Interpretable AutoML for Post-Earthquake RC Building Damage: Insights from the 2023 Türkiye Sequence

Published

July 17, 2026

Author(s)

Mohammadamin Hariri Ardebili, Siamak Sattar

Abstract

Post-earthquake reconnaissance produces rich yet heterogeneous information that is hard to synthesize quickly for triage and decision-making. We summarize a two-part study of the February 2023 Türkiye earthquake sequence: (i) structured field data collection on >240 reinforced-concrete buildings across the most affected provinces; and (ii) an interpretable automated machine learning (AutoML) pipeline that classifies damage and extracts physically meaningful drivers via SHAP (Shapley Additive Explanations). Using 15 building and shaking features, the multi-class task (light/moderate/severe) achieved 0.85 test accuracy with Random Forest and CatBoost after addressing class imbalance. A complementary binary task for critical damage utilized composite capacity–style indices (column/wall index and density) in conjunction with maximum PGA, PGV, and CAV (peak ground acceleration/velocity, and cumulative absolute velocity). It achieved test accuracy above 0.95. Interpretation indicates that building features dominate light damage, whereas ground motion measures — especially CAV and PGV — govern severe/critical damages. The overall recommendation is that if column and wall indices remain below 0.07 and 0.08, respectively, they indicate an increased likelihood of critical structural damage.

Conference Dates

July 13-17, 2026

Conference Location

Portland, OR, US

Conference Title

13th National Conference on Earthquake Engineering

Pub Type

Conferences

Download Paper

Local Download

Citation

Hariri Ardebili, M. and Sattar, S. (2026), Interpretable AutoML for Post-Earthquake RC Building Damage: Insights from the 2023 Türkiye Sequence, 13th National Conference on Earthquake Engineering , Portland, OR, US, [online], https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=961015 (Accessed September 2, 2026)

Additional citation formats