True multi-hazard analysis captures cascading and compounding interactions—rather than stacking independent layers (e.g., earthquakes triggering landslides; floods amplifying post-wildfire erosion; climate and geohazard interactions). Current practice remains overlay-based, black-box, static, siloed, and point-predictive. This session spotlights AI that encodes physics-based interactions, is interpretable and uncertainty-aware, learns continually, future projections, overcomes data scarcity via transfer learning and simulation, enables federated multi-agent operations, and fuses multimodal foundation models for time–space integration.

We invite contributions that implement and validate:
1. Higher-order interaction modeling (hypergraphs, attention, Physics-Informed Neural Networks) with explicit cascade activation/propagation rules.
2. Interpretable architectures with process-consistent explanations.
3. Dynamic Bayesian networks with online learning for non-stationary hazards.
4. Forecasting integrating projected climate, land use, and topography changes.
5. Transfer/zero/few-shot learning and physics-constrained generative simulation for sparse cascades; calibrated adaptation.
6. Federated multi-agent learning with privacy-preserving aggregation for cross-agency model updates.
7. Uncertainty-aware decision support with probabilistic ensembles, cascade-aware uncertainty propagation, decision-centric intervals.
8. Multimodal foundation models unifying digital elevation models, InSAR, seismic, hydrometeorology, and other data sources.
9. Infrastructure cascade vulnerability analysis via dependency graphs and higher-order networks with intervention prioritization.

We welcome susceptibility; nowcasting/forecasting/early-warning systems; quantitative risk for applications spanning but not limited to earthquake-induced cascades, flood-induced cascades, climate-driven sequences, and earthquake-flood compounded landslides.