Riassunto analitico
We have developed a machine learning (ML) approach to accelerate the convergence of the Random Phase Approximation (RPA) linear response functions of 2D semiconductors, a crucial quantity for the GW approximation in the framework of electronic structure Green's function methods. This development applies ML concepts within many-body theory and works as a convergence accelerator, allowing for the simulation, at a reduced cost, of larger and more computationally demanding systems. To design and train the ML model, we have studied three prototype systems (hBN, MoS2 and black phosphorene), treating the response functions as images of different size and quality. Interesting and promising results have been obtained by comparing different neural networks (NNs), especially in terms of generalization to other 2D materials.
|
Abstract
We have developed a machine learning (ML) approach to accelerate the convergence of the Random Phase Approximation (RPA) linear response functions of 2D semiconductors, a crucial quantity for the GW approximation in the framework of electronic structure Green's function methods. This development applies ML concepts within many-body theory and works as a convergence accelerator, allowing for the simulation, at a reduced cost, of larger and more computationally demanding systems. To design and train the ML model, we have studied three prototype systems (hBN, MoS2 and black phosphorene), treating the response functions as images of different size and quality. Interesting and promising results have been obtained by comparing different neural networks (NNs), especially in terms of generalization to other 2D materials.
|
