Using Machine Learning to Understand Mechanical Loss in LIGO Mirror Coatings

The sensitivity of future generations of gravitational wave detectors such as LIGO will be limited by thermal noise from the mirror coatings on the test masses. A major source of thermal noise is the mechanical loss of the coatings. The mechanical loss at temperatures higher than 10 K is believed to be the result of thermally activated transition between two level systems (TLSs) in the amorphous coating. By using a multilayer perceptron neural network to understand the atomic structures of many computer-generated two-level systems (TLS) of amorphous tantala (a-Ta₂O₅), a prospective coating material, we aim to identify the structural motifs that contribute to mechanical loss. From an analysis of many independent TLSs, we can identify which atoms are most likely to relax by comparing changes in atomic energies and positions. Values that describe the structural features of these atoms can then be computed using angular and radial distribution functions. Using these predictions, a common pattern of relationship between structure and relaxation behavior can be extracted. The results will likely aid in understanding what gives rise to the measured values of mechanical loss and will help determine the best materials and deposition parameters for lower mechanical loss mirror coatings.