American
International
University-
Bangladesh

Students & Supervisors

Abstract

Natural disasters such as floods, earthquakes, and pandemics often outstrip traditional response capabilities, because they rely on non-integrated modelling systems, long turnarounds, and unconnected data analyses. Physics-based simulations whether hydrodynamic for flooding, seismic for earthquakes, or epidemiological for pandemics, tend to be computationally demanding and takes too long to develop and can’t be used quickly enough to inform decisions. This is starting to change, with AI breaking down these limitations. One such example is a neural surrogate model of coastal ocean circulation that accelerates the 12-day forecasts, from a runtime of 9,908 seconds on 512 CPU cores to 22 seconds on a single A100 GPU, while maintaining high-fidelity results, at a speedup of 450× or higher. Likewise, an ML emulator for predicting flood inundation across Chicago reached R-squared values over 0.96 and median relative errors close to 1 percent and ran in about 0.006 s/event. Following such advancements, in this study we introduce an AI directed “crisis twin” that integrates fast physics-based surrogates with reinforcement learning and social-economic knowledge. In this framework, the flood dynamics are simulated alongside structural fragilities and disease spread. During post-disaster recovery and reconstruction, proposed deep reinforcement learning model outperform that of Monte Carlo algorithms with 33.6% reduction on delays, 46.4% in comparison to genetic algorithms and overall cost reduced by up to 18.6%. At the other end of the spectrum, social media-enhanced RL frameworks have enabled the attainment of up to 30% increase in resource satisfaction and equity scores as high as 0.5 as against traditional allocation schemes. The crisis twin framework grounds these components as a single pipeline: AI-powered simulations for near real-time high-fidelity, high-speed hazard modeling; reinforcement learning for adaptive resource allocation in dynamic environments; and socio-economic data for equitable, targeted, context-aware response strategies. This allows decision-makers to virtually test out many “what-if” interventions, compare alternate outcomes, fine-tune policies before deploying them in the real world, and therefore be better prepared, with fewer casualties, less economic damage, and better long-term resilience. In the end, this approach realises disaster response scenarios from reactive, silo models to proactive and integrated simulations, by exploiting the speed of neural surrogates, flexibility of reinforcement learning, and realistic socio-economic context. Attendees will learn about how AI can enable faster, more intelligent and more equitable crisis response and thereby offer a glimpse of its potential as a life-saving, decision-support technology for the layered, multi-hazard emergencies of today’s world.

Keywords

Publication Details

• Type of Publication:
• Conference Name: 1st International Conference on Science and Humanities for Sustainable Development
• Date of Conference: 23/10/2025 - 23/10/2025
• Venue: Dhaka University of Engineering & Technology, Gazipur Gazipur-1707, Dhaka, Bangladesh
• Organizer: Dhaka University of Engineering & Technology