nmr-quant

Least-squares fitting of NMR peak lineshapes (single Voigt, Voigt with satellites, and annotation-driven multiplets) to 1D spectra, with standard fit metrics and quantities (peak areas, FWHM, optional normalised concentration).

The package is general-purpose: it takes plain NumPy arrays in and returns a plain pandas table out. It knows nothing about any specific data source, LIMS, or plotting. Reference deconvolution and other row processing live in python-nmr-spectra-processing; visualisation lives in python-nmr-quant-dashboards.

Status

Feature-complete. Public API:

• fit_spectrum(x, y, ...) — fit one spectrum -> FitRecord. Single Voigt, Voigt with satellites, or an annotation-driven multiplet, with an optional simultaneous polynomial baseline and LS or Huber loss.
• fit_dataset(x, spectra, ...) — fit a 2-D batch -> QuantResults. Warm-starts on the mean spectrum, then fits each row with optional per-spectrum FWHM/centre priors; norm yields a quant column and ids are carried through.
• QuantResults — the fit table + model metadata, with save/load (parquet + json, no dataset needed) and model_curve for reconstruction.
• fit_metrics — standard goodness-of-fit and residual diagnostics (r2, noise-adjusted r2_adj, srr, sse, rmse, mae, residual_std, noise_std, aic, bic, Durbin-Watson dw, chi2_red, snr).
• build_quality_map / apply_quality_map / load_quality_map / kde_score — the generic quality-map rules engine (thresholds are caller config).
• nmr_quant._shapes — Voigt / multiplet lineshape primitives (bridge module; these become a re-export from python-nmr-spectra-processing.lineshapes once that module is published).

34 tests, ~94% coverage.

Install (dev)

uv sync
uv run pytest

Design

• Arrays in, table out. No coupling to any loader or to matplotlib.
• Inputs coerced to C-contiguous float64; batches are one 2-D (n_spectra, n_points) array.
• Metrics are computed for every fit, not left to the caller.

License

MIT.