DiPAS 1.2

DiPAS is a framework for differentiable simulations of particle accelerators.

DiPAS

DiPAS is a program for differentiable simulations of particle accelerators. It acts as a framework and thus supports a wide range of use cases such as particle tracking or optics calculations such as closed orbit search or computation of Twiss parameters.

The involved computations are backed by the PyTorch package which also provides the relevant functionality for differentiation of user-defined quantities as well as a variety of gradient-based optimizers that integrate with the thus derived quantities.

The DiPAS package can parse MADX lattice definitions and hence allows for zero-overhead importing of existing lattices. In addition it supports custom lattice definitions from provided element classes.

Relevant links

• Documentation

• Examples

• PyPI Project

Example usage

Minimizing loss along beamline by tuning quadrupoles:

import numpy
from dipas.build import from_file
from dipas.elements import Quadrupole
import torch

lattice = from_file('example.madx')

for quad in lattice[Quadrupole]:
    quad.k1 = torch.nn.Parameter(quad.k1)

optimizer = torch.optim.Adam(lattice.parameters(), lr=1e-3)

particles = torch.from_numpy(numpy.load('particles.npy'))

while True:
    tracked, loss_val = lattice.linear(particles, recloss='sum')
    lost = 1 - tracked.shape[1] / particles.shape[1]
    if lost < 0.01:  # Fraction of particles lost less than 1%.
        break
    optimizer.zero_grad()
    loss_val.backward()
    optimizer.step()