DeepFMKit: A Toolkit for Deep Frequency Modulation Interferometry

DeepFMKit is a high-performance Python framework for the end-to-end simulation, processing, and analysis of data from Deep Frequency Modulation Interferometry (DFMI) experiments. It provides a complete computational laboratory for researchers and engineers to design virtual experiments, study systematic errors, and develop and benchmark novel data analysis techniques.

About The Project

Deep Frequency Modulation Interferometry (DFMI) is a laser-based metrology technique used for high-precision displacement sensing and absolute distance measurement. This toolkit is designed to handle the entire experimental workflow, from simulating complex interferometric signals with realistic noise sources to performing robust, high-speed parameter estimation.

The novelty of DeepFMKit lies in its integrated, object-oriented approach. It combines a detailed physics engine—capable of modeling time-of-flight delays, colored noise sources, and arbitrary modulation waveforms—with a modular set of interchangeable fitting algorithms. This allows users to not only simulate realistic data but also to benchmark the performance of different readout schemes on identical, well-defined datasets.

Key Features

• High-Fidelity Physics Engine: Simulate realistic DFMI signals with a model that includes prescribed noise sources, dynamic path length changes, and arbitrary laser modulation waveforms.
• Comprehensive Noise Modeling: Inject realistic, colored noise based on user-defined Amplitude Spectral Densities (ASDs) for laser frequency, laser amplitude, modulation amplitude, and arm length fluctuations.
• Modular & Interchangeable Fitters:
  • NLS Fitter: A highly optimized, parallelized Non-Linear Least Squares (NLS) fitter working in the frequency domain.
  • EKF Fitter: A time-domain Extended Kalman Filter (EKF) for real-time state tracking and analysis of dynamic systems.
• High-Throughput Experimentation: A declarative framework for defining and running large-scale, parallelized parameter sweeps and Monte Carlo simulations to systematically characterize system performance.
• Data Handling & Visualization: A suite of tools for loading/saving data, managing experimental configurations, and plotting results.

Getting Started

Prerequisites

• Python 3.9+
• Standard scientific libraries: NumPy, SciPy, Pandas, Matplotlib
• For high-performance noise generation: Numba

Installation

You can either install directly from GitHub or clone the repository locally.

Option 1 – Install directly from GitHub (latest version):

pip install git+https://github.com/mdovale/DeepFMKit.git

Option 2 – Clone and install manually:

1. Clone this repository

   git clone https://github.com/mdovale/DeepFMKit.git
   cd DeepFMKit
   

2. (Recommended) Create and activate a virtual environment

   python -m venv .venv
   source .venv/bin/activate     # On Windows: .venv\Scripts\activate
   

3. Install dependencies

   pip install --upgrade pip
   pip install -r requirements.txt
   

Quick Start: A Complete Workflow

The following example demonstrates the primary workflow: defining a physical system, simulating data, running a fit, and plotting the results.

import deepfmkit.core as dfm
import matplotlib.pyplot as plt
# --- 1. Define the interferometer ---
ifo = dfm.IfoConfig()
ifo.ref_arml = 0.10 # Reference armlength (m)
ifo.meas_arml = 0.15 # Measurement armlength (m)
# --- 2. Define the laser source ---
las = dfm.LaserConfig()
las.fm = 1e3 # FM frequency (Hz)
las.set_df_for_m(ifo, 6.) # Target m (rad)
# --- 3. Compose the main channel ---
sim = dfm.SimConfig("ch0",
    laser_config=las,
    ifo_config=ifo,
    f_samp=200e3) # Acquisition frequency (Hz) 
# --- 4. Instantiate DeepFrame ---
df = dfm.DeepFrame(sim_config=sim)
# --- 5. Simulate ---
df.simulate("ch0",
    n_seconds=10) # Simulation length (s)
# --- 6. Run readout algorithm ---
df.fit("ch0")
# --- 7. Plot readout results ---
axes = df.plot(which=['phi', 'm', 'ssq'])
plt.show()

Advanced Usage

High-Fidelity Simulation with Colored Noise

The 'asd' simulation mode enables realistic noise modeling. Configure the noise properties of the LaserConfig and IfoConfig objects before simulating.

from deepfmkit.physics import LaserConfig, IfoConfig, SimConfig
# Assume 'dff' and 'label' exist from the Quick Start example
# and laser/ifo variables are still in scope.

# Configure laser noise (ASD at 1 Hz, and 1/f^alpha exponent)
laser.f_n = 100.0      # 100 Hz/sqrt(Hz) of laser frequency noise
laser.f_n_alpha = 1.0  # Pink (1/f) frequency noise

# Configure interferometer noise
ifo.arml_n = 1e-12     # 1 pm/sqrt(Hz) of arm length noise
ifo.arml_n_alpha = 2.0 # Red (1/f^2) arm length noise

# Simulate using the 'asd' mode
dff.simulate(label=label, n_seconds=5, mode='asd', trial_num=42) # Use trial_num for reproducibility

Using the Extended Kalman Filter (EKF)

To track a dynamic system, configure a time-varying signal and use the 'ekf' fitter.

from deepfmkit.physics import LaserConfig, IfoConfig, SimConfig
# Assume 'dff', 'laser', 'ifo' exist from previous examples.

# Add a 5 Hz sinusoidal motion to the measurement arm
ifo.arml_mod_f = 5.0
ifo.arml_mod_amp = 10e-9 # 10 nm amplitude

# Re-create the SimConfig and simulate with dynamic motion
sim_dynamic = SimConfig("dynamic_channel", laser, ifo, f_samp=200e3)
dff.add_sim(sim_dynamic)
dff.simulate(label="dynamic_channel", n_seconds=2, mode='asd')

# Fit using the EKF
dff.fit(main_label="dynamic_channel", method='ekf', fit_label="ekf_fit")

# Plot the EKF's phase estimate and compare to the ground truth
ax = dff.plot(labels=['ekf_fit'], which=['phi'])
raw_data = dff.raws["dynamic_channel"]
fit_data = dff.fits["ekf_fit"]
ax.plot(fit_data.time, raw_data.phi_sim_downsamp, 'r--', label='Ground Truth')
ax.legend()
plt.show()

High-Throughput Experiments

The Experiment framework (experiments.py) is designed for large-scale studies. It automates the process of running thousands of "simulate-and-analyze" trials in parallel.

1. Create an ExperimentFactory in factories.py: This class contains your logic for setting up a single trial. For example:

class VairableAmplitudeOffset(ExperimentFactory):
    def __init__(self, opd_main: float = 0.1):
        self.opd_main = opd_main
        if self.opd_main == 0:
            raise ValueError("opd_main cannot be zero in the factory.")

    def _get_expected_params_keys(self) -> Set[str]:
        """Declares the top-level parameters consumed by this factory's __call__ method.
        """
        return {'m_main', 'nominal_amplitude', 'amplitude_offset'}

    def __call__(self, params: dict) -> dict:
        """Generates the physics configurations for one standard DFMI trial.
        """
        m_main = params['m_main']
        nominal_amplitude = params['nominal_amplitude']
        amplitude_offset = params['amplitude_offset']
        laser_config = physics.LaserConfig(label="ExperimentLaser")
        laser_config.amp = nominal_amplitude + amplitude_offset 
        laser_config.df = (m_main * sc.c) / (2 * np.pi * self.opd_main)
        main_ifo_config = physics.IfoConfig(label="main_ifo")
        main_ifo_config.ref_arml = 0.1
        main_ifo_config.meas_arml = main_ifo_config.ref_arml + self.opd_main
        return {
            'laser_config': laser_config,
            'main_ifo_config': main_ifo_config
        }

2. Define and run the Experiment:

from deepfmkit.experiments import Experiment
from deepfmkit.factories import VairableAmplitudeOffset

import numpy as np
from functools import partial

def amplitude_random_offset_generator(nominal_amplitude: float, relative_noise_std: float) -> float:
    """
    Generates a random amplitude offset whose standard deviation scales
    with the `nominal_amplitude`. This simulates proportional noise.

    Parameters
    ----------
    nominal_amplitude : float
        The nominal signal amplitude, which dictates the scale of the random offset.

    Returns
    -------
    float
        A single random offset value.
    """
    noise_std = nominal_amplitude * relative_noise_std
    return np.random.normal(loc=0.0, scale=noise_std)

my_factory_instance = VairableAmplitudeOffset(
    opd_main=0.1
)

# Create an Experiment object.
experiment = Experiment(description="Example using VairableAmplitudeOffset factory")
experiment.set_config_factory(my_factory_instance)
experiment.n_cycles = 200
experiment.f_samp = 200e3

# Define sweep axis: the nominal signal amplitude.
nominal_amplitudes = np.linspace(0.5, 2.0, 10)
experiment.add_axis('nominal_amplitude', nominal_amplitudes)

# Define static parameters for all trials.
experiment.set_static({
    'm_main': 5.0,                  # Keep modulation depth constant
})

# Add stochastic variable: a random offset to the signal amplitude.
# It 'depends_on' the 'nominal_amplitude' axis.
gen_fun = partial(amplitude_random_offset_generator, relative_noise_std=0.05)

experiment.add_stochastic_variable(
    name='amplitude_offset',
    generator_func=gen_fun,
    depends_on='nominal_amplitude' # This links the noise scale to the nominal amplitude
)

# Specify the number of Monte Carlo trials for each point on the axis.
experiment.n_trials = 100 # 50 trials per nominal_amplitude point for good statistics

# Define the analysis to be performed on each simulated data set.
experiment.add_analysis(
    name='NLS_Fit',
    fitter_method='nls',
    result_cols=['amp', 'm', 'phi', 'psi', 'ssq'], # Parameters I want to collect
    fitter_kwargs={
        'n': 20,       # 20 modulation cycles per fit buffer
        'n_harmonics': 10,   # Use this many harmonics for NLS
    }
)

# --- Run the Experiment ---
print(f"Starting experiment: '{experiment.description}'...")
experiment.results = experiment.run()
print("Experiment completed.")

Citing DeepFMKit

If you use DeepFMKit in your research, please cite this paper:

@article{Dovale2025,
  doi = {10.48550/ARXIV.2507.23183},
  url = {https://arxiv.org/abs/2507.23183},
  author = {Dovale-\'Alvarez,  Miguel},
  keywords = {Optics (physics.optics),  Applied Physics (physics.app-ph),  Instrumentation and Detectors (physics.ins-det),  FOS: Physical sciences,  FOS: Physical sciences},
  title = {Fundamental Limitations of Absolute Ranging via Deep Frequency Modulation Interferometry},
  publisher = {arXiv},
  year = {2025},
  copyright = {arXiv.org perpetual,  non-exclusive license}
}

License

This project is licensed under the BSD 3-Clause License - see the LICENSE file for details.