How andreasleitherer/ai4stem works
AI-STEM characterizes atomic-resolution STEM images. Its primary workflow loads an image, divides it into local windows, computes FFT-HAADF descriptors, and applies a pretrained CNN to produce crystal-class predictions and uncertainty; a separate analysis workflow embeds neural-network representations for unsupervised exploration. The repository also includes CNN training and lattice reconstruction utilities. Edges are limited to relationships visible in the sampled source or SOURCE INDEX; reconstruction is shown as a capability, but no caller was sampled.
Main entry points: utils_data.py Data loaders · utils_prediction.py Image prediction · ai4stem_prediction.py Prediction example · utils_unsupervised.py Representation analysis
Components 16
Supervised analysis
- Prediction example
scripts/ai4stem_prediction.py - Image prediction
ai4stem/utils/utils_prediction.py - Local windows
ai4stem/utils/utils_prediction.py - FFT descriptors
ai4stem/utils/utils_fft.py
Models and data
- CNN inference
ai4stem/utils/utils_nn.py - CNN training
ai4stem/utils/utils_nn.py - Data loaders
ai4stem/utils/utils_data.py - Example image
ai4stem/data/experimental_images/Fe_bcc_100.npy - Pretrained CNN
ai4stem/data/pretrained_models/optimal_model.h5 - Class definitions
ai4stem/data/class_definitions
Unsupervised analysis
- Embedding example
scripts/ai4stem_unsupervised_learning.py - Representation analysis
ai4stem/utils/utils_unsupervised.py - UMAP embedding
ai4stem/utils/utils_unsupervised.py
Lattice analysis
- Atomic-column reconstruction
ai4stem/utils/utils_reconstruction.py - Reference lattices
ai4stem/data/reference_images
Other components
- Microscopy analyst
Connections 23
- Microscopy analyst → Prediction example: runs
scripts/ai4stem_prediction.py - Prediction example → Data loaders: loads image
scripts/ai4stem_prediction.py - Prediction example → Data loaders: loads model
scripts/ai4stem_prediction.py - Prediction example → Image prediction: calls
scripts/ai4stem_prediction.py - Prediction example → Data loaders: loads labels
scripts/ai4stem_prediction.py - Data loaders → Example image: reads
ai4stem/utils/utils_data.py - Data loaders → Pretrained CNN: loads
ai4stem/utils/utils_data.py - Data loaders → Class definitions: reads
ai4stem/utils/utils_data.py - Image prediction → Local windows: extracts windows
ai4stem/utils/utils_prediction.py - Image prediction → FFT descriptors: computes descriptors
ai4stem/utils/utils_prediction.py - Image prediction → Data loaders: loads model
ai4stem/utils/utils_prediction.py - Image prediction → CNN inference: predicts with uncertainty
ai4stem/utils/utils_prediction.py - CNN inference → Pretrained CNN: loads model
- Microscopy analyst → Embedding example: runs
scripts/ai4stem_unsupervised_learning.py - Embedding example → Data loaders: loads image
scripts/ai4stem_unsupervised_learning.py - Embedding example → Representation analysis: calls
scripts/ai4stem_unsupervised_learning.py - Representation analysis → Image prediction: gets descriptors
ai4stem/utils/utils_unsupervised.py - Representation analysis → Data loaders: loads model
ai4stem/utils/utils_unsupervised.py - Representation analysis → CNN inference: gets representations
ai4stem/utils/utils_unsupervised.py - Representation analysis → UMAP embedding: fits and transforms
ai4stem/utils/utils_unsupervised.py - UMAP embedding → Microscopy analyst: returns embedding
- CNN training → Pretrained CNN: saves model
- Atomic-column reconstruction → Reference lattices: uses reference lattice