pyFTLE 1.0.3

Python package to compute Finite Time Lyapunov Exponents (FTLE).

pyFTLE: A Python Package for Computing Finite-Time Lyapunov Exponents

pyFTLE computes hyperbolic Lagrangian Coherent Structures (LCS) from velocity flow field data using Finite-Time Lyapunov Exponents (FTLE).

---

OVERVIEW

pyFTLE is a modular, high-performance package for computing FTLE fields. It tracks particle positions over time by integrating trajectories in a velocity field. Then, the flow map Jacobian is computed, and the largest eigenvalue of the Cauchy-Green deformation tensor determines the FTLE field.

Figure 1: FTLE field over an airfoil.

Figure 2: FTLE field of a 3D ABC flow.

Key Features

• Supports both 2D and 3D velocity fields (structured or unstructured).
• Parallel computation of FTLE fields.
• Flexible particle integration strategies.
• Multiple velocity interpolation methods for particle positions.
• SIMD-optimized C++ backend for efficient 2D and 3D interpolations on regular grids.
• Extensible design supporting multiple file formats.
• Modular, well-structured codebase for easy customization and extension.

---

INSTALLATION

Requirements

• Python 3.10+

Using UV (Recommended)

UV is a modern Python package and project manager that simplifies dependency management.

Installation Steps:

1. Clone the repository:

   git clone https://github.com/las-unicamp/pyFTLE.git
   cd pyFTLE
   

2. Install the package (this automatically installs the SIMD-optimized C++/Eigen backend):

   uv tool install .
   

After installation, pyftle is available globally. This means you can import it in Python scripts or notebooks using:

import pyftle

and also run the CLI tool directly from the terminal with:

pyftle -c config.yaml

To uninstall, just run uv tool uninstall pyftle

Using Docker

You can run pyFTLE inside a Docker container without installing dependencies locally. The Docker image includes all required dependencies and the compiled C++ extension.

Building the Docker Image

From the repository root directory:

docker pull lasunicamp/pyftle:latest

Refer to section Running pyFTLE with Docker for more details about the Docker workflow.

---

USAGE

The code features both a clean, CLI-oriented architecture (utilizing configuration files and file-based I/O) and a lightweight, notebook-friendly API. The latter allows you to run small-scale examples entirely in memory, eliminating the need to handle intermediate files, which makes it perfect for demonstrating in Jupyter notebooks. Several such notebooks, located in the examples/ folder, combine analytical velocity fields with visual explanations to illustrate the FTLE solver’s execution.

[!TIP] For production runs, it is often more practical to read velocity and coordinate data directly from the file system (HD/SSD). In this case, the file-based CLI offers greater convenience and flexibility.

[!NOTE] At present, the solver accepts MATLAB (.mat) files as input and exports results in MATLAB (.mat) or VTK (.vts, .vtp) formats. File I/O is implemented using SciPy, so MATLAB itself is not required. The modular I/O subsystem allows developers to integrate additional file formats with minimal changes.

[!IMPORTANT] Input data is provided in MATLAB file format, grouped into three types: velocity, coordinate, and particle files.

• Velocity files contain the velocity field data, where each scalar component (e.g., velocity in the x, y, and z directions) is provided in separate columns. Each column header must be properly labeled (velocity_x, velocity_y, and velocity_z for 3D cases), with the corresponding scalar velocity values at each point in the grid.

• Coordinate files specify the positions where the velocity measurements were taken. The headers must correspond to the spatial coordinates (coordinate_x, coordinate_y, and coordinate_z for 3D cases). These coordinates map directly to the points where the velocity field data in the corresponding velocity file is measured.

• Particle files define groups of neighboring particles used to calculate the FTLE field and more precisely compute the deformation of the Cauchy-Green tensor. In contrast to the other files, each row in the particle file contains a set of coordinates (a tuple of [float, float] for 2D, or [float, float, float] for 3D). The columns specify the relative positions of particles in the group, and the values represent the coordinates of neighboring particles. These tuples help to define the spatial relationships that are critical for computing tensor deformations in the flow field. The neighboring particles are illustrated in the accompanying figure.

This structure ensures that the velocity data, coordinate information, and neighboring particle relations are clearly organized and ready for FTLE computation.

Figure 3: A single group of neighboring particles.

Figure 4: Particles centroids being tracked.

Instead of passing individual MATLAB files directly to the solver, the interface expects a set of plain text (.txt) files—one for each data type: velocity, coordinate, and particle data. Each of these .txt files should contain a list of file paths to the corresponding .mat files, with one path per line. For example, the velocity .txt file will list all the velocity MATLAB files (one per line), and similarly for the coordinate and particle .txt files. This approach allows the solver to process sequences of time-resolved data more easily and keeps the input interface clean and scalable.

[!TIP]

• The create_list_of_input_files.py facilitates the creation of these .txt files.
• An complete example of file-based I/O workflow is provided in the Jupyter Notebooks in the example/ folder.

Running the code via CLI

The script requires several parameters, which can be passed through the command line or a configuration file (config.yaml) located in the root directory. Among these parameters are .txt files that indicates the location of the input files in Matlab format (the velocity field, coordinates and particles).

[!TIP] Once the parameters are properly set, the solver can be executed from the root directory with the following command:

pyftle -c config.yaml

Alternatively, you can run the script from the CLI as:

pyftle \
    --experiment_name "my_experiment" \
    --list_velocity_files "velocity_files.txt" \
    --list_coordinate_files "coordinate_files.txt" \
    --list_particle_files "particle_files.txt" \
    --snapshot_timestep 0.1 \
    --flow_map_period 5.0 \
    --integrator "rk4" \
    --interpolator "cubic" \
    --num_processes 4 \
    --output_format "vtk" \
    --flow_grid_shape 100,100,100 \  # comment this line for unstructured data
    --particles_grid_shape 100,100,100  # comment this line for unstructured data

For VSCode users, the script execution can be streamlined via .vscode/launch.json.

To see the complete list of command-line options and their descriptions, simply run:

pyftle --help
# or equivalently
pyftle -h

This will display all available parameters, their default values, and usage examples directly in the terminal.

⚙️ Full List of CLI Parameters (click to expand)

Required Parameters

Parameter	Type	Description
experiment_name	str	Name of the subdirectory where the FTLE fields will be saved.
list_velocity_files	str	Path to a text file listing velocity data files.
list_coordinate_files	str	Path to a text file listing coordinate files.
list_particle_files	str	Path to a text file listing particle data files.
snapshot_timestep	float	Timestep between snapshots (positive for forward-time FTLE, negative for backward-time FTLE).
flow_map_period	float	Integration period for computing the flow map.
integrator	str	Time-stepping method (euler, ab2, rk4).
interpolator	str	Interpolation method (cubic, linear, nearest, grid).
num_processes	int	Number of workers in the multiprocessing pool. Each worker computes the FTLE of a snapshot.
output_format	str	Output format (mat, vtk).

Optional Parameters

Parameter	Type	Description
flow_grid_shape	list[int]	Grid shape for structured velocity measurements. It must be a comma-separated list of integers.
particles_grid_shape	list[int]	Grid shape for structured particle points. It must be a comma-separated list of integers.

Interpolation behavior depends on whether your velocity data is structured or unstructured:

• If flow_grid_shape is not provided, the velocity field is treated as unstructured. In this case, you can use the cubic, linear, or nearest interpolators, which rely on Delaunay triangulations. This approach offers flexibility but comes with higher computational cost.

• If flow_grid_shape is provided, the velocity field is considered structured. You can still choose cubic, linear, or nearest, but interpolation becomes significantly faster because it exploits the rectilinear grid structure of the data.

• For maximum performance on regular structured grids, pyFTLE includes custom bi- and trilinear interpolators implemented in C++/Eigen, achieving up to 10× speedup compared to SciPy’s implementation. To use this optimized backend, specify flow_grid_shape and set interpolator to grid.

The parameter particles_grid_shape is optional and mainly affects how results are written to disk. If the particle centroids form a regular grid, defining this parameter enables structured output—making post-processing and visualization more straightforward.

FTLE Computation Details

The parameters snapshot_timestep and flow_map_period together define the temporal window used to compute each FTLE field. The number of FTLE fields that will be produced is determined by the number of available velocity snapshots and the integration period of the flow map:

N_FTLE = N_snapshots - (flow_map_period / snapshot_timestep)

For example, suppose:

• list_velocity_files lists 100 velocity files,
• snapshot_timestep = 0.01, and
• flow_map_period = 0.1.

In this case, each FTLE field requires 10 consecutive velocity snapshots to integrate the flow map, so only 90 FTLE fields will be computed (one for each valid starting snapshot).

This ensures that the temporal integration for each FTLE remains consistent with the specified flow map period.

Running pyFTLE with Docker

(click to expand)

Run the pyftle CLI tool directly:

docker run --rm lasunicamp/pyftle:latest --help

Run with a configuration file (mount your data directory):

docker run --rm \
    -v /path/to/your/data:/data \
    -w /data \
    lasunicamp/pyftle:latest -c config.yaml

Run with command-line arguments:

docker run --rm \
    -v /path/to/your/data:/data \
    -w /data \
    lasunicamp/pyftle:latest \
    --experiment_name "my_experiment" \
    --list_velocity_files "velocity_files.txt" \
    --list_coordinate_files "coordinate_files.txt" \
    --list_particle_files "particle_files.txt" \
    --snapshot_timestep 0.1 \
    --flow_map_period 5.0 \
    --integrator "rk4" \
    --interpolator "cubic" \
    --num_processes 4 \
    --output_format "vtk"

Interactive Usage

Start an interactive shell inside the container:

docker run -it --rm --entrypoint sh lasunicamp/pyftle:latest

Inside the container, you can run:

pyftle --help
python -c "import pyftle; from pyftle import AnalyticalSolver"

[!TIP] Use volume mounts (-v) to access your data files and configuration files from inside the container. The -w flag sets the working directory inside the container.

Examples:

# Mount current directory and set it as working directory
docker run --rm \
    -v $(pwd):/data \
    -w /data \
    lasunicamp/pyftle:latest -c config.yaml

# Mount specific data directory
docker run --rm \
    -v /path/to/velocity/data:/data/velocity \
    -v /path/to/output:/data/output \
    -w /data \
    lasunicamp/pyftle:latest \
    --experiment_name "my_experiment" \
    --list_velocity_files "velocity/velocity_files.txt" \
    --list_coordinate_files "velocity/coordinate_files.txt" \
    --list_particle_files "velocity/particle_files.txt" \
    --snapshot_timestep 0.1 \
    --flow_map_period 5.0

# Mount configuration file from host
docker run --rm \
    -v $(pwd)/config.yaml:/app/config.yaml \
    -w /app \
    lasunicamp/pyftle:latest -c config.yaml

---

REFERENCES

A list of scientific works using pyFTLE includes:

1. de Souza, Miotto, Wolf. Active flow control of vertical-axis wind turbines: Insights from large-eddy simulation and finite-time resolvent analysis. Journal of Fluids and Structures, 2025.
2. de Souza, Wolf, Safari, Yeh. Control of Deep Dynamic Stall by Duty-Cycle Actuation Informed by Stability Analysis. AIAA Journal, 2025.
3. Lui, Wolf. Interplay between streaks and vortices in shock-boundary layer interactions with conditional bubble events over a turbine airfoil. Physical Review Fluids, 2025.
4. Lui, Wolf, Ricciardi, Gaitonde. Analysis of Streamwise Vortices in a Supersonic Turbine Cascade. AIAA Aviation Forum and Ascend, 2024.

---

LICENSE

This project is licensed under the MIT License.

---

CONTRIBUTING

When contributing to this repository, please make sure to follow the guidelines from the CONTRIBUTING file.

To make sure the Language Server Protocol (LSP) is going to work as expected, one can install the package as follows:

1. Install dependencies using UV:

   uv sync --all-extras
   

2. Install src/ directory as an editable package:

   uv pip install -e '.[dev,test]' --verbose
   

   • This installs src/ as an editable package, allowing you to import modules directly and modify the code during development.
   • The command also automatically installs the SIMD-optimized C++/Eigen backend.
   • Installing in editable mode helps avoid common import issues during development.

We use pytest for unit tests. To run the entire test suit, we recommend the following command in the base directory of the repository:

PYTHONPATH=${PWD} uv run python -m pytest

For Developers Using Docker

If you are using Docker, you might want to build the docker image locally:

docker build -t pyftle:latest .

All commands shown in section Running pyFTLE with Docker will work, but we need to remove the lasunicamp/ from it to tell Docker to use the source code version and not the version from the Docker Hub.

---

FUNDING

The authors acknowledge Fundação de Amparo à Pesquisa do Estado de São Paulo, FAPESP, for supporting the present work under research grants No. 2013/08293-7, 2019/17874-0, 2021/06448-0, 2022/09196-4, 2022/08567-9, and 2024/20547-9. Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) is also acknowledged for supporting this research under grant No. 304320/2024-2.

---

MAIN DEVELOPERS

• Renato Fuzaro Miotto
• Lucas Feitosa de Souza
• William Roberto Wolf

---

For bug reports, feature requests, or contributions, please open an issue or submit a pull request on GitHub.