Research-stage implementation plan

Obsidian Implementation Details

Purpose

Project Scope

Reduce the labeling needed to detect and localize semiconductor defects while preserving wafer- or lot-level split integrity.

  • Build reproducible wafer and SEM loaders
  • Benchmark low-label training
  • Surface model uncertainty and novel defects

Stack

Tools & Technologies

PyTorchVision models
MONAISegmentation pipeline
OpenCVImage processing
MLflowExperiment evidence

Architecture

System Design

Self-supervised pretraining feeds classification and segmentation heads. A k-center selector prioritizes informative samples for annotation and records every dataset split.

  • Dataset and provenance registry
  • Self-supervised encoder
  • Active annotation selector
  • Defect and uncertainty reports

Validation

Evidence Plan

Results remain research claims until these checks produce reviewable artifacts.

  • Wafer/lot-isolated data splits
  • Full-label and low-label baselines
  • Rare-class precision-recall analysis
  • Out-of-distribution rejection tests