Data and code for "Predicting Seismic Site Amplification from KiK-net Recordings Using a Hierarchical FiLM-Conditioned Fourier Neural Operator"

Reproducibility package (Tier 1, Minimum Mandatory Standard under the "Policy on EESD Papers Utilizing Data-Driven Approaches") for the manuscript "Predicting Seismic Site Amplification from KiK-net Recordings Using a Hierarchical FiLM-Conditioned Fourier Neural Operator." Contents:(i) Held-out test set of 39,692 KiK-net records (105 spectral periods spanning 0.01-10 s; 30-node Vs profiles);(ii) Trained model checkpoints for the proposed FNO-FiLM operator and every reproduced baseline (DNN-XGB after Nguyen et al., SSRNet after Li et al., Random Forest after Kim et al.);(iii) A runnable inference pipeline (Python 3.9-3.10, TensorFlow 2.10) that regenerates the paper's performance figures (Figures 9, 10, 11, 13, 14, 16) with a single command;(iv) Fitted preprocessing scalers, split indices, SHA-256 checksums, and full README. Runtime: approximately 5 minutes on a single GPU, 15 minutes on CPU only. Source data derive from the KiK-net strong-motion network operated by the National Research Institute for Earth Science and Disaster Resilience (NIED), Japan (https://www.kyoshin.bosai.go.jp). Only processed quantities are redistributed; users should acknowledge NIED as the original source of the recordings.