LARMOR is a desktop application for fitting solid-state NMR spectra, written as an open successor to dmfit. It keeps dmfit's fitting workflow and adds two things dmfit does not have: an uncertainty on every fitted number, and fits that can be reproduced later from a plain text file.
The physics comes from mrsimulator and the fitting from lmfit; LARMOR adds the interactive fitting UI, file readers, batch workflows and figure export on top. Spectra are read directly from instrument folders (Bruker, Varian/Agilent, legacy dmfit files); the acquired files are never modified. A fit is saved as a JSON "recipe" that references the data by path and hash, together with the acquisition parameters read from the instrument files and the software versions that produced the fit, so it can be re-run, diffed, and shared — typically right next to the spectrum it fits — and the Experimental paragraph and Table S1 of a paper come out of the same file as the fit.
The figure above is rendered by LARMOR itself from the example data in examples/ — a real ¹¹B and ²⁷Al dataset of one glass that Tutorials 1–3, 5 and 6 run against. Tutorials 4 and 7 walk through a composition series and a static wideline dataset that are described but not shipped.
Windows, no Python needed. Download LARMOR-<version>-setup.exe (or the
-win64.zip) from the Releases page and run it: it installs for your own
account (no administrator rights), adds LARMOR to the Start menu and registers
an uninstaller under Settings ▸ Apps. The build is not code-signed, so Windows
SmartScreen says "unknown publisher" — choose More info ▸ Run anyway. The
installed folder carries INSTALL.txt with the first steps and the two
self-checks (LARMOR.exe --selftest, LARMOR.exe --distcheck) to run if
anything misbehaves. A ¹¹B / ²⁷Al example data set ships inside it
(_internal\examples\pCABS2-4), so the tutorials can be followed without data
of your own.
From source (Windows, macOS, Linux). On Windows, double-click install.bat
in this folder: it creates the Python environment and puts a LARMOR.bat
launcher on your Desktop; update.bat updates an existing install. Full
instructions for all platforms, including the usual troubleshooting, are in
INSTALL.md.
Manual install, from inside the repository:
conda env create -f environment.yml # or: pip install -r requirements.txt
conda activate larmor
larmor desktop
Use Python 3.10–3.12 (3.11 recommended). On 3.13 mrsimulator currently has no wheels and the install fails; see INSTALL.md if that happens to you.
Fitting. Drag components directly on the spectrum (position, amplitude, width handles), or edit them in a parameter table with per-parameter fix/bounds/links. Eighteen lineshape models are available, from plain Gauss/Lorentz, Voigt and J-multiplets through Czjzek (also with a free dimensionality d), a correlated shift–C_Q Czjzek, extended Czjzek, dmfit's "Amorphous" Gaussian disorder, second-order quadrupolar, quad+CSA and CSA powder patterns, two-site chemical exchange, to external spectra used as backgrounds. The quadrupolar models take the nucleus and field from the data, so there is nothing special to do for a new isotope. Fit windows can be a union of regions, dmfit-style, and a multi-start "Auto fit" helps with awkward starting points. Spinning-sideband manifolds are detected from the data itself — the repeat at ±νrot, tolerant to a recorded rate off by 2 % — and offered as a linked manifold or a shifted copy of the spectrum in one click.
Uncertainties. Every fit reports errors, three ways: from the covariance matrix, by Monte-Carlo resampling, or from χ² profiling (proper 1σ/2σ confidence intervals). The latter two run on all CPU cores. Parameters that finish pinned at a bound are flagged rather than reported with a meaningless error bar, and the fit warns when two parameters are so correlated the data cannot separate them. Every such flag — unphysical values, degenerate pairs, a structured residual, parameters at bounds, a missing covariance, unsupported populations — is gathered into one always-visible fit-health strip under the spectrum, each chip opening its detail.
Series of spectra. A batch fit applies one shared model to many spectra at once, with amplitudes free per spectrum and any parameter optionally "released" to drift between samples; components can be excluded per sample. A series mode walks a composition or temperature series spectrum by spectrum in the main window: a spectrum without lines takes a copy of its neighbour's model, a fitted one is never overwritten unless you seed it explicitly, a "Keep this fit" lock protects finished members, components are paired by their label, and an auto sweep fits the whole series in forward and backward passes. Multi-field datasets can be fit simultaneously, which is often the only way to separate Cq from δiso. Results export to a long-format CSV with errors and populations, and with each value's status — fixed, linked or finished at a bound — so a table never presents a held value as a fitted one.
2D and MQMAS. Interactive 2D fitting (click to place a site, the fitted model drawn as contours over the data), hypercomplex phasing, shearing, and contour figures with projections and computed CS/QIS reference lines.
Processing. The usual TopSpin/ssNake operations: apodization windows, zero-filling, Fourier transform, phasing (manual or automatic), baseline correction, linear prediction, Hilbert reconstruction, whole-echo processing, region extraction, spectrum algebra and alignment, a time ↔ frequency toggle that returns to the FID for re-apodization without reloading, and real / imaginary / magnitude display channels for phasing. Processing steps are stored in the recipe and replayed whenever the data is reloaded.
Referencing audit. One session folder at a time, every stored SR is compared
with the value indirect referencing from the session's ¹H adamantane spectrum
gives (IUPAC Ξ ratios, the xiref rule); forgotten or stale references are
listed as TopSpin-ready sr values, and old and new values go to a permanent
log so the correction stays reversible (larmor srcheck).
Session inventory. One month folder read into a sample × nucleus grid: the
production EXPNO of every block is pre-picked by the operator's rule (the
highest EXPNO with a pdata/1/1r, demoted when its NS is a small fraction of
the block's maximum or its title says power check / test / failed), sample
names come from the folder convention rather than the title's pulse note,
titles whose rotor ID or sample contradict their folder are flagged, and one
action hands the picks to Batch fit (larmor inventory).
Before and after a series fit. A comparability check reads every member's
acqus, procs and TopSpin command trail and flags what differs from the series
majority (line broadening, TDeff, recycle delay…), with a one-click reprocess
of all members from their fids through one common chain (larmor compare).
A series table names, groups and orders the spectra and joins composition
columns from a CSV, so the series plot can run against a real x axis with
replicate averaging and a fitted line. Lines can be tagged with a structural
family (BO₃/BO₄, Al[4]/Al[5]/Al[6], Qⁿ…) to report summed populations and named
ratios such as N₄ with errors propagated on a stated basis. Every table marks
held (†), at-bound (‡) and linked (§) parameters, the fit-health strip judges
the recycle delay and flip angle against a sibling T₁ measurement, and the MAS
rate is taken by majority from the acqus, the title and the booking sidecar.
Relaxation and dipolar experiments. T1/T2 extraction from arrayed experiments (saturation/inversion recovery, CPMG, T1ρ), including per-site decomposition; REDOR curves to dipolar couplings and distances; QCPMG echo trains to spikelet or sum-echo spectra with a guided, step-by-step processing dialog.
Figures. A plotting studio produces publication figures (overlays, stacked plots, deconvolution grids from a batch fit, species distributions, relaxation series) from a JSON spec that can be saved and re-rendered identically. Journal style presets, PNG/SVG/PDF output.
Other. DFT tensor import from CASTEP/Quantum ESPRESSO .magres files
(spin-aware seeding, symmetry-equivalent atoms merged per crystallographic
site, a multi-reference shielding-to-shift calibration line with covariance),
a command palette (Ctrl+Shift+P) that reaches any menu entry by fuzzy search,
an optional bridge to SIMPSON for exact recoupling simulations, and a larmor
CLI (info, import, fit, batchfit, seqfit, multifit, satrec,
redor, magres, shiftcal) for scripted use — the whole thing is an
ordinary Python package underneath.
dmfit is the reference point throughout, and LARMOR reads its .fxml/.fxmla
fit files directly. The practical differences: uncertainties are computed for
everything rather than being a separate tool, fits are plain JSON instead of an
opaque format, batch and series fitting are built in, everything is scriptable
from Python, and it runs on any platform. If you have years of dmfit fits, they
import — that was one of the design goals, and 20 fits from a published
¹¹B study import with no warnings and match the paper's parameters to its own
rounding (validation report, §5.7).
A fitting program should not be trusted on faith. The physics is checked
against analytic theory, direct ensemble simulations, and real published
data: see docs/validation.md. The figures in that report
are generated by a script in the repo, not drawn. The test suite (about 1800
tests, pytest tests/) guards the physics constants and conventions and, on
machines that have them, fits real Bruker datasets; the published-paper
reproductions above are documented in the validation report. The Windows build
checks itself: LARMOR.exe --distcheck imports every module the program can
reach, opens every bundled resource, runs every engine, export format, the
process pool and the command line, and with --gui every window and menu
action, inside the installed copy.
The desktop app is the primary interface and where all development happens.
About 70k lines of Python, of which the fitting core is Qt-free and usable as a
library. Known gaps are tracked in docs/roadmap.md. The
Windows builds come from packaging/build.bat (PyInstaller + Inno Setup) and
are not code-signed; there is no CI yet. A FastAPI web variant exists in the
tree (larmor app) but is unmaintained and is not part of the Windows build.
Anyone modifying the code should start with docs/development-notes.md: the conventions the numbers depend on, the validation anchors a change must not break, and an index of the bugs already fixed here. docs/architecture.md maps the modules.
Tutorials live in docs/tutorials/:
- A first fit — ²⁷Al Czjzek
- Constraints — fix, bound, link
- Plotting studio figures
- Batch fitting a composition series
- MQMAS — 2D processing and fitting
- Error analysis — covariance, Monte-Carlo, χ² profile
- Static wideline — ⁸¹Br WURST-CPMG
They are also reachable in the app under Help ▸ Tutorials. Raw instrument
data is never committed to this repository and never written to — data/
holds only notes about where data lives.
MIT. If you use LARMOR in published work, a citation of the repository is appreciated, alongside dmfit (Massiot et al., Magn. Reson. Chem. 2002) whose design this follows and mrsimulator which does the quantum mechanics.
LARMOR itself is MIT. The graphical interface uses Qt 6 through PySide6 and
shiboken6 (© The Qt Company), used under the GNU Lesser General Public License
v3 (LGPL-3.0). In the Windows build the Qt libraries ship unmodified as
separate DLLs under _internal\PySide6 and can be replaced with another build
of the same Qt major version. The full third-party list with every licence
text the wheels carry is THIRD_PARTY_LICENSES.txt next to LARMOR.exe
(packaging/THIRD_PARTY_LICENSES.txt here), and the LGPL-3.0 / GPL-3.0 texts
are in LICENSES\; PySide6 licensing: https://doc.qt.io/qtforpython-6/licenses.html,
source: https://code.qt.io/cgit/pyside/pyside-setup.git.
LARMOR works offline and sends nothing anywhere. The one exception is Help ▸ More…, a just-for-fun card that fetches a picture from a public image service when you open it.
LARMOR is developed by Sami Soudani in the Nuclear, Optical, Magnetic, and Electronic Materials Laboratory (NOME) of Prof. John McCloy, School of Mechanical and Materials Engineering, Washington State University, with support from the U.S. Department of Energy (DOE).