Nitida.
User Guide

Section 01

Getting started

Nitida is a local-first desktop application: every spectrum, fit and model stays on your machine. Nothing is uploaded, no telemetry is collected.

Installation

First launch & activation

On first start Nitida asks for your license key (received by email after purchase). Paste it in the License key field and press Activate. The key is validated online once, then cached: the app works fully offline and only re-validates every 7 days (with a 3-day grace window if you happen to be offline when a re-check is due). You can manage billing, seats and keys at any time from the customer portal linked on the activation screen and in Settings → License.

The loading screen

Each time Nitida starts, its scientific engine (a local Python backend) boots in the background. The progress bar on the splash screen tracks the real state of the engine, and Enter workspace unlocks the moment it is ready — typically a few seconds.

The pipeline

Nitida is organized as a guided pipeline, visible in the left sidebar. Steps 01–05 follow the natural order of a Raman analysis; the two tools at the bottom are available at any time. A check mark appears next to each completed step.

Keeping Nitida up to date

When a new version is available, a coral banner appears at the bottom of the sidebar. Updating is a deliberate two-step process — the app never restarts by itself:

1. Download update — fetches the new version in the background. You can keep working; nothing is interrupted.
2. Install and restart — appears once the download is complete. Click it when you are ready; the app restarts immediately into the new version.

Section 02

Loading data

Drop one file — Nitida detects the format, the layout and whether it is a single spectrum, a collection or a spatial map. No format pickers, no import wizards.

Supported formats

Loading a file

1. On the Data page, drag your file onto the drop zone (or click it to browse). A preview card shows the file name and size.
2. Press Load file. A progress bar tracks the upload (with speed and ETA), then the parsing phase. Large maps may take a moment — keep the window open.
3. On success, a toast reports the number of spectra and the detected format, and the page shows the dataset overview.

Alternatively, switch to path mode and paste the absolute path of a file already on disk — fastest for multi-gigabyte maps, since nothing is copied.

Data modes

Advanced options

Under Advanced options on the load card:

• Downsample — keep every N-th spectrum of a map. Invaluable for a first fast look at very large maps (e.g. a 121×121 map at downsample 4 loads ~915 spectra instead of 14,641).
• Limit lines — parse only the first N lines of a text file (text formats only).

Section 03

Preprocessing — background & crop

Two light steps that decide the quality of everything downstream: remove what is not signal, keep the window you care about.

Processing order

Every spectrum used anywhere in Nitida goes through the same deterministic pipeline:

raw  →  cosmic-ray removal  →  background subtraction  →  crop

Cosmic-ray removal

The checkbox Remove cosmic rays (Background page, on by default) despikes every spectrum before any other processing. Detection uses a modified Z-score on the discrete Laplacian of the signal:

Points with z > 6 spanning at most 7 samples are treated as spikes and replaced by linear interpolation from clean neighbours (two passes). Real Raman peaks are wider and survive untouched.

Background subtraction

Pick a method on the left, tune its parameters, and watch the Live preview update in real time on the current sample spectrum (Original, Background and Corrected traces). When it looks right, press Apply subtraction — the settings are stored and applied to every spectrum from then on.

Excluded regions

Strong peaks can drag any baseline upward. Add excluded regions over them — the baseline algorithms then ignore those windows. Two ways: type Start/End in cm⁻¹ and click Add region, or click Draw exclusion in the header and tap two points directly on the chart. Regions appear as red bands on the preview and can be deleted individually.

Force non-negative result

Clamps the corrected spectrum at zero: I′ = max(I − b, 0). Useful before non-negative decompositions (NMF/MCR, Section 06).

Cropping

The Crop page restricts every later analysis to a Raman-shift window (fingerprint region, a single band family, …). Set Min/Max (cm⁻¹) numerically, or click Select on chart and tap two points on the spectrum — the kept range is highlighted in green. Press Apply crop; Reset restores the full range at any time.

Section 04

Peak fitting

Decompose a spectrum into physical peaks — positions, amplitudes and widths — with a fitting engine built for stubborn Raman data.

The Peak Fit page fits the current sample spectrum (with your background, crop and cosmic-ray settings already applied). You choose an engine, a peak shape and where the initial peak guesses come from, then press Run fit.

Peak shapes

All three models are normalized to peak height (the amplitude a is the height at the centre, for any mixing), so amplitudes are directly comparable across shapes.

Lorentzian — the Raman default

b = centre, γ = half-width at half-maximum. The natural line shape of most Raman bands.

Gaussian — instrumentally broadened peaks

Pseudo-Voigt — a tunable mix

L and G share one FWHM; the mixing fraction η is fitted (η = 1 pure Lorentzian, η = 0 pure Gaussian). This is a linear pseudo-Voigt, not a true convolution Voigt.

Where the initial peaks come from

Engines

Enhanced (lmfit) — the default. Builds a composite model and tries up to five solvers in turn (Levenberg–Marquardt, trust-region least-squares, Powell, Nelder–Mead, differential evolution), scoring each result and keeping the best. It is robust against poor starting points.
Standard (curve_fit) — SciPy's curve_fit with a three-tier adaptive multi-start; faster, no extra dependency. If a first pass fits poorly it automatically retries with perturbed starts and relaxed bounds.

Rigid constraints

Tick Rigid constraints (Manual mode) to impose hard [min, max] bounds per peak on amplitude, position and width. These are never loosened by the solver — useful to lock a known band. Widths are entered as FWHM and converted internally to the shape's native width.

Reading the results

After a successful fit the right column shows the goodness of fit and a per-peak table.

The publication figure

Every fit produces a customizable figure: edit the title and axis labels (LaTeX-style tokens like $I_D/I_G$ are supported), font sizes, data and fit colours, per-peak colours and fill opacity, and rename each peak with its assignment (e.g. “D band”, “G band”). Export as PNG (2400×1600, publication-ready), SVG (vector) or interactive HTML.

Section 05

Map analysis

Fit every spectrum of a 2D map and turn any fitted quantity into a spatial image. Available for datasets with (x, y) coordinates.

Fitting the whole map

Configure and run one single-spectrum fit first (Section 04) so Nitida knows the peak model to propagate. Then, on the Map Analysis page, click Apply fit to entire map. Every point is fitted with the same background, crop, shape and peak count. A progress panel and a floating widget track the job; you can navigate away and it keeps running.

Stored parameters

For each fitted peak i the map stores, at every pixel:

2D heatmap

Select any parameter and it is rendered as a heatmap over the stage grid. You control the colorscale (Viridis, Plasma, Inferno, Turbo, RdYlGn, Spectral, Blues, Reds, Jet), clamp the colour range with Min/Max value, and choose Interpolation: None, Smooth (Bicubic) (3× cubic upsampling) or Smooth (Gaussian) (σ = 1 blur). Labels accept LaTeX, and Free zoom unlocks the 1:1 aspect ratio. Click any pixel to open the spectrum-and-refit dialog for that point.

Custom parameters

Add custom parameter evaluates a math expression over the parameter maps and stores the result as a new named map — usable everywhere. Type parameter names directly and combine them with + − * / ^ and functions sqrt, log, exp, sin, … and constants pi, e. One-click Quick formulas insert common templates. For example, the analytic peak areas are:

The classic graphitic-carbon ratio I_D/I_G is simply peak1_amp / peak2_amp (or an area ratio using the formulas above).

3D surface

The 3D Surface tab renders the same parameter as a height field, with colorscale, opacity, projected contour lines, camera presets (Publication, Top, Side, Front, Isometric or a free camera), colour-range and axis overrides, and automatic downsampling for very large maps.

Profile cuts & single-point refit

The Profile cut tab extracts a 1D line — a vertical/horizontal cut at a fixed coordinate (bilinear interpolation) or a row/column mean, optionally restricted to a rectangular region of interest and a value window. Clicking any heatmap pixel opens Spectrum at (x, y), where you can view the Original / Processed / Fitted spectrum and refit that single point with adjusted guesses and bounds; the change updates every heatmap instantly, and Revert restores the original values for that pixel.

Export

The Export tab downloads the fitted data — CSV (one parameter grid) or XLSX (every parameter, one sheet each) — and the heatmap as a high-resolution PNG (1920×1440) or interactive HTML.

Section 06

Multivariate analysis & machine learning

Treat a whole map or collection as a dataset: reduce it, cluster it, correlate it, classify it. This is the deepest part of Nitida — this section spells out the mathematics of each method.

Step 1 — Build the feature matrix

Every method operates on a matrix X (rows = spectra, columns = features). Three ways to turn each spectrum into a feature vector:

Step 2 — Preprocess

Transforms applied to X right before every run. The most important, per spectrum i or per column j:

Column scaling: None · Mean-center · Autoscale (z-score) · Pareto · SNV (row) · MSC (row) · Vector norm (row). Row normalization: total area (L1), max, or L2. NaN policy: drop rows, impute mean/median, or zero-fill. Force non-negativity shifts X so it is ≥ 0 — required for NMF/VCA/MCR (Nitida enables it automatically when needed).

Step 3 — Unsupervised methods

PCA — Principal Component Analysis

Finds orthogonal directions of maximum variance via the SVD X = U S V^T: the loadings are the rows of V^T, the scores are T = X·V, and the variance explained by component a is λ_a = s_a²/(n−1). Nitida also reports two diagnostics per spectrum,

T² measures leverage inside the model; Q (residual) measures how much of a spectrum the model fails to explain. Two rank heuristics — Kaiser (λ > 1) and broken-stick — suggest how many components to keep. Views: 2D/3D scores (colored by map Y), scree plot, and loadings.

NMF, VCA, MCR-ALS — parts-based decomposition

All three factor X into non-negative components — abundance maps and endmember spectra — ideal for unmixing overlapping chemical phases.

NMF minimizes the Frobenius (or Kullback–Leibler) reconstruction error. VCA (Vertex Component Analysis) extracts pure endmembers as vertices of the data simplex, then unmixes by pseudoinverse. MCR-ALS alternates non-negative least squares on C and Sᵀ; lack-of-fit LOF% = 100·‖X − CSᵀ‖_F/‖X‖_F. (VCA needs umap-learn-free; MCR needs the optional pymcr package.)

UMAP & t-SNE — nonlinear embeddings

Project high-dimensional spectra to 2–3D for visualization (PCA-preconditioned to 50 components by default). UMAP minimizes the fuzzy-set cross-entropy between the high- and low-dimensional neighbour graphs; t-SNE matches Gaussian affinities in high-D with Student-t affinities in low-D by minimizing

UMAP keeps n_neighbors (local vs global structure) and min_dist (cluster tightness); t-SNE keeps perplexity. UMAP is intentionally non-deterministic for speed; t-SNE is reproducible. UMAP requires the optional umap-learn package.

Step 4 — Clustering

Group spectra by similarity, on either the feature matrix or any reduction/embedding. Every run reports the silhouette, Davies–Bouldin and Calinski–Harabasz scores, and renders a spatial cluster map, a colored reduced-space scatter and per-cluster centroid spectra.

The three validation metrics:

Tick Sweep k to scan a range of k and plot the silhouette (plus inertia for k-means or BIC for GMM) so you can pick the number of clusters. HDBSCAN needs no k.

Step 5 — Correlation

Correlate each feature (or reduction score) against one or more targets (fitted parameters or derived expressions). Choose Pearson (linear), Spearman (rank) or Kendall τ, and optionally apply Benjamini–Hochberg FDR correction to obtain q-values. With no target selected, the full feature×feature matrix is computed. Results: an r ∈ [−1, 1] heatmap (red positive, blue negative) and a table of the strongest correlations with p (and q).

Step 6 — Supervised classification

Label rows into classes (split evenly, threshold a fitted parameter, or manual row ranges), then train one of six classifiers with cross-validation:

• PLS-DA — the chemometric standard. One-hot targets, PLS latent variables maximizing X–Y covariance; predicted class = argmax. Reports VIP scores (importance per Raman shift), an RMSECV-vs-components curve and a latent-variable scores plot.
• SVM, Random Forest, Logistic Regression, k-NN, Gradient Boosting — with per-method parameters (kernel/C, trees/depth, penalty, neighbours, learning rate).

The VIP score, which tells you which bands drive the classification:

where p is the number of features, w̃_a the normalized X-weights, and SS_a the Y-variance explained by component a. VIP > 1 conventionally marks an important variable. Every run reports training and cross-validated accuracy, both confusion matrices, and an optional permutation test giving a p-value for the CV accuracy. Cross-validation is (repeated) stratified k-fold.

Section 07

AI research assistant

A fully local research agent: it searches the literature, compares your fitted peaks with published values and interprets your maps and multivariate models — all on your machine, with no API keys and no data leaving the computer.

How it runs — Ollama

The assistant is powered by a local large language model served by Ollama (a free, open companion app). The first time you open the AI Assistant page, a four-step setup wizard walks you through it:

1. Install Ollama — the wizard links the correct installer for your OS. Install it and return.
2. Start server — one click launches Ollama in the background (and it starts automatically next time).
3. Download the model — a one-time, several-gigabyte download from Ollama's registry, with a progress bar.
4. Load into memory — warms the model in RAM. The chat is ready.

What it can do

The assistant is agentic: it decides which of its tools to call to answer you, showing each step in-line. Its tools include:

Section 08

Settings, license & updates

Save and share reproducible workflows, manage your license, and keep the app current.

Reproducible workflows

The Settings page shows your current configuration — background method, exclusions, cosmic-ray flag, crop, peak shape, peak count, engine and constraints — and lets you save it as a JSON file (Export / Save preset) or load one (Import). Share the JSON with a colleague and they reproduce your exact preprocessing and fitting setup. Imports are previewed and only take effect when you press Apply import.

License

Nitida is licensed per user through Polar. The License panel shows your status, the masked key, customer, and validity, and lets you re-validate, open the customer portal (manage billing and seats), or remove the key from this device (which frees a seat so you can activate elsewhere).

Nitida validates online once, then works fully offline, re-checking roughly every 7 days with a 3-day offline grace window. All timing uses the server clock, so changing your computer's date does not extend a license.

Updates

See Getting started (Section 01): updates are a deliberate two-step Download → Install and restart. The app never restarts on its own, so a running analysis is never lost.

Quitting

Quit Nitida (Settings → About) shuts the scientific engine down cleanly and closes the window.

Appendix

Formats & troubleshooting

A · Formats that need an export step

A few native instrument formats have no open reader anywhere. For these, export from the vendor software once — Nitida reads the exported file natively:

B · Common questions

My file is not recognized

Nitida identifies files by content, so a correct file loads even if renamed. If a vendor file genuinely can't be parsed, the error names the export path to use (see table A). For plain text, make sure columns are numeric and separated by tab, comma or spaces.

A big map is slow to load or fit

Use Downsample (Data → Advanced) for a fast first look, and remember the map fit only needs peaks you actually care about — keeping the peak count low speeds it up dramatically. The fit runs across all CPU cores.

The Multivariate tab is greyed out

It requires a collection or map (≥ 3 spectra) with background and crop applied. Complete those two steps and it unlocks.

A fit won't converge

Check the sample point actually has signal (corners of a map are often off-sample). Subtract the background first, add exclusion regions over strong peaks, and if needed switch to Manual mode to place peaks by hand or add rigid constraints. The Compare lmfit vs curve_fit button helps pick the better engine.

Heatmap looks patchy

Downsampled maps do not fill every pixel. Load the full map for a dense heatmap, or use the Smooth interpolation option on the 2D Heatmap tab.

C · Keyboard & interaction

Nitida — Raman Analysis Suite · Local-first, no telemetry · © 2026 Michele Zanotti. All rights reserved. · Support: [email protected]