[Depfis] Seminario de inteligancia artificial aplicada a cristales moleculares - Hoy 12:30 INQUIMAE

[Uriel Nicolas Morzan]

Buen dia, les comparto este seminario, que se llevara a cabo hoy en el  
aula de seminarios del INQUIMAE a las 1230.

Abrazos,

Uriel

===================================================
Machine learning as a tool for understanding crystallization from  
first principles

Crystallization is a process of key importance for many modern  
technologies, such as the manufacturing of pharmaceuticals, and it  
also plays a central role in geological, planetary, and climate  
sciences. Over the years, molecular dynamics simulations have provided  
key insights into the microscopic mechanisms underlying crystal  
nucleation and growth processes. However, classical molecular dynamics  
simulations based on semi-empirical force fields are limited in  
accuracy and cannot describe important physical phenomena such as bond  
forming and breaking. On the other hand, ab initio molecular dynamics  
overcome some of these limitations, yet are limited to very small  
system sizes and short total simulation times. These shortcomings of  
classical and ab initio molecular dynamics have resulted in a  
significant knowledge gap in the microscopic mechanisms of  
crystallization. In this talk, I will discuss a new approach to this  
problem based on using machine learning to train interatomic  
potentials on large datasets of ab initio calculations. Such  
potentials can be used to drive large-scale, highly-accurate and  
reactive molecular dynamics simulations of crystallization phenomena.  
I will illustrate this approach with several examples. First, I will  
show that we can leverage this tool to compute homogeneous ice  
nucleation rates from first principles which are in remarkable  
agreement with experiment [1]. Furthermore, I will present some  
results about the formation of ice on feldspar, the most important ice  
nucleating particle in the atmosphere [2] . Finally, I will discuss  
the application of this technique to study the crystallization of  
calcium carbonate from aqueous solution, a process where reactivity  
plays an essential yet poorly understood role [3]. Taken together,  
these results show the great promise of machine learning as a tool to  
bridge time and length scales, and to provide insight into complex  
phenomena which were thought to be out of reach for molecular  
simulation.

[1] Piaggi, Weis, Panagiotopoulos, Debenedetti, and Car, Proc. Natl.  
Acad. Sci. 119, 33 (2022)
[2] Piaggi, Selloni, Panagiotopoulos, Car, and Debenedetti, Faraday  
Discuss. 249, 98 (2024)
[3] Piaggi, Gale, and Raiteri, arXiv:2409.18562 (2024)