Workshop Proceedings

X. Pang1, S. Thulasidasan2, L. Rybarcyk2 1Apple, 2Los Alamos National Laboratory Machine Learning for Engineering Modeling, Simulation, and Design Workshop at Neural Information Processing Systems 2020 Abstract We describe an approach to learning optimal control policies for a large, linear particle accelerator using deep reinforcement learning coupled with a high-fidelity physics engine. The framework consists of an AI controller that uses deep neural networks for state and action-space representation and learns optimal policies using reward signals that are provided by the physics simulator. For this work, we only focus on controlling a small section of the entire accelerator. Nevertheless, initial results indicate that we can achieve better-than-human level performance in terms of particle beam current and distribution. The ultimate goal of this line of work is to substantially reduce the tuning time for such facilities by orders of magnitude, and achieve near-autonomous control. ...

A. L. Edelen1, S. G. Biedron2, J. P. Edelen3, S. V. Milton4, P. J. M. van der Slot5 1Colorado State University, 2Element Aero, 3Fermi National Accelerator Laboratory, 4Los Alamos National Laboratory, 5University of Twente Workshop on Deep Learning for Physical Sciences at the Conference on Neural Information Processing Systems 2017 Abstract Free Electron Laser (FEL) facilities often must accommodate requests for a varietyof electron beam parameters in order to supply scientific users with appropriatephoton beam characteristics. This usually requires skilled human operators to tunethe machine. In principle, a neural network control policy that is trained on a broadrange of machine operating states could be used to quickly switch between theserequests without substantial need for human intervention. We present preliminaryresults from an ongoing simulation study in which a neural network control policyis investigated for rapid switching between beam parameters in a compact THzFEL that exhibits nonlinear electron beam dynamics. To accomplish this, we firsttrain a feed-forward neural network to mimic a physics-based simulation of theFEL. We then train a neural network control policy by first pre-training it as aninverse model (using supervised learning with a subset of the simulation data) andthen training it more extensively with reinforcement learning. In this case, thereinforcement learning component consists of letting the policy network interactwith the learned system model and backpropagating the cost through the modelnetwork to the controller network. ...