Interpreting the socio-technical interactions within a wind damage-artificial neural network model for community...

Area of Science:

• Engineering
• Computer Science
• Social Science

Background:

• Machine learning (ML) adoption is widespread, yet its 'black box' nature raises concerns.
• Modeling community-level building damage from extreme wind events is complex, often focusing solely on structural capacity.
• Existing models struggle to cohesively integrate factors like wind loading, debris, and social vulnerability.

Purpose of the Study:

• To explore artificial neural networks (ANNs) combined with graph theory for modeling and interpreting spatial building damage from extreme wind events.
• To address the 'black box' problem in ML by analyzing ANN internal patterns.
• To identify key predictors of structural damage at a community level.

Main Methods:

• Developed and analyzed two distinct artificial neural network (ANN) models for predicting the spatial distribution of structural damage.
• Applied graph theory to investigate the internal workings and patterns within the ANNs.
• Integrated data on wind loading, debris impact, and social characteristics to assess their influence on damage prediction.

Main Results:

• The study successfully modeled the spatial distribution of building damage using ANNs.
• Graph theory analysis revealed the internal decision-making processes of the ANN models.
• Social parameters were identified as critical factors in predicting structural damage, offering insights beyond physical loading models.

Conclusions:

• Artificial neural networks, when analyzed with graph theory, can effectively model complex phenomena like wind-induced building damage.
• Social characteristics play a significant role in community vulnerability and structural damage, a factor often underrepresented in traditional models.
• This research provides a method to interpret 'black box' AI models, enhancing trust and applicability in disaster risk assessment.