11.27. AI-Assisted Processing

pycsamt.ai.processing applies trained models to eight data-preparation tasks. Four already have a deterministic, rule-based counterpart in EM Tools Guide – quality control, noise suppression, anomaly detection, and dimensionality classification – and each tool here is a complement, not a replacement, for it. Gap filling and uncertainty calibration have no rule-based equivalent at all: nothing else in pyCSAMT reconstructs a genuinely missing impedance cell or re-estimates an inversion error floor from signal quality. Distortion triage is a third kind: it routes each station toward whichever existing correction tool (pycsamt.emtools.ss or pycsamt.emtools.gb) its regime actually calls for, rather than complementing either directly. Raw Time-Series Denoising (MMF-SVM-K-SVD) is the odd one out in a different way: every other tool here works on the frequency-domain impedance tensor, while it works one stage further upstream, on the raw field time series before it is even Fourier-transformed. Start with AI Processing Overview for how the tools relate to each other, the shared estimator interface, and when to reach for a learned model over a rule-based one.

• 11.27.1. AI Processing Overview
  • 11.27.1.1. Shared building blocks
• 11.27.2. Machine-Learning Quality Scoring
  • 11.27.2.1. Fitting and scoring
  • 11.27.2.2. Reading the score chart
  • 11.27.2.3. Sites in, masked sites out
  • 11.27.2.4. Parameters
• 11.27.3. Learned Spectral Denoising
  • 11.27.3.1. Training data and validation protocol
  • 11.27.3.2. Sites in, corrected sites out
  • 11.27.3.3. Parameters and limitations
• 11.27.4. Raw Time-Series Denoising (MMF-SVM-K-SVD)
  • 11.27.4.1. Separating signal from noise with a support vector machine
  • 11.27.4.2. Denoising only the flagged windows with K-SVD
  • 11.27.4.3. Validating against a known-clean synthetic record
  • 11.27.4.4. Applying the pipeline to a real field recording
  • 11.27.4.5. Parameters and limitations
• 11.27.5. Profile-Level Anomaly Detection
  • 11.27.5.1. Fitting and scoring
  • 11.27.5.2. Sites in, filtered sites out
  • 11.27.5.3. Parameters
• 11.27.6. Learned Dimensionality Classification
  • 11.27.6.1. Self-training from the rule-based labels
  • 11.27.6.2. Supervising the strike head with a classical estimate
  • 11.27.6.3. Parameters
• 11.27.7. Learned Gap Filling
  • 11.27.7.1. A survey with real gaps
  • 11.27.7.2. Quantitative validation on held-out cells
  • 11.27.7.3. Sites in, filled sites out
  • 11.27.7.4. Parameters and limitations
• 11.27.8. Learned Uncertainty Calibration
  • 11.27.8.1. Fitting and a held-out quantitative check
  • 11.27.8.2. Working directly with site collections
  • 11.27.8.3. Parameters and limitations
• 11.27.9. Learned Distortion Triage
  • 11.27.9.1. Building the feature table
  • 11.27.9.2. Self-training and a real limitation
  • 11.27.9.3. Routing to the real correction tools
  • 11.27.9.4. Parameters and limitations