SANS Model Fitter

A flexible, model-agnostic Python template for fitting Small-Angle Neutron Scattering (SANS) data using the SasModels library.

Features

• Model-Agnostic Design: Works with any model from the SasModels library (cylinder, sphere, core_shell, etc.)
• Multiple Fitting Engines: Supports both BUMPS (default) and LMFit optimization engines
• Flexible Data Loading: Reads CSV, XML, and HDF5 formats via sasdata
• Example Data Built In: A curated set of ready-to-fit datasets plus a simulator that generates data — examples.load_fitter('silica_spheres') or examples.simulate('sphere', radius=50)
• Q-Range Restriction: Fit only a chosen [qmin, qmax] window (e.g. trim beam-stop or background-dominated points)
• Dataset Arithmetic: Add, subtract, multiply, and divide datasets (or scale by constants) with propagated uncertainties via data_ops — e.g. background subtraction and transmission correction before fitting
• User-Friendly Parameter Management: Easy-to-use interface for setting parameter values, bounds, and fitting flags
• Interactive Visualization: Automatic plotting of data, fitted model, and residuals with Plotly
• Bayesian Analysis: Posterior sampling with BUMPS DREAM (MCMC) plus corner, marginal, predictive-band, correlation, and trace plots
• P(r) Inversion: Model-free pair distance distribution analysis (indirect Fourier transform) via pr_inversion — D_max exploration, automatic regularization/term selection, and Rg/I(0)/positivity diagnostics
• Result Export: Save fitted parameters and curves to CSV files

Installation

Option 1: Using pip (recommended for users)

# Clone the repository
git clone https://github.com/ai4se1dk/SANS-fitter.git
cd SANS-fitter

# Install the package
pip install -e .

# Or install with development dependencies
pip install -e ".[dev]"

Option 2: Using Pixi (recommended for development)

# Clone the repository
git clone https://github.com/ai4se1dk/SANS-fitter.git
cd SANS-fitter

# Install dependencies with Pixi
pixi install

# Run tests
pixi run test

# Run demo notebook
pixi run run-demo

Quick Start

No data of your own yet? Start from a bundled example:

from sans_fitter import examples

examples.describe()                              # see what's available
fitter = examples.load_fitter('silica_spheres')  # data + model + parameters
result = fitter.fit()
fitter.plot_results()

Or simulate data, so you can check the fit recovers it:

data = examples.simulate('sphere', radius=50, noise=0.02, seed=0)
data.truth['radius']    # 50.0

See the Example Data guide for the full collection.

With your own data:

from sans_fitter import SANSFitter

# Create fitter instance
fitter = SANSFitter()

# Load your data
fitter.load_data('my_sans_data.csv')

# Set the model (any model from SasModels!)
fitter.set_model('cylinder')

# Optionally restrict the Q range used for fitting
fitter.set_q_range(qmin=0.01, qmax=0.3)

# View initial parameter values
fitter.get_params()

# Configure parameters for fitting
fitter.set_param('radius', value=20, min=1, max=100, vary=True)
fitter.set_param('length', value=400, min=10, max=1000, vary=True)
fitter.set_param('scale', value=0.1, min=0, max=1, vary=True)
fitter.set_param('background', value=0.01, min=0, max=1, vary=True)

# View current parameters
fitter.get_params()

# Perform the fit (using BUMPS by default)
result = fitter.fit(engine='bumps', method='amoeba')

# Visualize results
fitter.plot_results(show_residuals=True)

# Save results
fitter.save_results('fit_results.csv')

Switching Models

The fitter is completely model-agnostic. Simply load a different model:

# Try with a sphere model instead
fitter.set_model('sphere')
fitter.get_params()  # See different parameters!

fitter.set_param('radius', value=25, min=5, max=100, vary=True)
result = fitter.fit()

Switching Fitting Engines

# Use BUMPS (default)
result = fitter.fit(engine='bumps', method='amoeba')

# Or use LMFit
result = fitter.fit(engine='lmfit', method='leastsq')

Working with Structure Factors

Combine any SasModels form factor with an interaction model to capture correlated systems.

fitter.set_model('sphere')

# Apply a structure factor (creates sphere@hardsphere product model)
fitter.set_structure_factor('hardsphere', radius_effective_mode='link_radius')

# Inspect linked parameters and run the fit as usual
fitter.get_params()
result = fitter.fit()

# Remove the structure factor to go back to the pure form factor
fitter.remove_structure_factor()

• Supported structure factors: hardsphere, hayter_msa, squarewell, stickyhardsphere.
• Radius handling: use radius_effective_mode='link_radius' to keep radius_effective equal to the form-factor radius, or leave the default unconstrained to fit it independently.
• State helpers: get_structure_factor() returns the active structure factor so notebooks/scripts can branch as needed.

Combining Models

Describe a dataset with several models at once — for example a low-Q diffuse feature plus a high-Q peak — fitted simultaneously against one dataset:

fitter = SANSFitter()
fitter.load_data('data.csv')

# Combine models; parameters get friendly per-model prefixes
fitter.set_models('dab', 'peak_lorentz')
fitter.set_param('dab_cor_length', value=50, min=1, max=500, vary=True)
fitter.set_param('peak_lorentz_peak_pos', value=0.1, min=0.01, max=0.5, vary=True)
fitter.set_param('background', value=0.001, min=0, max=0.1, vary=True)

result = fitter.fit(engine='bumps')

# See which feature each model accounts for
fitter.plot_results(show_components=True)

The combined intensity is

$I(q) = \text{scale} \cdot \sum_i \text{scale}_i \cdot I_i(q) + \text{background}$

so the global scale and background are shared by all components natively, while each component has its own <model>_scale.

Custom names (monikers) — for long model names, duplicates, or physics labels:

fitter.set_models(small='sphere', large='sphere', shared=['sld', 'sld_solvent'])
fitter.set_param('small_radius', value=20, vary=True)
fitter.set_param('large_radius', value=200, vary=True)

Sharing parameters — any name in shared=[...] that exists in ≥ 2 components becomes a single unprefixed parameter driving all of them (sld above). Polydispersity configuration stays per-component under the prefixed names.

Component curves — after fitting a '+' mixture, plot_results(show_components=True) overlays one dashed curve per component, each drawn as scale · part_scale · I_part(q) (background shown implicitly in the total).

Advanced: the raw sasmodels expression syntax is also available and keeps sasmodels' native A_/B_ names: fitter.set_model('dab+peak_lorentz'). For after-the-fact or asymmetric sharing, use equality links: fitter.link_params('large_sld', to='small_sld') and fitter.unlink_params('large_sld'). Composite models and parameter links currently require the bumps engine.

Bayesian Analysis

Sample the full posterior distribution of the varying parameters with the DREAM MCMC sampler (built into BUMPS — no extra dependencies):

fitter.set_model('sphere')
fitter.set_param('radius', value=50, min=10, max=200, vary=True)
fitter.set_param('scale', value=0.1, min=0.01, max=1.0, vary=True)

# Run the Bayesian fit (prints point estimates + credible intervals + diagnostics)
result = fitter.fit_bayesian(samples=10000, burn=200)

# Corner plot of the posterior
fitter.plot_posterior_pairs()

# More displays
fitter.plot_param_distribution('radius')      # marginal posterior
fitter.plot_posterior_predictive()            # 95% credible band over the data
fitter.plot_param_correlations()              # correlation heatmap
fitter.plot_trace()                           # MCMC chain traces

# Raw chain access / export
posterior = fitter.get_posterior()
posterior.save_posterior_csv('posterior_chain.csv')

See the User Guide for details.

Available Methods

BUMPS methods:

• 'amoeba' - Nelder-Mead simplex (default, robust)
• 'lm' - Levenberg-Marquardt
• 'newton' - Newton's method
• 'de' - Differential evolution

LMFit methods:

• 'leastsq' - Levenberg-Marquardt (default)
• 'least_squares' - Trust Region Reflective
• 'differential_evolution' - Global optimizer
• 'powell', 'nelder', etc.

Demo Notebooks

• notebooks/sans_fitter_demo.ipynb — comprehensive demonstration of the fitting workflow with examples.
• notebooks/bayesian_sampling.ipynb — Bayesian posterior sampling API (fit_bayesian()) and the associated posterior plots.
• notebooks/pr_inversion_demo.ipynb — model-free P(r) inversion: D_max exploration, automatic inversion, and diagnostics.

Design Philosophy

This implementation follows a template pattern where:

1. The core fitting logic is abstracted into a reusable class
2. Models are loaded dynamically from SasModels - no hardcoded model assumptions
3. Parameters are discovered automatically from the model definition
4. Multiple optimization engines are supported through a unified interface
5. The user maintains full control over parameter initialization and bounds

Implementation Details

Engine Adapters

The fitter implements adapter patterns for both BUMPS and LMFit:

• BUMPS: Uses native sasmodels.bumps_model integration
• LMFit: Uses sasmodels.direct_model.DirectModel with a custom residual function

Parameter Management

Parameters are stored internally with:

• value: Current/initial value
• min, max: Bounds
• vary: Fitting flag
• description: From model metadata

This allows the fitter to work with any model without prior knowledge of its parameters.

License

BSD 3-Clause License. See LICENSE for the full text.

References

• SasModels: https://github.com/SasView/sasmodels
• BUMPS: https://github.com/bumps/bumps
• LMFit: https://lmfit.github.io/lmfit-py/
• Plotly: https://plotly.com/python/