Determining the drivers of tsunami damage to roads in the 2015 Illapel tsunami, Chile, and transferability to the 2018 Sulawesi tsunami, Indonesia

Tsunamis pose a considerable threat to coastal communities, causing widespread destruction to the built environment, including roads, which are critical for disaster response and recovery. This study investigates the features influencing road damage in Coquimbo, Chile, following the 2015 Illapel tsunami and tests the transferability of these data to Palu, Indonesia, after the 2018 Sulawesi Tsunami. Machine learning models, Random Forest (RF) and Extremely Randomised Trees (XT), are utilised to assess the relative importance of hydrodynamic and non-hydrodynamic tsunami hazard metrics, roading asset attributes and shielding features. In feature importance testing, Scour Presence shows high importance followed by flow depth and road elevation. In contrast, flux and road orientation have negligible influence. Within Coquimbo, both models achieve high accuracy (99%) in predicting road damage levels. However, transfer analysis reveals low cross-regional performance when trained on Coquimbo (58%). The models achieve higher accuracy when trained on Palu (81%). These findings highlight the need for tailored damage models incorporating local characteristics. This study can inform disaster risk management by improving understanding of road vulnerability, tsunami resilience planning, and emergency response planning in tsunami-exposed regions. Not applicable.