Add U-Net project for seismic data interpolation and reconstruction

[Copilot]

Implements a complete PyTorch U-Net pipeline for reconstructing missing traces in 2-D seismic sections from scratch on an otherwise empty repository.

Architecture

• unet_model.py — Standard encoder-decoder U-Net with skip connections. 2-channel input (masked data + binary acquisition mask); 1-channel output (reconstructed section).
• dataset.py — SeismicDataset supporting .npy and SEG-Y inputs, random/regular subsampling masks, patch extraction, and per-section normalisation.
• utils.py — ReconstructionLoss (weighted L1 + MSE with configurable up-weighting on missing-trace locations), SNR/PSNR metrics, checkpoint helpers, visualisation.
• train.py — AdamW + cosine LR schedule training loop. Bundles a synthetic seismic data generator (dipping plane-wave events) so the pipeline runs end-to-end without real data.
• predict.py — Inference script with arbitrary-size input handling (reflect-pad → infer → crop), per-sample SNR/PSNR reporting, and optional side-by-side comparison figures.

Quick usage

# Zero-setup demo — generates synthetic data automatically
python train.py --epochs 50 --batch_size 8

# Train on real data
python train.py --train_data data/train.npy --val_data data/val.npy --missing_ratio 0.5

# Reconstruct
python predict.py --checkpoint outputs/best_model.pth --input data/incomplete.npy --output data/reconstructed.npy --plot

Tests

38 unit tests in tests/test_seismic_unet.py covering model shapes, gradient flow, mask generators, dataset I/O, loss behaviour, metrics, and padding round-trips.