties, which renders traditional interatomic potentials unable to model the rupture or formation of bonds and the heavy distortion of the lattice structure at high temperatures, with the appearance of many defects. Hence, their applications to AlNPs might provide misleading results. Following the seminal works by Parrinello and Behler Behler and Parrinello 2007 and Bartók Bartók et al. 2010 , machine-learnt potentials (MLPs) trained on DFT calculations have emerged as an alternative that bridges quantum and classical methods. MLPs promise to offer accuracy at the level of ab initio methods at fractions of their computational cost while exhibiting favourable scaling with system size, unlocking simulations of large systems at the timescale of nanoseconds, and have already been successfully used for cMD of Au Jindal and Bulusu 2020 and Al Chapman and Ramprasad 2020 MNPs (however in the latter a small set of nanoparticles was considered). Among the various MLPs, Bayesian Force Fields (BFFs) possess many desirable features, most notably the ability to learn efficiently from small datasets and an intrinsic measure of predictive uncertainty, which enables active learning. FLARE —Fast Learning of Atomistic Rare Events Vandermause et al. 2022 ; Vandermause et al. 2020 ; Owen et al. 2023a — is a state-of-the-art package to perform active learning and training of BFFs and has been successfully applied to a variety of nanosystems, such as Pt and Au nanoparticles and aluminium and gold surfaces Owen et al. 2023b ; Zeni et al. 2021b . Here, we present a robust BFF for AlNPs trained using FLARE. We compare the performance in reproducing low-index surface energies and bulk melting temperature with respect to available experimental data, providing a comprehensive understanding of the behaviour of AlNPs. We then study the energy stability of AlNPs, predicting icosahedra as the most favourable geometry up to 2000 atoms ( \sim 2 nm) and of FCC-like motifs behind 25000 atoms. Finally, we simulate the melting/freezing of AlNPs, estimating the presence of a hysteresis loop and the shift of the icosahedral range coming from the annealing of a liquid droplet. II Methodology We study the complete thermodynamical cycle, melting/freezing, of Al nanoparticles with various initial geometries between 10 2 -10 4 atoms. We perform iterative Molecular Dynamics (itMD) simulations using LAMMPS. Thompson et al. 2022 itMD is a standard procedure Delgado-Callico et al. 2021 ; Rossi et al. 2018 where the system temperature is modified iteratively in finite quantities. After equilibrating the system at an initial temperature, the system temperature is raised or lowered by a fixed amount \Delta T after a specific time interval t_{itMD} , long enough to sample the target ensemble at each temperature significantly. The procedure is repeated until the desired final temperature is reached. The heating/cooling rate is simply \lambda=\Delta T/t_{itMD} . Here, we consider various starting AlNPs configurations, built as geometrical structures with small random displacements, which are initially equilibrated at 500 K, heated up to 1100 K and then cooled back down to 500 K. A Langevin thermostat regulates the system temperature with a relaxation time \tau=20 ps. We fix \Delta T=50\text{K} , and then the heating rates can be chosen by adjusting t_{itMD} . Here we report results obtained using \lambda=100 K/ns. During each t_{itMD} interval, the system is free to evolve according to Newton’s equations of motion, which are integrated using a Velocity-Verlet algorithm Verlet 1967 with a timestep of 2 fs. For sizes up to 2057 atoms, we average our results over three independent simulations. Interactions between atoms are modelled with a Bayesian Force Field (BFFs), trained on ad hoc databases, including bulk, surfaces, and small nanoparticles, as summarised in Table 1 . II.1 Training and validating Al-BFFs We train Al-BFFs using the FLARE package Vandermause et al. 2020 , which benefits from integrated active learning capabilities. Training is performed on energy, force, and stress. Active learning trajectories are initiated from an ab initio calculation on an initial configuration over which a preliminary BFF is trained. The systems evolve according to the preliminary BFF until the predictive uncertainty of at least one local environment exceeds a fixed threshold, at which point DFT calculations are performed, and the potential is updated, then used again to perform MD until the uncertainty threshold is crossed again. Active learning efficiently samples the system’s configurational space by including configurations only when needed, based on the available database. We use Quantum Espresso (QE) Giannozzi et al. 2017 as ab initio reference, selecting the PBEsol functional (QE-PBEsol). Generation of the database and model training are better detailed in Suppl. Info. Featurization of local environments is performed via the Atomic Cluster Expansion (ACE) 3-body B2 descriptors Drautz 2019 using Chebyshev polynomials of the first kind as the radial basis. The kernel function is taken as the

Read the complete text