xpcsjax

JAX-native NLSQ fitting for X-ray Photon Correlation Spectroscopy (XPCS).

xpcsjax consolidates the homodyne and heterodyne analysis pipelines — both now deprecated in its favor — into one package with a shared engine and config-driven physics-model dispatch. It implements the transport-coefficient framework of He et al. PNAS 2024 and He et al. PNAS 2025 for characterizing nonequilibrium dynamics in flowing soft-matter systems.

---

Install

To get every feature and avoid missing-dependency issues, install with the all extra (pulls every optional dependency — GUI, fast-viz, plus the dev/docs/packaging tooling).

With uv (recommended):

uv pip install "xpcsjax[all]"

In a mamba/conda environment — or any other virtualenv (venv, virtualenv, pyenv) — use pip:

pip install "xpcsjax[all]"

For a minimal core install (NLSQ fitting only, no GUI/fast-viz), drop the extra: uv pip install xpcsjax or pip install xpcsjax. Individual extras are also available: xpcsjax[gui] (desktop workbench) and xpcsjax[viz-fast] (datashader-accelerated plots).

Python 3.12+ required, CPU-only in v0.1 (GPU support is v0.2+). Runtime dependencies are managed via pyproject.toml and mirror what the source homodyne package pins (versions of jax, nlsq, evosax, h5py, interpax, jaxopt, psutil, scikit-learn, tqdm, etc.).

From source (development): clone the repo and use an editable install — uv sync && uv pip install -e ".[dev]". See docs/source/development/contributing.rst.

---

Quickstart

from xpcsjax import load_xpcs_data, fit_nlsq

data   = load_xpcs_data("config.yaml")
result = fit_nlsq(data, "config.yaml")
print(result.parameters)

The YAML config's analysis_mode field selects the physics model and parameter set:

analysis_mode	Lineage	Model	Physics params
static_isotropic	homodyne	Equilibrium diffusion, angle-collapsed	3
static_anisotropic	homodyne	Same physics; angle-resolved data prep	3
laminar_flow	homodyne	Diffusion + sinc-shear	7
two_component (or heterodyne)	heterodyne	Two-component: reference + sample + velocity + mixing	14

fit_nlsq returns a single OptimizationResult for every mode (result.parameters is a NumPy array). Heterodyne (two_component) fits all phi angles jointly and packs the multi-angle result into that same object. Two per-angle scaling parameters (contrast, offset) are appended automatically for every azimuthal angle in all modes.

Data flow

YAML config --> XPCSDataLoader(HDF5) --> HomodyneModel / HeterodyneModel --> NLSQ engine --> Results (JSON + NPZ)

save_results (xpcsjax/service/persist.py) writes nlsq_result.json (fitted parameters, uncertainties, χ², diagnostics) and nlsq_result.npz (correlation and residual arrays) to the configured output directory.

---

Running xpcsjax

Three front-ends drive the same YAML config and NLSQ engine — pick whichever fits your workflow.

Command line. The xpcsjax console script (alias xj) runs a flag-driven fit:

xpcsjax --config analysis.yaml                          # run an NLSQ fit
xpcsjax --config analysis.yaml --output ./results       # override the output dir
xpcsjax --config analysis.yaml --multistart --multistart-n 16
xpcsjax --config analysis.yaml --plot-experimental-data # plot only, skip the fit

Exit codes: 0 converged · 2 ran but did not converge (outputs still written) · 1 error · 130 interrupted. xjexp / xjsim are experimental-/simulated-data plotting shortcuts. See the CLI commands table and the CLI guide for the full reference.

Interactive workbench (GUI). Needs the gui extra (pip install "xpcsjax[gui]"). Launch the PySide6 analysis workbench with:

xpcsjax-gui      # or: xj-gui

Load a config, run fits, and browse per-angle results and residual maps. The fit runs in a JAX-free worker process so the UI stays responsive; xpcsjax-gui -platform offscreen does a headless smoke run.

Shell completion. After install, wire up tab-completion (bash/zsh; fish is a non-fatal no-op) and optional XLA flags into your environment:

xpcsjax-post-install   # interactive: completion + XLA_FLAGS into the venv activate script
xpcsjax-cleanup        # remove what post-install added

---

Physics models

xpcsjax fits two-time intensity correlation functions $c_2(\vec{q}, t_1, t_2)$. All time integrals are evaluated numerically via cumulative trapezoid on the discrete time grid — no analytical antiderivatives — so the general power-law forms stay correct.

Homodyne (static_*, laminar_flow)

Single-component scattering where correlation decay encodes diffusion and shear. The laminar-flow two-time kernel (xpcsjax core/physics_nlsq.py; derived in docs/source/theory/homodyne_model.rst) is

$$c_2(\vec{q}, t_1, t_2) = c_{\text{offset}}(\phi) + \beta(\phi) \exp\left(-q^2\int_{t_1}^{t_2} J(t') dt'\right) \mathrm{sinc}^2\left(\frac{q h \cos(\phi - \phi_0) \Gamma(t_1, t_2)}{2\pi}\right)$$

with $\mathrm{sinc}(x) = \sin(\pi x)/(\pi x)$, accumulated strain $\Gamma(t_1, t_2) = \int_{t_1}^{t_2}\dot{\gamma}(t) dt$, rheometer gap $h$ (instrument geometry, not fitted), flow angle $\phi_0$, and per-angle scaling $\beta(\phi)$ / $c_{\text{offset}}(\phi)$ (the contrast / offset parameters). The static modes drop the shear term ($\mathrm{sinc}^2 \to 1$). Transport and shear follow power-law forms:

$$J(t) = D_0 t^{\alpha} + D_{\text{offset}} \qquad \dot{\gamma}(t) = \dot{\gamma}0 t^{\beta} + \dot{\gamma}{\text{offset}}$$

Registry parameter names (xpcsjax/config/parameter_registry.py) and their defaults:

Group	Parameter	Description	Default	Units
Diffusion	D0	Diffusion prefactor	1e3	Ų/s
	alpha	Transport exponent (0 = Wiener, 1 = ballistic)	0.5	—
	D_offset	Transport rate offset	10	Ų/s
Shear (laminar_flow only)	gamma_dot_t0	Shear-rate prefactor	0.01	s⁻¹
	beta	Shear-rate exponent (0 = constant shear)	0.5	—
	gamma_dot_t_offset	Shear-rate offset	0.0	s⁻¹
Flow angle	phi0	Flow angle offset relative to q-vector	0.0	degrees
Per-angle scaling	contrast	Optical (speckle) contrast	0.5	—
	offset	Baseline offset	1.0	—

The static modes use the 3 diffusion parameters; laminar_flow adds the 3 shear parameters and the flow angle (7 physics parameters total).

Heterodyne (two_component)

Two-component scattering (PNAS 2025 SI Eqs. S-77–S-98): light from a moving sample interferes with a static reference, and the cross-term oscillates at a frequency set by the sample velocity. The two-time correlation (Eq. S-95) is

$$c_2(\vec{q}, t_1, t_2) = 1 + \frac{\beta}{f^2}\left[C_{\text{ref}} + C_{\text{sample}} + C_{\text{cross}}\right]$$

$$C_{\text{ref}} = [x_r(t_1)x_r(t_2)]^2 \exp\left(-q^2\int_{t_1}^{t_2} J_r dt'\right) \qquad C_{\text{sample}} = [x_s(t_1)x_s(t_2)]^2 \exp\left(-q^2\int_{t_1}^{t_2} J_s dt'\right)$$

$$C_{\text{cross}} = 2 x_r(t_1)x_r(t_2)x_s(t_1)x_s(t_2) \exp\left(-\tfrac{1}{2}q^2\int_{t_1}^{t_2}[J_s + J_r] dt'\right)\cos\left[q\cos(\varphi)\int_{t_1}^{t_2}\mathbb{E}[v] dt'\right]$$

where $x_s(t)$ is the sample fraction, $x_r = 1 - x_s$ the reference fraction, $\varphi$ the angle between velocity and $\vec{q}$, and $f^2 = [x_s(t_1)^2 + x_r(t_1)^2][x_s(t_2)^2 + x_r(t_2)^2]$ normalizes so that $c_2(t, t) = 1 + \beta$ on the diagonal. The fit wraps the correlation with per-angle scaling, $c_2^{\text{model}} = \text{offset} + \text{contrast}\times(C_{\text{ref}} + C_{\text{sample}} + C_{\text{cross}})/f^2$.

Each transport coefficient and the velocity follow power laws, and the sample fraction is time-dependent — 14 physics parameters in five groups:

Registry parameter names (xpcsjax/config/parameter_registry.py, canonical order) and their defaults:

Group	Parameters	Rate function	Defaults	Units
Reference transport (3)	D0_ref, alpha_ref, D_offset_ref	$J_r(t) = D_{0,r} t^{\alpha_r} + D_{\text{offset},r}$	1e4, 0, 0	Ų/s, —, Ų/s
Sample transport (3)	D0_sample, alpha_sample, D_offset_sample	$J_s(t) = D_{0,s} t^{\alpha_s} + D_{\text{offset},s}$	1e4, 0, 0	Ų/s, —, Ų/s
Velocity (3)	v0, v_beta, v_offset	$v(t) = v_0 t^{\beta} + v_{\text{offset}}$	1e3, 1, 0	Å/s, —, Å/s
Sample fraction (4)	f0, f1, f2, f3	$f_s(t) = f_0 \exp\big(f_1(t - f_2)\big) + f_3$	0.5, 0, 0, 0	—, —, —, —
Flow angle (1)	phi0_het	—	0	degrees

The velocity exponent is v_beta and the flow angle phi0_het in the registry — deliberately distinct from homodyne's beta / phi0 to avoid name collisions. As with homodyne, 2 per-angle scaling parameters (contrast, offset) are tracked per azimuthal angle but live outside the 14-element physics vector. Per-angle scaling modes constant / averaged / individual (resolved from auto) reach full parity with the source heterodyne package's fit_nlsq_multi_phi.

---

Citation

xpcsjax implements the transport-coefficient framework introduced in:

@article{He2024,
  author  = {He, Hongrui and Liang, Heyi and Chu, Miaoqi and Jiang, Zhang and
             de Pablo, Juan J and Tirrell, Matthew V and Narayanan, Suresh
             and Chen, Wei},
  title   = {Transport coefficient approach for characterizing nonequilibrium
             dynamics in soft matter},
  journal = {Proceedings of the National Academy of Sciences},
  volume  = {121},
  number  = {31},
  year    = {2024},
  doi     = {10.1073/pnas.2401162121}
}

@article{He2025,
  author  = {He, Hongrui and Liang, Heyi and Chu, Miaoqi and Jiang, Zhang and
             de Pablo, Juan J and Tirrell, Matthew V and Narayanan, Suresh
             and Chen, Wei},
  title   = {Bridging microscopic dynamics and rheology in the yielding
             of charged colloidal suspensions},
  journal = {Proceedings of the National Academy of Sciences},
  volume  = {122},
  number  = {42},
  year    = {2025},
  doi     = {10.1073/pnas.2514216122}
}

To cite the software itself:

@software{xpcsjax,
  author      = {Chen, Wei},
  title       = {xpcsjax: JAX-native NLSQ fitting for X-ray Photon Correlation Spectroscopy},
  year        = {2026},
  version     = {0.1.1},
  institution = {Argonne National Laboratory},
  url         = {https://github.com/imewei/xpcsjax}
}

License

MIT. See LICENSE.

Authors

• Wei Chen (weichen@anl.gov) — Argonne National Laboratory