RESEARCH ON BUILDING A MULTIMODAL DEEP LEARNING MODEL FOR URBAN-SCALE EARTHQUAKE DAMAGE PREDICTION USING NUMERICAL ANALYSIS DATA

This study proposes a multimodal deep learning model for urban-scale earthquake damage prediction that integrates seismic waveform data and building–ground attributes. Using numerical simulations of about 30,000 buildings in Kobe’s Nada Ward under five earthquake scenarios, a 1D Convolutional Neural Network (CNN) processes waveform data, while a Multi-Layer Perceptron (MLP) handles structural and geotechnical attributes. Feature-wise Linear Modulation (FiLM) is tested for modality fusion. Results show that combining waveform and intensity measures enhances prediction accuracy and generalization, while FiLM provides limited improvement.