Should we consider Anaklasis as another profile calculator?

[rozyczko]

anaklasis as a third EasyReflectometry calculator backend

Question: Would it be useful to add anaklasis as a calculator backend in EasyReflectometry, alongside refnx and refl1d? Does it provide anything extra?

Recommendation: No. As a reflectivity kernel anaklasis adds nothing that refnx and refl1d do not already provide, and it brings real packaging and licence costs. Its genuinely novel parts live in the fitting layer, which EasyReflectometry would not use because it has its own parameter, constraint and fitting stack through easyscience.

Evaluation date: 2026-10-09.

---

1. What anaklasis is

• Open-source Python 3 package with Fortran 90 extensions for specular neutron and X-ray reflectometry.
• Author: Alexandros Koutsioubas (Koutsioumpas), Jülich Centre for Neutron Science.
• Published in J. Appl. Cryst. 54, 1857–1866 (2021), DOI 10.1107/S1600576721009262.
• Three public functions in the ref module: ref.calculate, ref.compare, ref.fit.
• Models are defined as nested Python lists: a system is a list of patches (models) with coverage fractions; each model is a list of layers [Re SLD, Im SLD, thickness, roughness, solvent fraction, name].
• Core kernel: anaklasis/fortran_ref_functions.f90 (~10 KB). Everything else: anaklasis/ref.py (~232 KB monolith).

2. The reflectivity kernel compared with refnx and refl1d

Aspect	anaklasis	refnx / refl1d
Algorithm	Abelès matrix method	Abelès / Parratt, equivalent physics
Roughness	Névot–Croce damping factor	Névot–Croce, plus micro-slicing of arbitrary profiles
Absorption	Complex SLD supported	Supported
Solvent / hydration	Per-layer volume-fraction blending	Supported (refnx Slab.vfsolv, refl1d mixtures)
Resolution	Fixed 17-point midpoint rule over roughly ±3.3σ; constant dQ/Q or pointwise from 4th column	Gaussian smearing with configurable quadrature; constant or pointwise
Vectorisation	None, loops per Q point in Fortran	Vectorised C / Numba kernels
Fallback engines	Numba (30–40 % slower), pure Python (20–30x slower)	n/a
Magnetism	None in the kernel. Non-spin-flip PNR is emulated by giving each spin curve a different SLD via a per-curve parameter. Spin-flip unsupported; paper points users to GenX	refl1d: full four-channel polarised calculation, already exposed through WrapperBase.calculate_polarized
Mixed-area models	Incoherent coverage-weighted sum of patches	refnx MixedReflectModel; refl1d can do the same; trivially done at the EasyReflectometry model level

Conclusion: at the kernel level anaklasis is a strict subset of what the current backends offer, and it would most likely be slower.

3. What is actually new in anaklasis

These features:

• Layer parameters as SymPy expressions in global parameters, per-curve multi-parameters and the layer index n.
• Inequality constraints written as strings, e.g. '2*p1/p0>1'.
• Co-refinement of several curves with shared global parameters and per-curve multi-parameters.
• Uncertainty estimation by three routes: numdifftools Hessian diagonal, bootstrap (1000 refits), emcee MCMC.
• Global optimisation by SciPy differential evolution with polish=True.
• List-based model definition aimed at non-programmers.

EasyReflectometry already provides parameters, constraints and fitting through easyscience. Wrapping anaklasis would mean wrapping only the Fortran routine, which duplicates what refnx and refl1d already give.

4. Practical blockers

• Licence: GPL-3.0. Adding it as an optional dependency of a BSD-licensed library is at best awkward.
• Build: requires gfortran. Only Windows wheel is for Python 3.9. The PyPI page could not be fetched during this evaluation. The CI matrix would inherit the Fortran packaging burden.
• Maintenance: single maintainer, 3 stars, 2 forks, 1 open issue. Substantive commits in September 2025, then a single "Add files via upload" on 2026-09-10.
• Validation: an ORSO validation check plus regression against the package's own 1.5.2 output.
• API shape: the public functions read data files and write plots and folders. A backend would have to call internal Reflectivity() helpers rather than a clean function API.
• Latent correctness issues observed in the Fortran:
  • The reflection coefficient r is declared COMPLEX*8 (single precision) while the rest of the code is double precision.
  • The hydration flags handle only two of the fronting/backing combinations. When both media are flagged as solvent, the relevant indices are never set, leaving the result undefined. The same gap exists in f_RealRef, f_profilecalc and f_solventcalc.

5. Ideas worth borrowing (implement on our side, do not link anaklasis)

• Expression-valued layer parameters that reference other parameters or the layer index.
• Bootstrap error estimation as an alternative to MCMC.
• Coverage-weighted mixed-area ("patch") models, if not already available at the EasyReflectometry model level.

Whether these are real gaps depends on what easyscience already exposes.

6. Decision

Do not add anaklasis as a calculator backend. Keep refnx and refl1d. Treat anaklasis only as a reference for fitting-layer features.

---

Sources

• anaklasis repository
• anaklasis Fortran kernel
• anaklasis paper, J. Appl. Cryst. 54 (2021), PMC8662969
• IUCr article page
• EasyReflectometryLib wrapper_base.py