Research
-
ML POTENTIALS
Neural networks potentials for materials and chemical reactions
The development of the machines learning potentials for phase transitions and reactive events is challenging due to the difficulty of including all relevant configurations in the training set. Using enhanced sampling techniques such as Metadynamics and multithermal-multibaric simulations we were able to expand the region of the potential energy surface covered by the reference configurations and thus to develop reliable neural networks potentials. This allowed us to perform ab-initio simulations of complex processes such as silicon crystallization [1], gallium nucleation and its phase diagram [2] and urea decomposition in water [3], which would otherwise be prohibitively expensive.[1] Physical Review Letters
(2018) , 121, 265701
[2] Nature Communications(2020) , 11, 2654
[3] arXiv:2011.11455(2020)
-
ENHANCED SAMPLING
Improve enhanced sampling methods with machine learning techniques
A popular strategy to overcome kinetic bottlenecks in atomistic simulations is to identify a number of key collective variables and to introduce an external bias potential that is able to accelerate sampling by favoring their fluctuations.
I developed a variant of the variationally enhanced sampling method, in which the bias potential is represented as a neural network [1]. The bias is optimized on-the-fly with a reinforcement learning-like scheme, which minimizes the Kullback-Leibler divergence between the sampled and the target distribution. Using a neural network rather than a linear basis expansion allows to better represent represent complex free-energy surfaces and to handle several collective variables.[1] Proceedings of the National Academy of Sciences
(2019) , 116 (36)
GitHub repository -
COLLECTIVE VARIABLES
Data-driven identification of collective variables for enhanced sampling
Fundamental to the success of several enhanced sampling methods is the identification of adequate collective variables, whose fluctuations should be related to all slow degrees of freedom of the rare event of interest. To facilitate their design in case of complex systems, I have developed a data-driven approach that builds these variables from the knowledge of metastable states alone [1]. Here a neural network provides a non-linear featurization of the possibly many inputs into a reduced latent space, where a classification technique is used to obtain the collective variables. This method has been successfully employed also for the calculation of ligand binding free energies, where the analysis of the relevance of the input descriptors helped to clarify the role played by water in the binding process [2]. A synthetic communication is available here [3].[1] Journal of Physical Chemistry Letters
(2020) , 11, 8
[2] Nature Communications(2021) , 12, 93
[3] arXiv:2101.07085(2021)
GitHub repository