On November 12, 2026, Carson Dudley, Reiden Magdaleno, Yutong Bi, Anthony Valencia, Xiaofeng Liu, and Marisa Eisenberg at the University of Michigan released a preprint titled Universal Infectious Disease Forecasting with Mantis 2. The research introduces an open-source epidemiological foundation model that forecasts outbreak trajectories up to 16 weeks ahead, enabling public health research groups and operational surveillance units to generate multi-pathogen projections without engineering customized models for every disease and jurisdiction.
What happened
Infectious disease forecasting historically relied on pathogen-specific statistical architectures or handcrafted compartmental simulations calibrated to individual jurisdictions. While effective locally, this approach requires modeling teams to rebuild and recalibrate models whenever a novel variant or pathogen emerges. General time-series foundation models learn broad temporal patterns, but they can project biologically implausible trajectories when applied to epidemiology because they lack mechanistic constraints such as susceptible population depletion.
To address this gap, researchers at the University of Michigan developed Mantis 2, an infectious disease time-series foundation model supported by the U.S. Centers for Disease Control and Prevention (CDC) Center for Forecasting and Outbreak Analytics (CFA) Insight Net cooperative agreement. Mantis 2 builds on Amazon’s pretrained Chronos-2 backbone and expands upon the original Mantis architecture. While the original model relied solely on simulated outbreaks to forecast a single outcome eight weeks ahead, Mantis 2 adopts a hybrid training curriculum pairing 20,000 synthetic epidemic trajectories from the mechanistic Reservoir simulator with over 1,000 observational real-world disease time series.
The updated system accepts up to 512 weeks of historical surveillance context and uses latent feedback recurrence to project up to 16 weeks ahead. Crucially, Mantis 2 operates multivariately, forecasting up to six surveillance streams simultaneously—including cases, hospital admissions, emergency department visits, deaths, and environmental covariates.
Methodology and key findings
The authors evaluated Mantis 2 across 30 empirical forecasting tasks spanning diverse respiratory, vector-borne, and enteric pathogens across varied regions, testing both in-domain holdouts and out-of-domain evaluation on unencountered diseases. Baselines included Chronos-2, a three-layer long short-term memory (LSTM) network, exponential smoothing (ETS), seasonal autoregressive integrated moving average (SARIMA), seasonal-naive, and persistence.
Mantis 2 achieved the lowest aggregate Weighted Interval Score (WIS) across all horizon tiers: 0.20913 over weeks 1 to 4, 0.28439 over weeks 1 to 8, and 0.36965 over weeks 1 to 16, consistently outperforming Chronos-2 and statistical models. In outbreak milestone detection, Mantis 2 missed approximately 30 percent fewer epidemic peaks than the strongest baseline model, identifying anticipated peaks more than six weeks in advance on average.
Ablation experiments across eleven model variations demonstrated that simulation-rich pretraining provides the greatest value when local surveillance history is short or sparse. As empirical records accumulate, observational data guides the trajectory, but mechanistic simulations consistently prevent catastrophic forecast errors by enforcing biologically plausible bounds.
Why it matters for research practice
For epidemiology groups and public health agencies, Mantis 2 bridges traditional epidemiological models and modern deep learning. Public health units often face trade-offs: compartmental models enforce epidemic dynamics but struggle with high-dimensional noise, whereas unconstrained neural networks overfit or project impossible exponential growth.
By combining mechanistic simulations with empirical surveillance, Mantis 2 validates the role of synthetic data in research pipelines. Simulated data teaches the network fundamental transmission mathematics, while observational data adapts it to reporting noise and nowcasting delays. This hybrid framework lowers the barrier to deploying automated forecasts in public health departments that lack the biostatistical resources to maintain bespoke models.
What changes for researchers
Open-source release of the model weights and codebase allows academic laboratories and public health modelers to deploy unified forecasting across pathogens without constructing distinct modeling pipelines. Modeling teams can supply up to six surveillance signals to generate probabilistic projections of healthcare burden.
Additionally, the work provides a reproducible blueprint for scientific foundation models in data-scarce domains. Consortia can use synthetic epidemiological simulations to generate training diversity for emerging pathogens before real-world surveillance data accumulates, establishing reliable baselines ahead of seasonal surges and outbreak threats.
Sources
- University of Michigan and CDC Center for Forecasting and Outbreak Analytics Insight Net: arXiv:2611.16935v1 (2026-11-12)
- Mantis Open Source Project Repository: GitHub Repository (2026-11-12)
