scibmad is a bridge between PyTorch and Julia for differentiable scientific workflows. It lets you define or load Julia functions, call them from Python, and keep gradients flowing through PyTorch autograd.
- PyTorch tensor to Julia value conversion
- Autograd-aware wrappers for Julia functions
- Beamline helpers for SciBmad element tracking
- A Julia-style object API for screenshot-style beamline workflows
- A small Python API for loading Julia code into
Main
- Python 3.11+
- Julia 1.9+
- A working SciBmad-compatible Julia environment
Install the package:
pip install git+https://github.com/ChanpiseyU/scibmad.gitCore Python dependencies:
pip install torch juliacall numpyIf your Julia environment does not already include the required packages, scibmad will bootstrap SciBmad during initialization.
from scibmad import core as scibmad
import torch
scibmad.define(
"""
f_square(x) = x^2
f_scaled_sum(scale, arr) = scale * sum(arr)
"""
)
x = torch.tensor(3.0, dtype=torch.float64, requires_grad=True)
y = scibmad.f_square(x)
y.backward()
print(float(y)) # 9.0
print(float(x.grad)) # 6.0The recommended beamline interface mirrors Julia SciBmad closely:
import scibmad as sb
import torch
# Make a FODO cell
qf = sb.Quadrupole(Kn1=0.36, L=0.5)
d = sb.Drift(L=1.2)
qd = sb.Quadrupole(Kn1=-0.36, L=0.5)
# 18 GeV electrons
fodo = sb.Beamline(
[qf, d, qd, d],
E_ref=18e9,
species_ref=sb.Species("electron"),
)
coords = torch.tensor(
[1e-3, 0.0, 1e-3, 0.0, 0.0, 0.0],
dtype=torch.float64,
requires_grad=True,
)
qf.Kn1 = torch.tensor(1.2, dtype=torch.float64, requires_grad=True)
qd.Kn1 = -qf.Kn1
result = sb.track(fodo, v0=coords)
loss = result.v[0, :4, -1].sum()
loss.backward()scibmad also keeps the lower-level functional helpers for direct element-by-element tracking:
from scibmad import core as scibmad
import torch
coords = torch.tensor(
[1e-3, 0.0, 1e-3, 0.0, 0.0, 0.0],
dtype=torch.float64,
requires_grad=True,
)
line = scibmad.beamline(
scibmad.drift(1.0, p_over_q_ref=2.0),
scibmad.quadrupole(0.5, 1.2, p_over_q_ref=2.0),
)
result = line(coords)
loss = result.sum()
loss.backward()scibmad.define(code)evaluates Julia code inMainscibmad.load(path)includes a Julia source filescibmad.seval(code)evaluates arbitrary Julia expressionsscibmad.using(*packages)loads Julia packages withusingscibmad.Quadrupole(...),scibmad.Drift(...),scibmad.Beamline(...), andscibmad.track(...)provide a Julia-style beamline API- Julia functions defined in
Mainbecome available asscibmad.<name>
Run the focused test suite with:
python run_tests.pyMIT