Research
PhD research
Materials modelling for energy and optoelectronic applications
The main research topic in my PhD is to characterize the structural and optoelectronic properties of novel semiconductor materials, such as chalcohalide anti-perovskites. To achieve this, I use first-principles computational methods, like Density Functional Theory (DFT) and ab-initio Molecular Dynamics simulations. These methods allow for the computation of fundamental properties of solid-state crystal systems with a high level of accuracy and agreement with experiments. Some of the properties that first-principles methods can determine include (but are not limited to) lattice parameters, phonon computations and thermodynamics properties, electronic structure, and optical properties of the system.
The studied materials are expected to be promising candidates for use in energy and optoelectronic applications, such as photovoltaics, photocatalysis, thermoelectricity, solid-state batteries, and neuromorphic computing, among others. Consequently, they may contribute to the transition to a more sustainable and decarbonized future.
Electron-phonon coupling effects in optoelectronic properties
DFT formalism operates at zero temperature, but it is generally sufficient for correctly predicting the optoelectronic properties of condensed matter systems. However, some systems exhibit strong electron-phonon interactions, making it necessary to consider the thermal effects of the lattice to accurately model the optoelectronic properties of these systems. This is the case with chalcohalide anti-perovskite materials, where we have observed a strong dependence of the band gap and dielectric tensor on temperature, with a significant decrease in the band gap as temperature increases.
A screening of various materials was conducted to identify other systems with similarly large temperature-dependent renormalization effects, leading to the identification of several perovskite-like systems. However, in these systems, an increase in the band gap with rising temperature was observed.
The differing effects of temperature on the electronic bands are successfully explained in terms of the overlap of different electronic orbitals near the valence and conduction bands.
Crystal Graph Neural Networks for materials property prediction
Convolutional Graph Neural Networks (CGNNs) have emerged in recent years as an effective machine learning method for predicting the properties of materials. These models are trained on graphs, a mathematical structure used to represent systems of data with relationships, such as social networks, images, or molecules. These graphs consist of nodes and edges, i.e., connections between nodes. Both nodes and edges can have various features. In this way, a material can be represented as a graph where the nodes represent atoms and the edges represent the distances between atoms (i.e., chemical bonds). By allowing more than one edge between the same pair of nodes, we can account for the periodicity of the unit cell.
In our research, we are developing CGNN models to predict band gaps and thermal effects on band gaps, thus avoiding the need for expensive DFT calculations. This approach also enables the computational study of more complex systems, such as solid solutions.
Crystal structure prediction with ML interatomic potentials
The structure of the unit cell in crystal materials and the type of ions present in this unit cell fully determines
the physical and chemical properties of a given material. Thus, correctly predicting the crystal phases of materials is of utmost importance. Crystal Structure Prediction methods emerged to find stable and metastable phases of a given material using first-principles methods. However, these methods often rely on expensive DFT computations and face the challenge of efficiently exploring the energy (or Gibbs energy, when accounting for pressure and temperature) hypersurface, due to its large dimensionality and irregular morphology.
To find the minima of the energy hypersurface, we propose using a random search algorithm along with ionic relaxations using machine learning interatomic potentials. We developed PyMCSP (Python and Machine learning Crystal Structure Prediction) as a user-friendly software to find stable and metastable phases of materials. Additionally, the program can also search for high-pressure phases and perform phase prediction from X-ray and neutron diffraction experimental results.
The github repository to PyMCSP code can be found
here.
