This code applies local interpretable model-agnostic explanation (LIME) to a set of topologically inverse designed transverse mode multipelxers as presented Advancing Photonic Chip Design with Interpretable Machine Learning by Pira et. al (https://doi.org/10.48550/arXiv.2505.09266).
By applying the LIME algorithm to a small CNN trained to differentiate between two classes of high and low bandwidth multiplexers, we are able to use the LIME explanations to highlight particular regions within each class. Based on the identified regions we then modify the initial conditions of the inverse design optimisation in order to further improve the multiplexer bandwidths.
Inverse designed mode multiplexers used in the paper are contained within the multiplexer_combined_bandwidths/ folder. The multiplexers are categorised into two folders two class folders: combined_high_bw and combined_low_bw based on their simulated tranmission bandwidth of both modes.
Bitmaps are read as grayscale images with pixel values normalised to [0, 1], and resized 256×256 in size.
We train a simple sequential CNN using Adam optimiser with sparse categorical cross-entropy. The CNN is trained for 20 epochs using an 80/20 tran/test split.
CNN Structure
The final trained CNN model used in the paper is saved in cnn_v4_combined_model_v1.h5 and can be loaded in the jupyter notebook.
Model Accuracy
Model Loss
The LIME section:
- Loads the saved model (
cnn_v4_combined_model_v1.h5). - Defines a
predict_fnwrapper that converts LIME’s RGB input back into grayscale tensors for the CNN. - Uses
lime_image.LimeImageExplainer()with SLIC superpixels to perturb images and estimate feature importance. - Writes out:
- Original images (
Original_*.png) - LIME overlays (
LIME_Explanation_*.png) - Raw binary/importance masks (
Mask_*.png)
- Original images (
The helper functions mask_average and lime_class_heatmaps aggregate the per-image masks into class-level heatmaps to visualize which regions are generally important for each bandwidth class.
Output Heatmap
