This work presents the application of machine learning to the estimation of Adaptive Optics (AO) parameters using telemetry data from the AO system. To improve the estimation accuracy of the Strehl ratio, seeing, wind speed and outer scale, we leverage the capabilities of Recurrent Neural Networks (RNNs). Estimating these optical performance and atmospheric parameters is a key problem in AO. Another critical aspect is obtaining a measure of these quantities as seen from the Wavefront Sensor (WFS) itself, which provides a precise indication of their impact on the image quality. Analytical approaches to compute some of these quantities from WFS telemetry data have been proposed in the literature; however, some parameters are difficult to estimate and measure and may require approximations that limit the estimate's reliability. In recent years, other statistical approaches based on the large amount of data available in WFS telemetry have become feasible. In this contribution, we propose a novel approach based entirely on machine learning algorithms. These algorithms utilize the data produced by the WFS (mirror commands in our case) to perform the task. We then compare our results to the state-of-the-art methods. This preliminary approach uses simulated SOUL (LBT AO system) data generated by the PASSATA software: in the future we plan to implement a similar approach on real data.