National Disaster Management Authority

Contents

Executive Summary

This panel discussion at an AI summit examines how India and other nations can institutionalize AI within national disaster risk reduction (DRR) frameworks, moving beyond pilot projects to population-scale resilience systems. The core thesis is that the next frontier in DRR is not better algorithms alone, but embedding AI within sovereign, interoperable governance architectures that integrate space technology, meteorology, digital platforms, and community-centered dissemination. Speakers emphasize the critical need for computational infrastructure, human-in-the-loop governance, hybrid AI-physics models, and federated data architectures to enable early warning systems and actionable intelligence at the last mile.

Key Takeaways

AI ≠ Algorithms; AI = Infrastructure + Governance + Data: The bottleneck for disaster resilience in India is not AI sophistication but computational infrastructure (40 PF vs. 1,000+ PF needed), energy/cooling capacity, data interoperability across federal systems, and governance models that keep humans accountable for life-saving decisions.
Hybrid Physics-ML Models Are the Realistic Path: Institutions like the UK Met Office, IMD, and European centers are adopting hybrid approaches combining physical models (which encode scientific reasoning) with ML (which captures complex patterns in big data). Full replacement of physics models with ML is neither feasible nor trustworthy; blending is the standard.
Digital Twins Enable Real-Time Operational Intelligence: Creating bridged representations of physical infrastructure with real-time sensor layers (thermal, structural, asset, people) transforms disaster response from reactive to precision-targeted. This extends beyond weather to cyber-physical systems.
Federalism + Data Sovereignty = Architectural Challenge: India's federal structure (national + 28 state governments) requires decentralized but integrated data architectures. Central intelligence must be replicable to state/local agencies without centralizing control. Edge-based AI and federated learning address this.
Startups + Public Institutions + Private Tech = Necessary Triangle: No single stakeholder solves this. IMD and NDMA provide risk knowledge and operational authority; public institutions like universities contribute data and models; private sector (Google, HPE, etc.) provides infrastructure and scalability; startups innovate in niche applications (water body forecasting, risk extraction).

AI, Resilience, and Governance at Scale in India

Key Topics Covered

AI Institutionalization in DRR: Moving from isolated projects to national resilience architectures
Hybrid AI-Physics Models: Complementing physical weather forecasting with machine learning rather than replacement
Digital Twins: Creating bridged representations of physical and virtual spaces for operational response
Computational Infrastructure Gaps: India's acute shortage of high-performance computing relative to global standards
Sovereign Data Architectures: Ensuring interoperability across federal and state systems while protecting national data
Early Warning Systems: Technology and governance requirements for effective alert dissemination
Cyber-Physical Resilience: Extending disaster definitions to include cyber attacks and system vulnerabilities
Human-in-the-Loop Governance: Preventing over-automation in critical decision-making affecting human lives
Last-Mile Inclusion: Deploying AI in low-connectivity, high-risk environments using edge computing
Data Integration and Standardization: Aggregating fragmented data across multiple ministries and agencies
Private-Public Partnerships: Role of technology providers and startups in developing resilience systems
UN Early Warning for All Initiative: 2027 target of universal early warning coverage across all countries

Key Points & Insights

Infrastructure as Bottleneck, Not Algorithms: India has approximately 40 petaflops (PF) of computing capacity; the U.S. deploys exascale systems (1,000+ PF). The deficit is not in AI algorithm sophistication but in computational power, energy, cooling, and water infrastructure to process real-time geospatial and satellite data. Private-public partnerships are essential to bridge this gap, as government alone cannot fund $400M–$1B supercomputer systems.
Digital Twins as Operational Bridge: Mauritius's approach of creating digital twin architectures—connecting physical infrastructure maps with real-time sensor data (thermal imaging, heartbeat detection, structural plans)—enables emergency services to locate victims and resources precisely. This extends disaster definitions beyond floods and cyclones to include cyber attacks and system failures in critical infrastructure.
Human-in-the-Loop Non-Negotiable: Minister of Mauritius emphasized that early warning message verification cannot be 100% automated. Infected or false messages trigger adrenaline surges and societal disruption. Critical decisions (e.g., cell broadcast alerts) must involve human verification; machines should inform human decision-makers, not replace them.
Hybrid Models, Not Replacement: The UK Met Office explicitly rejects replacing physics-based weather models with machine learning alone. The path forward involves step-by-step blending—hybrid models, ensemble outputs, and co-developed benchmarking with partners like India (WCSSP India, Wiser Asia Pacific). Trust requires transparent, fair comparisons using metrics that matter to end users, not just ML performance indicators.
Data Fragmentation as Crisis: Critical data exists but is scattered across meteorological agencies, water boards, ministry systems, and social media. Real-time actionable intelligence requires federated architectures that centralize data integration while maintaining local agency autonomy and data sovereignty. AI's role is converting chaotic, noisy, unstructured data into structured hazard-specific insights.
Agentic AI for Disconnected Environments: Google Cloud's architecture demonstrates how central intelligence (single source of truth) can be replicated to edge devices, enabling tactical response in air-gapped, low-connectivity disaster zones. Applications can run on minimal infrastructure (basic GPUs) while maintaining federated connection to central systems when connectivity resumes.
Startups' Role in DPGs: Startups (e.g., Vasar Labs) leverage AI to address data gaps—translating satellite and radar data into real-time nowcasts for unmonitored water bodies. Near-real-time forecasts for ~1 million small, unregulated dams across India require AI-driven solutions that extract risk information from satellite data (30-minute intervals) and translate it into hydraulic models.
AI + Unstructured Data for Historical Risk Mapping: AI can extract location-specific hazard frequency, intensity, and damage information from news archives and unstructured sources to build parametric datasets. This feeds insurance, risk assessment, and location-specific vulnerability databases—historically unavailable for events like earthquakes and localized floods.
Observational Network Expansion + AI Processing: NDMA's roadmap includes deploying automated weather stations in every village, quadrupling seismometers, and expanding landslide instrumentation. Exponential data growth requires proportional advances in processing capacity and AI-based inference. The challenge is not data collection but converting raw observations into actionable citizen-facing early warnings.
Affordability via GPU-Based Models: For small island nations and low-resource countries, traditional supercomputers are unaffordable. AI-driven "box models" deployed on GPU clusters offer a path to early warning systems at 1/100th the cost of exascale HPC, though with trade-offs in model sophistication and lead time.

Notable Quotes or Statements

Minister of Mauritius (IT, Communication & Innovation):
"Machines cannot decide for humans; humans decide for machines."
— Emphasizing that human-in-the-loop governance is non-negotiable for life-affecting decisions.
Minister of Mauritius:
"Disaster is not just the flood, the cyclone, the drought. Disaster can also be the cyber security attacks."
— Expanding disaster definitions to cyber-physical resilience.
Beth Woodham (UK Met Office):
"We don't know what the answer to this solution is yet... we're not going to have a complete shift. We are going to do this step by step."
— Acknowledging the uncertainty and incremental approach to AI-physics model integration.
Som Sasangi (HPE):
"It's not a technology barrier; it's infrastructure, scale, and procurement processes."
— Redirecting the problem from algorithm sophistication to systems and policy constraints.
Dr. Britja Mahatra (IMD Director General):
"A box model has come up where you can give it to small island nations and with a few GPU nodes they can have the forecast."
— Highlighting affordability breakthroughs for resource-constrained nations.
Dr. Krishna Vasa (NDMA):
"The road map is not sufficiently clear... how do data centers and individual early warning agencies interact?"
— Identifying the unresolved architectural challenge of integrating central infrastructure with distributed agency operations.

Speakers & Organizations Mentioned

Government Representatives

Dr. Vin Ram Toul – Minister for Information Technology, Communication & Innovation, Republic of Mauritius
Dr. Britja Mahatra – Director General, India Meteorological Department (IMD)
Dr. Krishna Vasa – Member/Head of Operations, National Disaster Management Authority (NDMA), India
State Disaster Management Authorities – Representatives from Tamil Nadu, Andhra Pradesh, and Telangana

Multilateral & International Organizations

UN – Early Warning for All Initiative (2022–2027 target)
WMO (World Meteorological Organization) – Referenced in early warning coordination
UK Met Office – National meteorological agency (UK)
WCSSP India, Wiser Asia Pacific – International collaboration networks for weather forecasting

Technology & Private Sector

Beth Woodham – Senior Manager, UK Met Office (disaster risk reduction, forecasting innovation)
Som Sasangi – Former Senior Vice President & Managing Director, Hewlett Packard Enterprise (HPE) India (geospatial & climate analytics)
Pankaj Shukla – Head of Customer Engineering, Google Cloud India (cloud AI, hazard mapping, predictive analytics)
Nikil Kumar – CEO and Co-founder, Vasar Labs (AI for disaster risk reduction, water body forecasting)

Institutions & Datasets Referenced

National Supercomputer Mission – India's initiative (established 2015, ₹4,500 crore investment)
Survey of India – Geospatial and earthquake data provider
Oak Ridge National Laboratory, Argonne National Laboratory – U.S. exascale computing facilities (Frontier, Aurora systems)
European Center for Medium-Range Weather Forecasts (ECMWF) – Physics-based weather modeling reference

Technical Concepts & Resources

Computational Infrastructure

Petaflops (PF): Units of computing power (10^15 floating-point operations/second). India: ~40 PF; U.S. systems: 1,000–1,800 PF (exascale).
GPU-based AI Models: Lower-cost alternative to supercomputers for forecasting in resource-constrained regions.
Box Models: Lightweight, AI-driven meteorological models deployable on edge devices with minimal computational footprint.
U.S. Exascale Systems Referenced:
- Frontier (Oak Ridge): 1.3 exaflops
- Aurora (Argonne): 1.0 exaflop
- El Capitan: 1.8 exaflops

AI/ML Approaches

Hybrid Physics-ML Models: Combining traditional partial differential equation (PDE)-based weather models with machine learning for pattern recognition and data-driven adjustments.
Ensemble Methods: Combining outputs from physics-based and ML models for consensus forecasting.
Agentic AI: AI systems that autonomously reason over multimodal data and recommend or execute actions.
Multimodal Models: Gemini and other hyperscaler models referenced as capable of processing diverse data types (text, imagery, sensor data).
Federated Learning: Distributed model training while keeping data decentralized across agencies.
Zero-Trust Architecture: Security model for disconnected, air-gapped disaster response environments.

Data & Observational Networks

Real-Time Satellite Data: 30-minute intervals for water body, flood, and dam monitoring.
Radar Data: Nowcasting (0–6 hour forecasts) from India Meteorological Department.
Digital Twin Layers:
1. Data layer (observational, satellite, structural)
2. Modeling layer (hazard, vulnerability, risk models)
3. Asset & people layer (geospatial asset inventories, population distributions)
4. Workflow layer (decision support, dissemination)
Seismic Networks: Micro-earthquake monitoring (magnitude <3) for hazard pattern recognition.
Automated Weather Stations (AWS): Planned deployment in every village in India (next 5 years).
Parametric Risk Databases: Location-specific, AI-extracted hazard frequency, intensity, and damage datasets from historical records.

Forecasting & Early Warning

Nowcasting: Real-time or near-term (0–6 hour) hazard prediction.
Lead Time: Time between forecast issue and hazard impact (critical metric for evacuation effectiveness).
Impact-Based Warnings: Forecasts framed around societal consequence (e.g., "flooding of roads X, Y, Z") rather than technical metrics (e.g., "100mm rainfall").
Cell Broadcast Systems: Government-grade emergency alert systems for mobile devices (referenced as Mauritius deployment).

Governance & Architecture Frameworks

DPI (Digital Public Infrastructure): India's model of sovereign, interoperable systems (e.g., Aadhaar, India Stack) adapted for disaster resilience.
DPG (Digital Public Goods): Open-source, globally accessible systems addressing public challenges; India positioning DRR AI systems as DPGs.
Human-in-the-Loop (HITL) / Human-on-the-Loop (HOTL): Governance models where AI informs but humans authorize critical decisions.
Federated Architecture: Central data integration with distributed agency autonomy (resolves federalism in India context).
On-Prem, Edge, and Disconnected Deployment: Multi-tier deployment supporting central cloud, regional on-premises, and fully offline tactical operations.

Standards & Evaluation

Benchmarking & Standardization: Fair comparison metrics between physics-based and ML models (UK Met Office emphasis).
Explainability Standards: Requirements for AI systems informing life-saving decisions (referenced but not detailed in transcript).
Interoperability Standards: Enabling diverse AI models and platforms to work within sovereign data architectures.

International Frameworks & Targets

UN Early Warning for All (2022–2027): Universal early warning system coverage by 2027; currently <50% of countries have coverage.
WCSSP India, Wiser Asia Pacific: International collaboration networks for co-developing forecasting models and standards.
South-South Cooperation: Knowledge and resource sharing among developing nations on resilience.

Contextual Notes

The discussion reflects India's unique position as:

A mega-disaster-prone nation (Himalayan earthquakes, cyclones, floods, heat waves across diverse climates and urban centers)
A federal democracy with 28 states requiring decentralized governance
A DPI innovator (Aadhaar, India Stack) now scaling those lessons to resilience
A middle-income nation with world-class scientific institutions (IMD) but constrained computing infrastructure
A potential exporter of resilience models and DPGs to other developing and small island nations

The panel discussion positions AI not as a technology silver bullet but as one component of a broader transformation requiring governance reform, infrastructure investment, and institutional integration.