Getting Started

Installation

From PyPI

uv pip install tide-GPR

Or with pip:

pip install tide-GPR

From source

git clone https://github.com/vcholerae1/tide.git
cd tide
uv build

Requirements

Dependency	Version
Python	≥ 3.12
PyTorch	≥ 2.12
CUDA Toolkit	optional, for GPU support
CMake	≥ 3.28, optional, for building from source

First Success Criteria

Minimal 2D Forward Run

import torch
import tide

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
dtype = torch.float32

ny, nx = 96, 96
epsilon = torch.full((ny, nx), 4.0, device=device, dtype=dtype)
sigma = torch.zeros_like(epsilon)
mu = torch.ones_like(epsilon)

nt = 300
dt = 4e-11
src = tide.ricker(freq=8e8, length=nt, dt=dt, device=device, dtype=dtype).view(1, 1, nt)
src_loc = torch.tensor([[[20, 48]]], device=device, dtype=torch.long)
rec_loc = torch.tensor([[[20, 60]]], device=device, dtype=torch.long)

model = tide.EMModel(epsilon, sigma, mu)
operator = tide.MaxwellTM(
    tide.Discretization(0.02, dt, boundary=tide.CPML(10)),
    tide.Experiment(tide.Acquisition(src_loc, rec_loc), src),
    execution=tide.ExecutionOptions(fallback=tide.FallbackPolicy.REFERENCE),
)
receivers = operator(model).receiver_data

print(receivers.shape)  # [nt, n_shots, n_receivers]

Optional 3D Preview

import torch
import tide

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
dtype = torch.float32

nz, ny, nx = 32, 32, 32
epsilon = torch.full((nz, ny, nx), 4.0, device=device, dtype=dtype)
sigma = torch.zeros_like(epsilon)
mu = torch.ones_like(epsilon)

nt = 200
dt = 4e-11
src = tide.ricker(freq=1e8, length=nt, dt=dt, device=device, dtype=dtype).view(1, 1, nt)
src_loc = torch.tensor([[[16, 16, 16]]], device=device, dtype=torch.long)
rec_loc = torch.tensor([[[16, 16, 20]]], device=device, dtype=torch.long)

model = tide.EMModel(epsilon, sigma, mu)
operator = tide.Maxwell3D(
    tide.Discretization(
        [0.03, 0.03, 0.03],
        dt,
        boundary=tide.CPML(6),
    ),
    tide.Experiment(tide.Acquisition(src_loc, rec_loc), src),
    execution=tide.ExecutionOptions(
        backend=tide.BackendPreference.NATIVE,
        fallback=tide.FallbackPolicy.REFERENCE,
    ),
)
rec = operator(model).receiver_data

print(rec.shape)

Verify Backend Availability

from tide import backend_utils

print("backend available:", backend_utils.is_backend_available())
print("library path:     ", backend_utils.get_library_path())

Common Startup Issues

Shape mismatch

source_amplitude must be [n_shots, n_sources, nt].
source_location and receiver_location must be [n_shots, n_points, ndim].

Out-of-bounds indices

Coordinates must satisfy 0 <= index < model_size for each spatial dimension.

Instability warning

TIDE adjusts the internal time step using CFL and resamples time signals. Consider reducing dt or coarsening grid spacing.

Read the first result

The minimal example has one shot, one source, and one receiver. Its result shape is therefore [300, 1, 1]. The first axis is physical time sampled every 4e-11 seconds. The other axes identify shot and receiver.

The returned tensor is differentiable. Confirm this before building an inversion:

epsilon = epsilon.clone().requires_grad_(True)
result = operator(tide.EMModel(epsilon, sigma, mu))
objective = result.receiver_data.square().mean()
objective.backward()

print("loss:", float(objective.detach()))
print("gradient shape:", epsilon.grad.shape)
print("finite gradient:", bool(torch.isfinite(epsilon.grad).all()))

This is a computational check, not a physical validation. A finite gradient can still be wrong because of units, geometry, insufficient resolution, or a misinterpreted component.

Make one controlled change

Change a compact region of relative permittivity and compare traces:

epsilon_perturbed = epsilon.detach().clone()
epsilon_perturbed[45:55, 50:60] = 7.0

baseline = operator(tide.EMModel(epsilon.detach(), sigma, mu)).receiver_data
perturbed = operator(
    tide.EMModel(epsilon_perturbed, sigma, mu)
).receiver_data

difference = perturbed - baseline
print("maximum trace change:", float(difference.abs().max()))

This exercise establishes the core forward-modeling contract: the operator and experiment stay fixed, the material model changes, and receiver data records the effect.

Understand internal sub-stepping

The source and returned receiver tensors use the dt supplied to Discretization. TIDE may choose a smaller internal time step to satisfy the CFL condition:

from tide.cfl import cfl_condition

inner_dt, step_ratio = cfl_condition(
    grid_spacing=0.02,
    dt=dt,
    max_vel=299_792_458.0,
)
print(inner_dt, step_ratio)

When step_ratio > 1, propagation performs several internal updates per user sample. TIDE upsamples the source and downsamples receiver data automatically. The output still contains nt samples, but runtime increases with the ratio.

First physical checks

Before replacing the homogeneous model with survey data:

Estimate the direct-arrival time from source-receiver distance and expected material velocity.
Confirm the trace becomes active near that time.
Move the receiver farther away and confirm the arrival is later.
Increase CPML width and confirm the direct arrival stays fixed while late boundary energy changes.
Repeat with a finer grid or higher stencil order and compare phase.

These checks reveal coordinate, unit, boundary, and dispersion errors earlier than an inversion loop will.