Limitations =========== VANE states its assumptions explicitly so results are not over-interpreted. Input format and rotor ---------------------- * **Modern ``.lin`` only.** The legacy OpenFAST linearization format is rejected. * **Three-bladed rotors.** The MBC3 transform and the blade collective/regressive/progressive taxonomy assume a three-bladed rotor. Two- or many-bladed rotors are out of scope for the multi-blade coordinate path; purely non-rotating models (no blade triplets) are averaged without the transform. Averaged-model assumptions -------------------------- * The averaged non-rotating model approximates a linear time-periodic system by its azimuth average. It is exact for an isotropic rotor and degrades with anisotropy; the **azimuth-spread** uncertainty quantifies how much. A coherent sweep (distinct azimuths, constant rotor speed) is required and enforced. * Adequate azimuth coverage is the user's responsibility; VANE rejects degenerate sweeps but does not synthesise missing azimuths. Mode tracking ------------- The cross-operating-point tracker extracts each Campbell line as the **globally optimal single-path** continuity through the sweep, then removes it and repeats. This is deterministic and avoids greedy adjacent-point errors, but it is a *sequential* extraction, not a simultaneous globally optimal **multi-track** assignment: in dense spectra or at near-degenerate crossings, an earlier high-scoring path can still claim a node a later path would have preferred. Tracks flagged ``is_ambiguous`` mark exactly these hotspots, and the per-track ``confidence`` (the weakest-link correlation) exposes how trustworthy a line is. Treat ambiguous, low-confidence tracks as candidates for manual review. Resonance detection ------------------- Resonance detection reports **line crossings** between a mode curve and an ``nP`` excitation line, with the crossing rotor speed, frequency, and damping interpolated linearly, a **damping-aware severity** (a lightly damped crossing is more dangerous than a well-damped one), and optional **operating-window filtering**. The excitation family defaults to 1P/3P/6P/9P and can be set to any positive-integer harmonics. It does not yet provide closest-approach margin bands or uncertainty-propagated crossings. AI components are experimental ------------------------------ The Gaussian-process classifier and the ensemble fusion are **experimental**. The default model is bootstrapped on *synthetic* labelled modes, not a measured, multi-turbine corpus, so its calibrated probabilities are indicative rather than validated. Constructing the classifier — directly, or through the trainer, which builds one — emits an :class:`~vane.ai.ExperimentalWarning`; the ensemble fuses the predictions of an already-constructed classifier. Do not rely on the learned label alone for engineering decisions; the rule-based physical label and the deterministic taxonomy are the primary identification path. The PyTorch-based sequence/graph trackers are not part of the supported core. See the model card (:doc:`../reference/model_card`) for the default model's training data, intended use, and metrics. Scope ----- VANE post-processes linearization output; it does not run OpenFAST, set up models, or compute spatial (geometry-resolved) mode shapes. Visualisation renders modal participation and complex phasors, not deflected turbine geometry.