Skip to content

XLuminA, a highly-efficient, auto-differentiating discovery framework for super-resolution microscopy.

License

Notifications You must be signed in to change notification settings

artificial-scientist-lab/XLuminA

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

✨ XLuminA ✨

Nature Communications Downloads License: MIT Tests Python Versions PyPI version GitHub stars

XLuminA, a highly-efficient, auto-differentiating discovery framework for super-resolution microscopy

📄 Read our paper here:
Automated discovery of experimental designs in super-resolution microscopy with XLuminA
Carla Rodríguez, Sören Arlt, Leonhard Möckl and Mario Krenn

📰 Read the press release:
10,000 times faster than traditional methods: new computational framework automatically discovers experimental designs in microscopy

📚 Related works featuring XLuminA:

NeurIPS Machine Learning and the Physical Sciences Workshop - Poster

ICML AI4Science Workshop - Oral contribution

NeurIPS AI4Science Workshop - Oral contribution

💻 Installation:

Using PyPI:

Create a new conda environment and install xluminafrom PyPI. We recommend using python=3.11:

conda create -n xlumina_env python=3.11

conda activate xlumina_env

pip install xlumina

It should be installed in about 10 seconds. The package automatically installs:

  1. JAX (CPU only) and jaxlib (the version of JAX used in this project is v0.4.33),

  2. Optax (the version of Optax used in this project is v0.2.3),

  3. SciPy (the version of SciPy used in this project is v1.14.1).

Clone repository:

git clone https://github.com/artificial-scientist-lab/XLuminA.git

GPU compatibility:

To install JAX with NVIDIA GPU support (Note: wheels only available on linux), use CUDA 12 installation:

pip install --upgrade "jax[cuda12]"

XLuminA has been tested on the following operating systems:

Linux Enterprise Server 15 SP4 15.4,

and it has been successfully installed in Windows 10 (64-bit) and MacOS Monterey 12.6.2

👾 Features:

XLuminA allows for the simulation, in a (very) fast and efficient way, of classical light propagation through optics hardware configurations,and enables the optimization and automated discovery of new setup designs.

workflow

The simulator contains many features:

✦ Light sources (of any wavelength and power) using both scalar or vectorial optical fields.

✦ Phase masks (e.g., spatial light modulators (SLMs), polarizers and general variable retarders (LCDs)).

✦ Amplitude masks (e.g., spatial light modulators (SLMs) and pre-defined circles, triangles and squares).

✦ Beam splitters, fluorescence model for STED, and more!

✦ The light propagation methods available in XLuminA are:

📝 Examples of usage:

Examples of some experiments that can be reproduced with XLuminA are:

The code for each of these optical setups is provided in the Jupyter notebook of examples.ipynb.

➤ A step-by-step guide on how to add noise to the optical elements can be found in noisy_4f_system.ipynb.

🚀 Testing XLuminA's efficiency:

We evaluated our framework by conducting several tests - see Figure 1. The experiments were run on an Intel CPU Xeon Gold 6130 and Nvidia GPU Quadro RTX 6000.

(1) Average execution time (in seconds) over 100 runs, within a computational window size of $2048\times 2048$, for scalar and vectorial field propagation using Rayleigh-Sommerfeld (RS, VRS) and Chirped z-transform (CZT, VCZT) in Diffractio and XLuminA. Times for XLuminA reflect the run with pre-compiled jitted functions. The Python files corresponding to light propagation algorithms testing are scalar_diffractio.py and vectorial_diffractio.py for Diffractio, and scalar_xlumina.py and vectorial_xlumina.py for XLuminA.

propagation

(2) We compare the gradient evaluation time of numerical methods (using Diffractio's optical simulator and SciPy's BFGS optimizer) vs autodiff (analytical) differentiation (using XLuminA's optical simulator with JAX's ADAM optimizer) across various resolutions:

performance

(3) We compare the convergence time of numerical methods (using Diffractio's optical simulator and SciPy's BFGS optimizer) vs autodiff (analytical) differentiation (using XLuminA's optical simulator with JAX's ADAM optimizer) across various resolutions:

performance

➤ The Jupyter notebook used for running these simulations is provided as test_diffractio_vs_xlumina.ipynb.

➤ The Python files corresponding to numerical/autodiff evaluations are numerical_methods_evaluation_diffractio.py, and autodiff_evaluation_xlumina.py

If you want to run the comparison test of the propagation functions, you need to install Diffractio - The version of Diffractio used in this project is v0.1.1.

🤖🔎 Discovery of new optical setups:

With XLuminA we were able to rediscover three foundational optics experiments. In particular, we discover new, superior topologies together with their parameter settings using purely continuous optimization.

➤ Optical telescope (or 4f-correlator),

➤ Polarization-based beam shaping as used in STED (stimulated emission depletion) microscopy,

➤ The sharp focus of a radially polarized light beam.

The Python files used for the discovery of these optical setups, as detailed in our paper, are organized in pairs of optical_table and optimizer as follows:

Experiment name 🔬 Optical table 🤖 Optimizer 📄 File for data
Optical telescope four_f_optical_table.py four_f_optimizer.py Generate_synthetic_data.py
Pure topological discovery: large-scale sharp focus (Dorn, Quabis and Leuchs, 2004) hybrid_with_fixed_PM.py hybrid_optimizer.py N/A
Pure topological discovery: STED microscopy hybrid_with_fixed_PM.py hybrid_optimizer.py N/A
6x6 grid: pure topological discovery six_times_six_ansatz_with_fixed_PM.py hybrid_optimizer.py N/A
Large-scale polarization-based STED hybrid_sted_optical_table.py hybrid_optimizer.py N/A
Large-scale sharp focus (Dorn, Quabis and Leuchs, 2004) hybrid_sharp_optical_table.py hybrid_optimizer.py N/A

🦾🤖 Robustness and parallelized optimization of multiple optical tables with our noise-aware scheme:

✦ Importantly, to ensure simulations which approximate real-world experimental conditions we have included imperfections, misalignment, and noise sources in all optical components (during post-processing and/or during optimization). All the results presented in the paper are computed considering a wide variety of experimental errors.

➤ A step-by-step guide on how to setup the optimization using this scheme can be found in noisy_optimization.ipynb.

➤ A step-by-step guide on how to add noise to the optical elements can be found in noisy_4f_system.ipynb.

noise_aware

✦ The optimization procedure is as follows: for each optimization step, we execute $N$ parallel optical tables using vmap. Then, we sample random noise and apply it to all available physical variables across each of the $N$ optical tables. The random noise is uniformly distributed and includes:

  • Phase values for spatial light modulators (SLMs) and wave plates (WP) in the range of $\pm$ (0.01 to 0.1) radians, covering all qualities available in current experimental devices.
  • Misalignment ranging from $\pm$ (0.01 to 0.1) millimeters, covering both expert-level precision ($\pm$ 0.01 mm) and beginner-level accuracy ($\pm$ 0.1 mm).
  • 1% imperfection on the transmissivity/reflectivity of beam splitters (BS), which is a realistic approach given the high quality of the currently available hardware.

We then simulate the optical setup for each of the $N$ tables simultaneously, incorporating the sampled noise. The loss function is computed independently for each of the setups. Afterwards, we calculate the mean loss value across all optical tables, which provides an average performance metric that accounts for the introduced experimental variability (noise). The gradients are computed based on this mean loss value and so the update of the system parameters'.

Importantly, before applying the updated parameters and proceeding to the next iteration, we resample new random noise for each optical table. This ensures that each optimization step encounters different noise values, further enhancing the robustness of the solution. This procedure is repeated iteratively until convergence.

👀 Overview:

In this section we list the available functions in different files and a brief description:

  1. In wave_optics.py: module for scalar optical fields.

    Class Functions Description
    ScalarLight Class for scalar optical fields defined in the XY plane: complex amplitude $U(r) = A(r)*e^{-ikz}$.
    .draw Plots intensity and phase.
    .apply_circular_mask Apply a circular mask of variable radius.
    .apply_triangular_mask Apply a triangular mask of variable size.
    .apply_rectangular_mask Apply a rectangular mask of variable size.
    .apply_annular_aperture Apply annular aperture of variable size.
    .RS_propagation Rayleigh-Sommerfeld diffraction integral in z-direction (z>0 and z<0).
    .get_RS_minimum_z Given a quality factor, determines the minimum (trustworthy) distance for RS_propagation.
    .CZT Chirped z-transform - efficient diffraction using the Bluestein method.
    LightSource Class for scalar optical fields defined in the XY plane - defines light source beams.
    .gaussian_beam Gaussian beam.
    .plane_wave Plane wave.
  2. In vectorized_optics.py: module for vectorized optical fields.

    Class Functions Description
    VectorizedLight Class for vectorized optical fields defined in the XY plane: $\vec{E} = (E_x, E_y, E_z)$
    .draw Plots intensity, phase and amplitude.
    .draw_intensity_profile Plots intensity profile.
    .VRS_propagation Vectorial Rayleigh-Sommerfeld diffraction integral in z-direction (z>0 and z<0).
    .get_VRS_minimum_z Given a quality factor, determines the minimum (trustworthy) distance for VRS_propagation.
    .VCZT Vectorized Chirped z-transform - efficient diffraction using the Bluestein method.
    PolarizedLightSource Class for polarized optical fields defined in the XY plane - defines light source beams.
    .gaussian_beam Gaussian beam.
    .plane_wave Plane wave.
  3. In optical_elements.py: shelf with all the optical elements available.

    Function Description
    Scalar light devices -
    phase_scalar_SLM Phase mask for the spatial light modulator available for scalar fields.
    SLM Spatial light modulator: applies a phase mask to incident scalar field.
    Jones matrices -
    jones_LP Jones matrix of a linear polarizer
    jones_general_retarder Jones matrix of a general retarder.
    jones_sSLM Jones matrix of the superSLM.
    jones_sSLM_with_amplitude Jones matrix of the superSLM that modulates phase & amplitude.
    jones_LCD Jones matrix of liquid crystal display (LCD).
    Polarization-based devices -
    sSLM super-Spatial Light Modulator: adds phase mask (pixel-wise) to $E_x$ and $E_y$ independently.
    sSLM_with_amplitude super-Spatial Light Modulator: adds phase mask and amplitude mask (pixel-wise) to $E_x$ and $E_y$ independently.
    LCD Liquid crystal device: builds any linear wave-plate.
    linear_polarizer Linear polarizer.
    BS_symmetric Symmetric beam splitter.
    BS_symmetric_SI Symmetric beam splitter with single input.
    BS Single-side coated dielectric beam splitter.
    high_NA_objective_lens High NA objective lens (only for VectorizedLight).
    VCZT_objective_lens Propagation through high NA objective lens (only for VectorizedLight).
    General elements -
    lens Transparent lens of variable size and focal length.
    cylindrical_lens Transparent plano-convex cylindrical lens of variable focal length.
    axicon_lens Axicon lens function that produces a Bessel beam.
    circular_mask Circular mask of variable size.
    triangular_mask Triangular mask of variable size and orientation.
    rectangular_mask Rectangular mask of variable size and orientation.
    annular_aperture Annular aperture of variable size.
    forked_grating Forked grating of variable size, orientation, and topological charge.
    👷‍♀️ Pre-built optical setups -
    bb_amplitude_and_phase_mod Basic building unit. Consists of a sSLM (amp & phase modulation), and LCD linked via VRS_propagation.
    building_block Basic building unit. Consists of a sSLM, and LCD linked via VRS_propagation.
    fluorescence Fluorescence model.
    vectorized_fluorophores Vectorized version of fluorescence: Allows to compute effective intensity across an array of detectors.
    robust_discovery 3x3 setup for hybrid (topology + optical settings) discovery with single wavelength. Longitudinal intensity (Iz) is measured across all detectors. Includes noise for robustness.
    hybrid_setup_fixed_slms_fluorophores 3x3 optical table with SLMs randomly positioned displaying fixed phase masks; to be used for pure topological discovery; contains the fluorescence model in all detectors. (Fig. 4a of our paper)
    hybrid_setup_fixed_slms 3x3 optical table with SLMs randomly positioned displaying fixed phase masks; to be used for pure topological discovery. (Fig. 4b of our paper)
    hybrid_setup_fluorophores 3x3 optical table to be used for hybrid (topological + optical parameter) discovery; contains the fluorescence model in all detectors . (Fig. 5a and Fig. 6 of our paper)
    hybrid_setup_sharp_focus 3x3 optical table to be used for hybrid (topological + optical parameter) discovery. (Fig. 5b of our paper)
    six_times_six_ansatz 6x6 optical table to be used for pure topological discovery. (Extended Data Fig. 6 of our paper)
    🫨 Add noise to the optical elements -
    shake_setup Literally shakes the setup: creates noise and misalignment for the optical elements. Accepts noise settings (dictionary) as argument. Can't be used with jit across parallel optical tables.
    shake_setup_jit Same as shake_setup. Doesn't accept noise settings as argument. Intended to be pasted in the optimizer file to enable jit compilation across parallel optical tables.
  4. In toolbox.py: file with useful functions.

    Function Description
    Basic operations -
    space Builds the space where light is placed.
    wrap_phase Wraps any phase mask into $[-\pi, \pi]$ range.
    is_conserving_energy Computes the total intensity from the light source and compares is with the propagated light - Ref.
    softmin Differentiable version for min() function.
    delta_kronecker Kronecker delta.
    build_LCD_cell Builds the cell for LCD.
    draw_sSLM Plots the two phase masks of sSLM.
    draw_sSLM_amplitude Plots the two amplitude masks of sSLM.
    moving_avg Compute the moving average of a dataset.
    image_to_binary_mask Converts image (png, jpeg) to binary mask.
    rotate_mask Rotates the (X, Y) frame w.r.t. given point.
    gaussian Defines a 1D Gaussian distribution.
    gaussian_2d Defines a 2D Gaussian distribution.
    lorentzian Defines a 1D Lorentzian distribution.
    lorentzian_2d Defines a 2D Lorentzian distribution.
    fwhm_1d_fit Computes FWHM in 1D using fit for gaussian or lorentzian.
    fwhm_2d_fit Computes FWHM in 2D using fit for gaussian_2d or lorentzian_2d.
    profile Determines the profile of a given input without using interpolation.
    spot_size Computes the spot size as $\pi (\text{FWHM}_x \cdot \text{FWHM}_y) /\lambda^2$.
    compute_fwhm Computes FWHM in 1D or 2D using fit: gaussian, gaussian_2d, lorentzian, lorentzian_2.
    📑 Data loader -
    MultiHDF5DataLoader Data loader class for 4f system training.
  5. In loss_functions.py: file with loss functions.

    Function Description
    small_area_hybrid Small area loss function valid for hybrid (topology + optical parameters) optimization
    vMSE_Intensity Parallel computation of Mean Squared Error (Intensity) for a given electric field component $E_x$, $E_y$ or $E_z$.
    MSE_Intensity Mean Squared Error (Intensity) for a given electric field component $E_x$, $E_y$ or $E_z$.
    vMSE_Phase Parallel computation of Mean Squared Error (Phase) for a given electric field component $E_x$, $E_y$ or $E_z$.
    MSE_Phase Mean Squared Error (Phase) for a given electric field component $E_x$, $E_y$ or $E_z$.
    vMSE_Amplitude Parallel computation of Mean Squared Error (Amplitude) for a given electric field component $E_x$, $E_y$ or $E_z$.
    MSE_Amplitude Mean Squared Error (Amplitude) for a given electric field component $E_x$, $E_y$ or $E_z$.
    mean_batch_MSE_Intensity Batch-based MSE_Intensity.

⚠️ Considerations when using XLuminA:

  1. By default, JAX uses float32 precision. If necessary, enable jax.config.update("jax_enable_x64", True) at the beginning of the file.

  2. Basic units are microns (um) and radians. Other units (centimeters, millimeters, nanometers, and degrees) are available at __init.py__.

  3. IMPORTANT - RAYLEIGH-SOMMERFELD PROPAGATION: FFT-based diffraction calculation algorithms can be innacurate depending on the computational window size (sampling).
    Before propagating light, one should check which is the minimum distance available for the simulation to be accurate.
    You can use the following functions:

    get_RS_minimum_z, for ScalarLight class, and get_VRS_minimum_z, for VectorizedLight class.

💻 Development:

Some functionalities of XLuminA’s optics simulator (e.g., optical propagation algorithms, lens or amplitude masks) are inspired in an open-source NumPy-based Python module for diffraction and interferometry simulation, Diffractio. We have rewritten and modified these approaches to combine them with JAX just-in-time (jit) functionality. We labeled these functions as such in the docstrings. On top of that, we developed completely new functions (e.g., sSLMs, beam splitters, LCDs or propagation through high NA objective lens with CZT methods, to name a few) which significantly expand the software capabilities.

📝 How to cite XLuminA:

If you use this software, please cite as:

Rodríguez, C., Arlt, S., Möckl, L. and Krenn, M. Automated discovery of experimental designs in super-resolution microscopy with XLuminA. Nat Commun 15, 10658 (2024). https://doi.org/10.1038/s41467-024-54696-y

BibTeX format:

@article{NatCommun.15.10658,
  title={Automated discovery of experimental designs in super-resolution microscopy with {XLuminA}},
  author={Rodríguez, Carla and Arlt, Sören and Möckl, Leonhard and Krenn, Mario},
  journal={Nature Communications},
  volume={15},
  pages={10658},
  year={2024},
  publisher={Nature Publishing Group},
  doi={10.1038/s41467-024-54696-y}
}