S1.E1. AI for Uncertainty: Predicting epidemics and disasters with Alex Rodriguez

Scientific Concepts, Discoveries, and Real-World Phenomena

Complex dynamical systems

• Multiple interacting processes whose behavior is not fully understood.
• Example domains:
  • Epidemics (biological processes)
  • Weather (atmospheric processes)
  • Consequences for agriculture (e.g., crop impacts)

Epidemic dynamics and population-level forecasting

Predicting:

• The current state (who is infected, with potential skew toward certain populations/communities)
• Future trajectories under different scenarios and interventions

Key challenges:

• Data gaps
• Reporting delays
• Incomplete observations
• Limited public-health testing capacity

Disaster resilience and network failures

• Using AI to infer where network failures may occur after disruptions.
• Example framing:
  • Inferring power/water network failures following earthquakes or hurricanes
  • Deciding where to dispatch repair crews
• Goal: improve community resilience after disruptions, including biological threats (e.g., COVID-19)

Time series forecasting across domains

Applied to:

• Epidemics (case/hospitalization/case-burden curves)
• Weather (e.g., cyclone/hurricane forecasts)
• Other time-dependent quantities (temperature, humidity, economic indicators)

Emphasis: time series modeling is not only predictive, but tied to decision-making.

Uncertainty Quantification (UQ)

• Models should provide predictions plus confidence, enabling decision-makers to hedge risk.
• Importance of calibrated uncertainty:
  • When a model claims it is confident (or not), it should be statistically reliable
• Example of uncertainty failure:
  • During COVID-era forecasting, later-horizon uncertainty performed so poorly that it was removed from CDC display.

Conformal prediction

• Produces prediction bounds with mathematical performance guarantees (on average/under assumptions).
• Mismatch problem: when environments shift, assumptions can fail.

Adaptive conformal prediction (concept)

• Extends conformal prediction to non-stationary environments (changing regimes).
• Epidemic example:
  • Variant replacement (e.g., Delta → Omicron) changes transmissibility/behavior.

Multimodal data for better uncertainty

• Proposed conceptual approach:
  • Incorporate multiple data modalities to improve uncertainty estimation while preserving conformal/adaptive conformal guarantees.

Trajectory-level uncertainty (and temporal correlations)

• Instead of evaluating one-step-ahead predictions independently:
  • Treat outcomes as belonging to trajectories with temporal dependence.
• Weather application:
  • Cyclone/hurricane trajectories require preserving temporary/trajectory correlations.

Epidemic and surveillance modeling under confounding signals

Early pandemic challenge described:

• Influenza surveillance was symptom-based.
• COVID-like flu symptoms were included in flu data.
• This motivated leveraging/transitioning influenza-based models to handle COVID-influencing signals.

Domain adaptation

Conceptual approach:

• Train in one regime (pre-COVID influenza)
• Adapt to a new regime (COVID-era mixed signals)

Example mechanism named:

• Domain adaptation between changing data-generating processes.

CDC forecasting initiative / ensembling

• Models submitted repeatedly (e.g., weekly in influenza season).
• CDC combines team predictions into an ensemble intended to be robust for decision-making.

Foundation models for science

Core idea:

• Pretrain on broad data/time-series
• Then specialize/fine-tune for a domain/task (epidemics, weather, etc.)

Suggested directions:

• Multi-disease learning (learn from multiple diseases)
• Multi-task learning (predict future curves, peaks, outbreak trends/events)
• Incorporate multimodal inputs

Conceptual parallels mentioned:

• chemistry / protein folding / climate forecasting.

Simulators and physics-/theory-informed learning

• Use scientific simulators derived from domain experts/theory/equations when real data is scarce.
• Main issue:
  • Simulators are imperfect (assumptions may not match reality)
• Therefore, methods must balance trust between:
  • simulation outputs
  • real-world data
• Related framing:
  • AI for science
  • physics-informed machine learning
  • knowledge-guided machine learning

---

Methodologies / Workflows Mentioned

Epidemic/flu prediction workflow (as described)

• Develop deep learning models for influenza forecasting
• Submit weekly predictions to the CDC forecasting initiative
• CDC constructs an ensemble from multiple teams for decision-making
• During COVID:
  • Adapt focus as flu surveillance signals become mixed with COVID-like symptoms
  • Perform prediction and adaptation using:
    • new signals
    • historical influenza data
  • Use approaches analogous to domain adaptation

Uncertainty quantification framework

• Use conformal prediction to generate prediction bounds with performance guarantees (under assumptions)
• Under non-stationarity, use adaptive conformal prediction
• Improve uncertainty with multimodal data inputs
• Extend to trajectory-level uncertainty to reflect temporal correlations in confidence

Network failure inference after disasters

• After an event (earthquake/hurricane):
  • Use partial observations on network nodes
  • Infer probable failure locations
  • Decide where to dispatch repair crews to inspect/recover

Foundation-model style scientific learning (conceptual approach)

• Pretrain broadly across:
  • datasets, time-series, tasks, diseases
• Specialize to a domain (e.g., epidemics at specific locations)
• Employ multi-task learning (event detection, trend prediction, peak timing)
• Employ multi-disease learning

Hybrid learning with simulations

• Combine machine learning with scientific simulations and domain knowledge
• Balance reliance on:
  • imperfect simulators
  • limited real data
• Goal:
  • reduce data requirements while improving realism/consistency

---

Researchers or Sources Featured (Mentioned by Name)

• Professor Alex Rodriguez (University of Michigan) — primary featured researcher
• Georgia Tech — institution where he received his PhD (source institution)
• CDC (U.S. Centers for Disease Control and Prevention) — forecasting initiative and decision-making context
• ICML — referenced for a “best paper award” (venue; no specific author named)
• ACM — referenced for a dissertation award (no specific author named)
• Georgia Tech College of Computing — dissertation award institution (not a person)