The difficult problems being tackled in the accelerator community are those that are nonlinear, substantially unmodeled, and vary over time. These problems would be ideal candidates for model-based optimization and control if representative models could be developed. Such models must be inexpensive to deploy and maintain, and must remain valid throughout the operating region of the system and through variations in system dynamics. This project will develop methodology and algorithms for building high-fidelity mathematical representations of complex nonlinear systems via a combination of first-principles and neural network models. In Phase I, empirical data and first-principles information were used to train a combined a neural network model and nonlinear parametric model through constrained nonlinear optimization. The combined model was applied to three challenging problems (local orbit correction in electron storage rings and colliders, beamline model verification in accelerators, and gas composition model in gas-phase polymerization reactors), demonstrating both accuracy and computational efficiency. In Phase II, the combined model will be further developed and deployed in three important applications in particle accelerators: (1) minimizing longitudinal emittance for RF photocathode guns, (2) modeling beam loss at storage rings, and (3) to optimizing local orbit correction at storage rings and colliders. Algorithms for self-validation of these models will be developed, and the diagnostic value of such algorithms will be assessed.
Commercial Applications and Other Benefits as described by the awardee: The new software product should allow the modeler to easily use first-principles information, process data, and operator know-how to build high-fidelity models to address current and future needs in process industries and high-energy physics.