The app is designed for visual analysis and comparison of three deep learning architectures — LSTM, CNN, and hybrid CNN-LSTM — on the task of binary signal-blockage classification in THz/mmWave high-frequency wireless systems. Users can upload a custom CSV of power sequences (8×64 matrix) or use the built-in sample, then get predictions from all three models directly in the browser.
Technical decisions
All three neural networks were trained in Python with Keras and exported to TensorFlow.js format, enabling fully client-side inference with no server-side compute. Input data is normalised with MinMaxScaler and fed to the models as a tensor of shape [1, 8, 64]. A metrics table shows accuracy, precision, recall, and F1-score for each model. Signal visualisation is built with Chart.js, displaying power levels in dB over time.
Challenges and trade-offs
The main challenge was adapting Keras models to TensorFlow.js Converter v4.22.0 and correctly loading binary weights via Next.js static hosting. The architectures differ significantly in capturing temporal dependencies: pure LSTM achieved the highest accuracy (68 %), while CNN-LSTM reached the best recall (71 %), which is critical for minimising missed blockages in real-world scenarios.