Artificial intelligence unveils a golden (nano) secret

Gold nanoparticles’ unique properties are determined by their structure, which, in turn, shape-shifts as a function of temperature. We adopt data-driven algorithms to model and characterize these changes in an efficient, accurate, and automated fashion.
Artificial intelligence unveils a golden (nano) secret

Gold–what can it not do, and undo? - William Shakespeare

This question gets even more complex when looking at the phase stability of gold in the form of a nanoparticle, a system whose size ranges between 1 to 100 nms, corresponding to 1/100000th of the size of a human hair. 

Gold nanoparticles' unique properties, which are exploited in catalysis, biomedicine, and optics, dramatically change depending on whether these are solid or liquid.  Classical thermodynamics fails at the nano scale because of the non-trivial interplay between volume and surface effects towards the stabilization of a given phase. A mechanistic picture of the melting mechanism of small nanoparticles, as well as a tool to quantitatively predict their melting temperature, is key for their application in cutting-edge technologies.

To simulate what Gold nanoparticles do and don’t do when subject to heating, we resort to computational models which exploit quantum mechanics calculations, machine learning predictions, and molecular mechanics simulations. More specifically, we use machine learning to approximate the slow-but-accurate quantum mechanics predictions for forces acting among atoms. This way, atomic simulations that would take thousands of years if carried out using quantum mechanics first principles calculations, can be done in a few days, dramatically reducing their computational - and environmental - cost.

To analyse what happens inside the simulations we run, we apply unsupervised learning techniques to draw a data-driven criterion to define which ````"phase” each atom is in. Each atom is automatically identified as liquid or solid, and whether it belongs to the surface, the edge, or the inner part of the nanoparticle. This way, we can identify when and how the nanoparticles melt, and what happens at the surface at temperatures lower than the melting temperatures.

Our results, shown in Figure 1,  predict the melting temperatures in good agreement with experiments, and highlight that atoms on the surface of nanoparticles become liquid earlier than the inner ones, at temperatures ranges that agree with experiments.

Figure 1: (left) Melting temperatures for gold nanoparticles of different sizes, computed using our method (green triangles), a classical force field (gray pentagons), and estimated experimentally (pink squares and violet diamonds). (right) Experimental estimate for surface melting temperatures (pink diamonds), and temperature ranges where the surfaces are significantly more liquid than the nanoparticles’ cores, simulated using our method (green triangles).

We foresee the application of the machine learning tools here presented towards addressing the complexity of phase changes in other technologically-relevant systems.

The paper:

Claudio Zeni, Kevin Rossi, Theodore Pavloudis, Joseph Kioseoglou, Stefano de Gironcoli, Richard E. Palmer & Francesca Baletto Data-driven simulation and characterisation of gold nanoparticle melting Nature Communications volume 12, Article number: 6056 (2021) 

Please sign in or register for FREE

If you are a registered user on Nature Portfolio Chemistry Community, please sign in