Devito

Devito - Symbolic PDE Solver

Quick Reference

pythonCopy code
from devito import Grid, Function, TimeFunction, Eq, solve, Operator

# Create grid
grid = Grid(shape=(101, 101), extent=(1000., 1000.))

# Velocity model
v = Function(name='v', grid=grid, space_order=4)
v.data[:] = 1500.

# Wavefield
p = TimeFunction(name='p', grid=grid, time_order=2, space_order=4)

# Wave equation: d2p/dt2 = v^2 * laplacian(p)
stencil = Eq(p.forward, solve(p.dt2 - v**2 * p.laplace, p.forward))

# Compile and run
op = Operator([stencil])
op(time_M=100, dt=0.5)

Key Classes

Essential Operations

Grid and Fields

pythonCopy code
from devito import Grid, Function, TimeFunction

# 2D/3D Grid
grid = Grid(shape=(nx, nz), extent=(x_size, z_size))

# Velocity model (spatial field)
v = Function(name='v', grid=grid, space_order=4)
v.data[:] = 1500.

# Wavefield (time-dependent)
p = TimeFunction(name='p', grid=grid, time_order=2, space_order=4)

Source and Receivers

pythonCopy code
from examples.seismic import RickerSource, Receiver, TimeAxis

time_range = TimeAxis(start=0., stop=1000., step=dt)

# Source
src = RickerSource(name='src', grid=grid, f0=10., npoint=1, time_range=time_range)
src.coordinates.data[0, :] = [500., 20.]

# Receivers
rec = Receiver(name='rec', grid=grid, npoint=101, time_range=time_range)
rec.coordinates.data[:, 0] = np.linspace(0., 1000., 101)
rec.coordinates.data[:, 1] = 20.

Build and Run

pythonCopy code
# Wave equation
stencil = Eq(p.forward, solve(p.dt2 - v**2 * p.laplace, p.forward))
src_term = src.inject(field=p.forward, expr=src * dt**2 * v**2)
rec_term = rec.interpolate(expr=p)

# Compile and execute
op = Operator([stencil] + src_term + rec_term)
op(time_M=nt-1, dt=dt)

# Results
shot_record = rec.data        # (nt, nrec)
snapshot = p.data[0]          # Current wavefield

Symbolic Derivatives

Stability and Accuracy

CFL Condition: dt < dx / (v_max * sqrt(ndim))

Higher order = more accurate but slower. Use 4-8 for production.

When to Use vs Alternatives

Choose Devito when: You need high-performance finite-difference wave propagation with symbolic equation specification. It auto-generates optimized C/OpenMP/GPU code from Python-level math, making it ideal for FWI, RTM, and research prototyping.

Avoid Devito when: You need finite-element methods (use FEniCS), or simple pedagogical examples where NumPy suffices.

Common Workflows

Acoustic wave forward modelling with sources and receivers

• Define Grid with shape and physical extent matching the velocity model
• Create velocity Function and populate with model values
• Create TimeFunction for the wavefield (time_order=2, space_order=4+)
• Verify CFL condition: dt < dx / (v_max * sqrt(ndim))
• Build wave equation stencil: Eq(p.forward, solve(p.dt2 - v**2 * p.laplace, p.forward))
• Create source (RickerSource) and receivers, set coordinates
• Add source injection and receiver interpolation terms
• Compile Operator with stencil + source + receiver terms
• Run operator: op(time_M=nt-1, dt=dt)
• Extract shot record from rec.data and plot

References

• Operators and Stencils - Detailed operator construction
• Performance Optimization - GPU execution and tuning

Scripts

• scripts/acoustic_wave.py - Basic acoustic wave modeling