Understanding mechanical dissipation in glasses through machine learning molecular dynamics

The sensitivity of the present generation of gravitational waves detectors is limited by the magnitude of thermal noise affecting the mirrors used in such devices. These can be seen as light interferometers. The mirrors are made by alternating layers of vitreous silica (SiO2) and vitreous tantala (Ta2O5) doped with Titanium. From the fluctuation-dissipation theorem, thermal noise is determined by mechanical dissipations. In the frequency region interesting for gravitational waves detectors (10Hz-10 kHZ), the origin of them is still debated. They have been modelled as thermally activated transitions between two different structural configurations (TLC) of the atomic structures of the glass. So far, only atomic structures obtained from classical force-fields have been investigated. Unfortunately, their accuracy is quite poor. Density-functional theory (DFT) gives accurate results, but its computational cost makes it unpractical. With this thesis we want to overcome the limitations in terms of accuracy of the classical force-fields replacing them with machine learning potentials (MLPs) reaching the accuracy of state-of-the-art density functional theory at a fraction of the cost. We will focus on vitreous tantala as it gives the main source of noise. We will fit an extensive set of snapshots from DFT molecular dynamics simulations with the Allegro MLP. For assessing the degree of confidence of the obtained MLPs, we will first build with a quench-from-the melt scheme a model of vitreous Ta2O5 and then we will calculate for it a set of properties as the pair correlation functions, the spectrum of vibrations, the structure factors to be compare with DFT and experimental results. Using the validated MLPs we will generate few models of vitreous Ta2O5 and for them we will identify a set of TLCs. Their energy parameters together with the dependence on external strain will permit the calculation of the mechanical dissipation factor Q-1.