Detall

Conferenciant: Profesor Alberto Roldán.

Resum: Computational techniques have proven reliable tools for simulating physicochemical properties and thermokinetic energy profiles. Commonly, the computational work done on supported NPs considers their morphologies in their geometric ground state, i.e., the global energy minimum, which in many cases needs to be found using stochastic global optimisation (GO) and intelligent sampling techniques. The energies of these nanoscale structures are often evaluated using density functional theory (DFT) and interatomic potentials (IP). This is problematic because (i) small moieties (size ≤1.5 nm) have a relatively large number of low-lying energy configurations, making the number of DFT calculations to map their stabilities overly demanding even when using highly scalable software in HPC facilities, and (ii) the energy contribution from each atom depends on its position in the supported cluster, making the parametrisation of standard IPs unable to describe energies accurately. We have developed an innovative atom-centred machine-learned IP and a stochastic approach to map efficiently the energy and morphology of multi-metallic supported catalysts. The relative stabilities allow consideration of predominant morphologies under specific temperature conditions, leading to accurate representations of supported catalysts.

Progress on the simulation of nanostructures is routed on an accurate description of the atoms forming it. At the nanoscale, these atoms are surrounded by a variable environment, i.e., support, metal neighbours, and molecular species. They are also placed in sites with different electronic structures and properties, e.g., facets, edges, and corners. We investigated these media and derived a universal fingerprint that unequivocally describes each atom. The fingerprint is based on the symmetry function by Behler et al. and structural features resulting from graph theory. The description of atomic energies and forces assigned to each fingerprint forms the database to train a machine-learning model with a relatively simple neural network architecture. Validation of the learning model leads to mean absolute errors (MAE) of 0.003 eV.atom−1 and 0.035 eV.Å−1 in mono- and bimetallic structures. We used such accurate prediction and an unbiased genetic algorithm to generate, optimise, and evaluate thousands of supported metal structures. Continuous symmetry measurements allow the classification of clusters’ shapes, whereas analyses of the relative energies provide the most prevalent morphologies under specific temperatures, e.g., reaction temperature.,

Link: https://www.iqtc.ub.edu/events/iqtc-seminar-alberto-roldan/.