Optics correction using deep learning.
3 programs: to create dataset (using MAD-X), to find erroneous element (classification problem), to find errors in all elements (regression problem).
Elements: 26 quadupoles. Parameters: field gradients.
Input: beam x and y orbits, 108 points (54 BPMs x2) or beam x and y dbetatron functions in 108 locations (54 BPMs x2). Output: 26 error values.