Exploratory Nodes

Exploratory nodes reveal structure before supervised modeling.

PCA, Decomposition, and Curve Resolution

Node	Use When	Inputs	Outputs	Key Configuration
PCA (model.pca)	Explore variance, scores, loadings, outliers, and compressed features.	default: Array2D	scores; loadings; explained_variance; model	n_components; standardized; scaled. Uses Sherpa's native exact full-SVD authority. n_components="mle" uses Minka's automatic dimensionality method and requires n_samples >= n_features. Keep scaling choices consistent with your spectroscopy convention.
PCA Transform (model.pca_transform)	Project new spectra into an already fitted PCA model.	X_new: SpectralDataset; model: DecompositionResult	scores: ScoreMatrix	no parameters. Use the same preprocessing as the fitted PCA model.
NMF (model.nmf)	Resolve non-negative concentration-like and spectrum-like factors.	default: SpectralDataset	concentrations; spectra; reconstruction_error; model	n_components; solver; max_iter; tol. Input must be non-negative; use Clip Floor or baseline correction first if needed.
FastICA (model.ica)	Blind source separation when independent latent sources are plausible.	default: SpectralDataset	sources; mixing_matrix; components; model	n_components; algorithm; fun; max_iter; tol.
PARAFAC (model.parafac)	Resolve components in sample-first or spatial-first multiway data while keeping every physical mode distinct.	default: SpectralDataset with rank 3–6 and explicit mode roles	mode-1 sample_scores; component_weights; relative_reconstruction_error; data-free model	n_components; max_iter; tol; ridge. Uses deterministic native CP-ALS. A three-mode image cube may carry an exact binary mask over its first two spatial modes; every ALS update and the reconstruction error honor that mask. No hidden unfolding is performed.
MCR-ALS (model.mcr_als)	Resolve mixture concentration profiles and pure spectra with constraints.	default: SpectralDataset	C; St; residuals; ground_truth_comparison; model	n_components; non_negative_C; non_negative_St; max_iter; tol; normSpec; validation indices. Requires SpectroChemPy.
EFA (model.efa)	Estimate evolving rank/component count in ordered mixture or process data.	default: SpectralDataset	forward_eigenvalues; backward_eigenvalues; model	n_components. Requires SpectroChemPy.
SIMPLISMA (model.simplisma)	Estimate pure variables/components by purity maximization.	default: SpectralDataset	concentrations; spectra; purity_values; model	n_components; tol; noise. Requires SpectroChemPy.

SpectroChemPy's MCR-ALS and baseline documentation are useful background for constrained curve-resolution thinking: https://www.spectrochempy.fr/0.7.0/userguide/analysis/mcr_als.html and https://www.spectrochempy.fr/0.8.3/userguide/processing/baseline.html.

PARAFAC is deliberately an expert-only canvas node in 0.6.0: no analysis starter references it. A supported pixels-by-features Eigenvector image DSO exposes a second, explicitly named image-cube scientific result. Select that result and add model.parafac manually. The exact IASIM16 Test 1 source has a private, data-free qualification receipt, but that exploratory decomposition does not establish melamine detection or challenge performance.

Clustering

Node	Use When	Inputs	Outputs	Key Configuration
HCA (model.hca)	Build hierarchical clusters and dendrograms from spectra or scores.	default: Array2D	labels; cluster_summary; linkage_matrix; dendrogram_data; embedding; model	n_clusters; linkage; metric. Ward linkage expects Euclidean distance.
K-Means (model.kmeans)	Partition samples into a chosen number of compact clusters.	default: Array2D	labels; centroids; cluster_summary; embedding; model	n_clusters; n_init; max_iter; random_state.
DBSCAN (model.dbscan)	Find density-based clusters and noise/outlier samples.	default: Array2D	labels; cluster_summary; embedding; model	eps; min_samples; metric. Tune eps carefully after scaling.

Peak and Library Nodes

Node	Use When	Inputs	Outputs	Key Configuration
Peak Finding (analysis.peak_finding)	Detect candidate spectral peaks for interpretation, masking, or library workflows.	default: SpectralDataset	peaks (consensus PeakTable); plots; per_spectrum (measurement matrix)	height; threshold; distance; prominence; width; consensus_tolerance. Detection controls mirror SciPy find_peaks; consensus tolerance groups detections by their axis position.
Compare vs. Library (analysis.compare_library)	Rank a sample against selected reference spectra using HQI and cosine similarity.	sample: SpectralDataset; library: SpectralDataset	ranking dictionary with scores and diagnostics	top_n; library_filter; hqi_mode; diagnostic bands and overlap thresholds.

PeakTable includes the numbered consensus group, position statistics, detection fraction, labels, and complete detection membership. The former Salient Features port is consolidated into this table; existing editable-workflow connections migrate automatically, while historical run records remain unchanged. Re-run the workflow after upgrading to generate the merged table.

For each spectrum, the measurement matrix contains four columns per consensus group: peak_position_N, magnitude_at_peak_N, group_position_N, and magnitude_at_group_position_N. An undetected peak leaves the first two values missing; the nearest measured coordinate to the consensus guide and its magnitude remain available. Connect either table to Data Table and use Quick Plot's Column selector to plot any scalar column against its one-based row index. A downstream Plot node offers the same column choice. Missing values remain gaps, text values use a categorical axis, and nested membership lists should be inspected in Data View.

SciPy documents the find_peaks controls for height, threshold, distance, prominence, and width here: https://docs.scipy.org/doc/scipy-1.16.0/reference/generated/scipy.signal.find_peaks.html.

Peak-finding input and replay semantics

Numeric editors preserve typed decimals and scientific notation. They do not round values to the suggested step or clamp them to a bound. Invalid entries remain visible and block execution. Blank optional inputs are saved as null, which becomes Python None.

Field	Effective meaning
Height	Blank: height=None, no height filter. Otherwise a minimum in response units.
Threshold	Blank: threshold=None, no neighbor-threshold filter. Otherwise in response units.
Distance	New-node default: 10, shown as an entered value. Clearing it passes distance=None. Legacy 0 also explicitly disables it; other values must be at least 1 sample point.
Prominence	Blank: prominence=None, no prominence filter. Otherwise in response units.
Width	Blank: width=None. Legacy 0 also explicitly disables the filter. Otherwise minimum half-prominence width in sample points.
Consensus tolerance	Required; default 0, shown in the editor. Zero groups only identical positions. Positive values set the maximum consensus-bin span in feature-axis units. This is a Sherpa postprocessing setting, not a SciPy argument.

New results retain the exact scipy.signal.find_peaks keyword arguments, including wlen=None, rel_height=0.5, and plateau_size=None, together with the SciPy version. Executed peak-finding inputs shows the retained call in both the Inspector and expanded view. It describes the result's execution, independently of subsequent edits to the settings. Historical results without that record are explicitly marked unavailable; current settings are not used to reconstruct it.

Reported widths are measured separately using scipy.signal.peak_widths(spectrum, indices, rel_height=0.5, prominence_data=None, wlen=None) and mapped to the feature axis. To reproduce a run, use the same upstream spectral matrix, axis and sample ordering as well as the recorded calls and version. Consensus markers and full-height dotted vertical guides are visible in new plots. Rerun Peak Finding to obtain these guides and the new matrix; historical run outputs are unchanged.

The Per-spectrum Peak Matrix output connects directly to Data Table. It retains one row per spectrum and four columns for each consensus group x (numbered from 1 in ascending consensus-position order): peak_position_x, magnitude_at_peak_x, group_position_x, and magnitude_at_group_position_x. The first pair is NaN when no peak was detected for that spectrum. The second pair always uses the measured feature coordinate nearest the consensus guide and its actual magnitude, without interpolation. Equidistant coordinates use the lower coordinate. If a spectrum has multiple detections within one group, the nearest detection to the guide supplies the first pair (ties use the lower position); all detections remain in Peak List. Column-specific units and the exact guide positions are retained as metadata. The original Peak List remains the consensus/membership summary; connect the new matrix output for a rectangular per-spectrum table and CSV export. Data Table's Quick Plot offers each numeric column separately, with null detections left as gaps.

Practical Use

Use exploratory nodes to understand variation, outliers, clusters, pure-component estimates, and candidate spectral features before locking in a calibration or classification model. Prefer scores plots for sample structure, loadings or coefficients for variable interpretation, and residual/limit plots for model adequacy.

LLM-assisted peak interpretation is available through the privacy-gated Sherpa Advisor conversation after deterministic Peak Finding. Advisor responses are human-reviewed interpretation, not canonical DAG execution or independently reproducible scientific evidence.