Artificial intelligence applied to the automatic analysis of absorption spectra. Objective measurement of the fine structure constant

A new and automated method is presented for the analysis of high-resolution absorption spectra. Three established numerical methods are unified into one "artificial intelligence" process: a genetic algorithm (GVPFIT); non-linear least-squares with parameter constraints (VPFIT); and Bayesian Model Averaging (BMA). The method has broad application but here we apply it specifically to the problem of measuring the fine structure constant at high redshift. For this we need objectivity and reproducibility. GVPFIT is also motivated by the importance of obtaining a large statistical sample of measurements of $\Delta\alpha/\alpha$. Interactive analyses are both time consuming and complex and automation makes obtaining a large sample feasible. In contrast to previous methodologies, we use BMA to derive results using a large set of models and show that this procedure is more robust than a human picking a single preferred model since BMA avoids the systematic uncertainties associated with model choice. Numerical simulations provide stringent tests of the whole process and we show using both real and simulated spectra that the unified automated fitting procedure out-performs a human interactive analysis. The method should be invaluable in the context of future instrumentation like ESPRESSO on the VLT and indeed future ELTs. We apply the method to the $z_{abs} = 1.8389$ absorber towards the $z_{em} = 2.145$ quasar J110325-264515. The derived constraint of $\Delta\alpha/\alpha = 3.3 \pm 2.9 \times 10^{-6}$ is consistent with no variation and also consistent with the tentative spatial variation reported in Webb et al (2011) and King et al (2012).