Demonstration of machine learning-enhanced multi-objective optimization of ultrahigh-brightness lattices for 4th-generation synchrotron light sources

Fourth-generation storage rings enabled by multi-bend achromat lattices are being inaugurated worldwide and many more are planned for the next decade. These sources deliver stable ultra-high brightness radiation with unmatched levels of transverse coherence by virtue of their highly advanced magnetic lattices. Optimization of these challenging and strongly nonlinear lattices with many degrees of freedom bounded by extensive sets of constraints and multiple often conflicting optimization goals is highly demanding and requires application of the most advanced numerical tools available to the community. While multi-objective genetic algorithms have been very successful in supporting these optimization efforts, the algorithms suffer from a fundamental limitation of their stochastic nature: an exceedingly vast number of candidate lattices, most of which eventually are rejected, has to be fully evaluated. This comes at immense computational cost and thus drives excessive runtime despite use of large supercomputing clusters. We therefore propose to employ deep learning techniques and iterative retraining of neural networks to massively accelerate such lattice evaluation, thereby allowing lattice optimization to rely on far fewer a priori assumptions, open up to larger search ranges, and include right from the start and in parallel multiple error distributions to find truly global optima, all while completing a full optimization campaign in weeks rather than months. In this paper we present the neural network designs, the deep learning approach, iterative retraining procedures, and demonstrate how these machine learning techniques can be incorporated into existing state-of-the-art optimization workflows with only minimal changes applied to the optimization pipeline itself and none at all to the employed tracking codes.