Santamaria Garcia, A. Bründermann, E. Caselle, M. De Carne, G. Müller, A.-S. Scomparin, L. Xu, C. JACoW Publishing Geneva, Switzerland 2673-7035 978-3-95450-224-0 10.18429/JACoW-FLS2023-TH3D3 English 249-256 controls operation electron laser feedback Contribution to a conference proceedings 2024 2024-01 https://doi.org/10.18429/JACoW-FLS2023-TH3D3 http://jacow.org/fls2023/papers/th3d3.pdf Machine learning (ML) is one of the key technologies that can considerably extend and advance the capabilities of particle accelerators and needs to be included in their future design. Future light sources aim to reach unprecedented beam brightness and radiation coherence, which require challenging beam sizes and accelerating gradients. The sensitive designs and complex operation modes that arise from such demands will impact the beam availability and flexibility for the users, and can render future accelerators inefficient. ML brings a paradigm shift that can re-define how accelerators are operated. In this contribution we introduce the vision of ML-driven facilities for future accelerators, address some challenges of future light sources, and show an example of how such methods can be used to control beam instabilities.