Pith. sign in

REVIEW 3 major objections 6 minor 51 references

This paper presents MEmilio, an open framework that makes compartmental, metapopulation, and agent-based epidemic models share one description and one simulation core, with high-performance computing scale.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 00:14 UTC pith:ASFQKI3L

load-bearing objection A genuinely unified multi-paradigm epidemic simulation framework, but the headline ABM speedup claim compares unequal workloads and should be re-benchmarked. the 3 major comments →

arxiv 2602.11381 v2 pith:ASFQKI3L submitted 2026-02-11 q-bio.PE cs.MS

MEmilio -- A high performance Modular EpideMIcs simuLatIOn software for multi-scale and comparative simulations of infectious disease dynamics

Julia Bicker , Carlotta Gerstein , David Kerkmann , Sascha Korf , Ren\'e Schmieding , Anna Wendler , Henrik Zunker , Daniel Abele
show 28 more authors
Maximilian Betz Khoa Nguyen Lena Pl\"otzke Kilian Volmer Agatha Schmidt Nils Wa{\ss}muth Patrick Lenz Daniel Richter Hannah Tritzschak Ralf Hannemann-Tamas Julian Litz Paul Johannssen Marielena Borges Annika Jungklaus Manuel Heger Annalena Lange Elisabeth Kluth Kathrin Rack Vincent Wieland Jonas Arruda Sebastian Binder Margrit Klitz Martin Siggel Manuel Dahmen Achim Basermann Michael Meyer-Hermann Jan Hasenauer Martin J. K\"uhn
This is my paper
classification q-bio.PE cs.MS
keywords infectious disease modelingagent-based simulationmetapopulation modelscompartmental modelshigh-performance computingepidemic preparednesshybrid modelssurrogate models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper presents MEmilio, an open-source framework that aims to end the fragmentation of epidemic modeling software by letting population-based compartmental models, spatially structured metapopulation models, and individual-level agent-based models be specified, calibrated, and compared from one shared abstract description. The authors' central claim is that no existing framework offers this combination of multi-paradigm coverage, computational scalability, and software-engineering robustness, and that MEmilio fills that gap. They support the claim with benchmarks: a C++ core that outperforms a widely used statistical-language reference implementation by one to two orders of magnitude, an agent-based model that handles populations up to 256 million, near-ideal scaling for thousands of ensemble runs on a supercomputer, hybrid agent-compartmental models that retain extinction dynamics while cutting runtime by about 99%, and graph-neural-network surrogates that evaluate hundreds of scenarios in a fraction of a second. If the framework delivers what it claims, outbreak responders could switch between model granularities without rewriting code, run the same model on a laptop or a supercomputer, and quantify how modeling assumptions bias forecasts.

Core claim

The paper's central claim is that the absence of a harmonized abstraction across epidemic-model paradigms is a solvable software problem, and that MEmilio is the solution. Its load-bearing design is a flow-based formulation of compartmental dynamics: transitions are written as explicit flows between states rather than net derivative changes, giving ODE, SDE, Linear Chain Trick, and integro-differential-equation models a common numerical substrate. On top of this sits a graph-based metapopulation layer with pluggable mobility schemes, a mobility-based agent-based model with hourly time steps, and a temporal-hybrid model that switches from agents to compartments when case numbers cross a thres

What carries the argument

The flow-based ODE formulation is the central mechanism: each transition between compartments is a rate sigma_{j,i}, cumulative flows are integrated, and compartment sizes are reconstructed from net flows. This single abstraction lets ODE, SDE, Linear Chain Trick, and integro-differential models share the same simulation and analysis code, and it extends naturally to graph-based metapopulations by treating each node as a local flow model coupled through a mobility operator. The temporal hybrid switch — a threshold condition plus state-conversion routines — is the second load-bearing mechanism, translating agent-based states into compartmental states so that the efficient population model tak

Load-bearing premise

The load-bearing premise is that the benchmark comparisons are setup-equivalent: in Supplementary Information Section J.1, MEmilio's agent-based model runs 120 hourly steps (5 simulated days) while the comparison daily-step simulators run 120 daily steps (120 simulated days), so the reported per-step speedups compare steps of different simulated-time length; if that premise fails, the efficiency advantage over existing agent-based frameworks is not established.

What would settle it

Run an end-to-end comparison on a single 120-day outbreak: 2,880 hourly steps in MEmilio against 120 daily steps in the comparison agent-based simulators, measuring wall-clock time, peak memory, and whether the comparison simulators complete at 128 million agents under the same hardware; if MEmilio's total runtime or memory envelope is not better under this matched workload, the headline speedup and scaling claims are not the whole story.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If MEmilio works as described, a research group can instantiate the same epidemic process as an ODE, linear-chain, integro-differential, metapopulation, or agent-based model from one shared specification, making inter-model comparison a configuration change rather than a reimplementation.
  • Nonexponential waiting-time distributions become a default option through the Linear Chain Trick and integro-differential formulations, allowing peak-timing predictions that are known to differ from exponential-base compartmental models.
  • The temporal-hybrid model is claimed to preserve agent-based extinction dynamics while cutting simulation runtime by roughly 99%, which would make stochastic individual detail affordable in the early phase of an outbreak.
  • At scale, the framework claims to handle 128–256 million agents and near-ideal strong scaling for thousands of ensemble runs, which would allow extensive scenario and uncertainty exploration on high-performance computing systems.
  • Integration with simulation-based inference and optimal control is claimed to be straightforward, allowing calibrated ensembles and optimized intervention strategies in a single workflow.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Inference: the benchmark speedups for the agent-based model may shrink if workloads are matched end-to-end; the reported comparison uses 120 hourly MEmilio steps (5 simulated days) versus 120 daily steps in the other simulators (120 simulated days), so the per-step comparison does not establish total-runtime superiority.
  • Inference: the unified model description could be leveraged to build automated equivalence tests across paradigms, turning the framework into a platform for systematically probing which modeling assumptions drive forecast differences.
  • Inference: the hybrid switching threshold could be tuned adaptively to keep approximation error below a user-specified tolerance, converting the runtime-accuracy tradeoff into a controllable knob rather than a fixed design choice.
  • Inference: if surrogate models trained on this framework's ensembles generalize, they could make real-time web-based outbreak exploration feasible without dedicated supercomputer access, which would change how quickly scenario results reach decision-makers.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. This manuscript describes MEmilio, an open-source epidemic simulation framework with a C++ core and a Python interface. The paper claims to harmonize ODE-, SDE-, LCT-, and IDE-based population-based models (PBMs), metapopulation models (MPMs), agent-based models (ABMs), temporal hybrid models, and graph-neural-network surrogates under shared abstractions, and to support calibration, scenario analysis, ensemble runs, and HPC-scale execution. Applications include PMCMC-based calibration of influenza and Ebola models, BayesFlow calibration of spatially resolved SARS-CoV-2 models for Germany and Spain, an NPI scenario analysis, an inter-model comparison, and multiple performance/scaling benchmarks (C++ vs. R, regional scaling, ABM scaling vs. Covasim and OpenCOVID, hybrid speedups, GNN surrogates, and strong scaling up to 16,384 cores). The paper concludes that MEmilio fills the gap of a multi-paradigm, scalable epidemic-modeling framework.

Significance. If the framework works as described, this is a potentially important contribution: it is one of the few open frameworks spanning PBM, MPM, ABM, and hybrid paradigms with an HPC orientation and reproducible calibration pipelines. The authors are to be credited for making the code, simulation scripts, parameter tables, and data publicly available, and for using repeated benchmark runs (e.g., median of 40 runs in Fig. 6a) and detailed SI parameter tables. The calibration sections and the inter-model comparison are specific and credible, and the framework's functional breadth appears genuinely novel. However, the paper's strongest performance claim — that MEmilio's ABM is faster than Covasim/OpenCOVID and scales beyond 64 million agents — rests on a benchmark whose workloads are not equivalent. That issue, together with a few other comparability gaps, means the efficiency half of the central claim is not yet established as stated.

major comments (3)
  1. [Fig. 6c / SI J.1] The ABM speedup benchmark compares different amounts of simulated time. MEmilio's ABM ran 120 hourly time steps, i.e., 5 simulated days; Covasim and OpenCOVID each ran 120 daily time steps, i.e., 120 simulated days. The reported per-step ratios (1.2–1.8x vs. Covasim, 8–12x vs. OpenCOVID) therefore do not support a wall-clock or end-to-end advantage. One simulated day requires 24 MEmilio steps, so even the upper per-step bound implies roughly 24/1.8 ≈ 13x longer per simulated day than Covasim and 24/8 = 3x to 24/12 = 2x longer per simulated day than OpenCOVID. No 120-day end-to-end benchmark is provided. The statement that MEmilio 'scales to 128 and 256 million agents, whereas Covasim and OpenCOVID did not complete runs beyond 64 million' is an out-of-memory observation, not a scaling comparison on equal footing. Please rerun the comparison with equal simulated-time horizons (e.g., 120 da
  2. [Fig. 6b / SI I.1] The claim that the graph-based MPM 'scales modestly with respect to regions regardless of the specific PBM' is not supported for the IDE model. SI I.1 explicitly states: 'MEmilio's IDE-based MPM is not yet able to incorporate mobility, hence its simulations are run without any exchange of individuals between regions.' The ODE- and LCT-based MPMs use instant mobility with node degree 100. The reported runtime differences (IDE about 15x slower than ODE and 3x slower than LCT) therefore conflate model formulation with missing mobility coupling. Please qualify the claim or extend the IDE MPM to include mobility before drawing cross-model scaling conclusions.
  3. [Fig. 6e / SI H] The GNN surrogate speedups of 806–32,078x compare a sequential CPU ODE solver with GPU-based GNN inference. As reported in SI H.2, the ODE baseline is executed on the CPU while the GNN uses a Tesla V100 GPU. This is a platform/infrastructure comparison as much as an algorithmic one, and the benchmark does not include the cost of generating the 1,000 training simulations. Please report a GPU-vs-GPU or CPU-vs-CPU comparison, or explicitly state that the 'speedup' includes the GPU advantage and that training-data generation is excluded. The claim that a trained surrogate provides fast on-the-fly predictions is reasonable; the headline 'speedup' should be qualified to avoid overstatement.
minor comments (6)
  1. [SI B.1.2] Typo: 'initival value problems' should be 'initial value problems'.
  2. [Results, 'Efficient and scalable implementations'] The phrase 'one time step in MEmilio's ABM is 1.2–1.8 (Covasim) and 8–12 (OpenCOVID) times faster than one time step' should explicitly state that MEmilio's time step is one hour while the comparators' time steps are one day, to avoid misleading readers who do not consult SI J.1.
  3. [SI J.2] 'shown in 11' should read 'shown in Table 11'.
  4. [Fig. 6b caption] The caption should state that the IDE-based MPM is simulated without mobility, per SI I.1, so that readers can interpret the comparison correctly.
  5. [Main text, NPI scenario analysis] The text reports selecting a representative simulation based on 'root mean square deviation' without defining the quantity or the time window; please add a pointer to SI E.
  6. [SI Table 10] For the IDE row, ρIDE, ξC, and ξI are listed as ConstantFunction(1), whereas the ODE/LCT rows use different numeric values. Please explain whether these are equivalent parameterizations or a deliberate calibration choice; otherwise the regional scaling comparison in Fig. 6b is confounded.

Circularity Check

1 steps flagged

Minor circularity in the inter-model comparison: the S1.1 agreement is by construction because the PBM transmission probability was harmonized to the ABM in that same scenario; the central framework and HPC claims are otherwise externally benchmarked.

specific steps
  1. fitted input called prediction [Supplementary Information Section F.2 (setup for Fig. 5d,e); main text Results, 'Model ensembles and quantification of model-specific assumptions', Fig. 5e S1.1]
    "The transmission probability on contact used in the ODE-, LCT- and IDE-based models is harmonized with the transmission rates of the ABM based on S1.1 and subsequently used for all other scenarios. ... Given Setup S1 with one location per location type (S1.1), we first observe that all considered models can be aligned under the same simplifying assumptions, giving almost identical epidemic dynamics with respect to the number of Exposed, ICU, and deceased cases (Fig. 5e S1.1)."

    The ODE/LCT/IDE transmission parameter is fitted to reproduce the ABM dynamics specifically in Setup S1.1, and the paper then presents the close agreement in S1.1 as an observed result. The cross-model match in that scenario is therefore by construction rather than an independent, out-of-sample comparison. The circularity is limited to this validation step: the later S2 and multi-location comparisons vary distributions and location structures after the same harmonized parameter was fixed, so those comparative conclusions retain independent content. The main claims of the paper — a unified multi-paradigm framework and computational efficiency — rest on external benchmarks (R/deSolve, Covasim, OpenCOVID) and public code, not on this fitted agreement.

full rationale

The paper is largely self-contained and externally benchmarked, so the central claims are not circular. The cross-framework ABM speed comparison (SI J.1) compares MEmilio's 120 hourly steps (5 simulated days) with Covasim's and OpenCOVID's 120 daily steps (120 simulated days); this is a benchmark-equivalence concern about the strength of the performance claim, not a circular derivation, so it does not increase the circularity score. The main circular flavor is the inter-model comparison in SI F.2: the ODE/LCT/IDE transmission probability is harmonized to the ABM using the S1.1 setup, and then the S1.1 agreement is reported as an observation. That agreement is thus partly built in by the parameter choice. However, it is a minor, non-load-bearing validation step: the framework's novelty, the hybrid-model speedups, the GNN surrogate speedups, and the large-scale ABM scaling results are all presented with explicit implementations and external or reproducible benchmarks rather than being derived from this harmonization. Self-citations appear (e.g., Bicker et al. for the hybrid model, Schmidt et al. for GNN accuracy), but they point to separate published works or are supplemented by in-paper descriptions, so they do not constitute a load-bearing self-citation chain. Overall circularity is low: score 2 reflects the one fitted-input-as-result step while recognizing that the central scientific and software claims stand on independent evidence.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

This paper contributes software rather than a new derivation, so the ledger is mostly about benchmark and configuration honesty rather than hidden fitted constants. Two disclosed but hand-chosen parameters affect reported results: rho harmonized to the ABM for the inter-model comparison, and the hybrid switching threshold tau. The modeling rests on standard stochastic-process mathematics (diffusion approximation, Erlang chain, Gillespie sampling) plus one ad hoc choice (the dABM single-well potential). The many fits in Tables 1-3 (influenza, Ebola, SARS-CoV-2) are openly disclosed calibration demonstrations and are not used to derive the central framework claim.

free parameters (2)
  • Transmission probability on contact rho for ODE/LCT/IDE models in inter-model comparison = 0.0409163 to 0.0300238 by age group (Table 7)
    Harmonized to match ABM transmission rates in Setup S1.1 (SI Section F.2), then reused in S1.2/S2.1/S2.2. The near-identical S1.1 trajectories in Fig. 5e are therefore partly guaranteed by this fitting rather than independently emergent.
  • Temporal-hybrid switching threshold tau = 3 infected per 10,000 (Fig. 5b-c); 2%, 5%, 10% of agents (Fig. 6d)
    Hand-chosen design parameter governing when the hybrid switches ABM to PBM. The reported ~99% runtime reduction and the extinction-capture result depend on this threshold; results at other thresholds differ by up to an order of magnitude in speedup (Fig. 6d).
axioms (5)
  • standard math Diffusion approximation validity of the SDE formulation, Eq. (5), following Allen et al. (2008)
    The SDE-SIRS influenza calibration in Fig. 4b assumes the sqrt(flow)-scaled noise correctly approximates the underlying continuous-time Markov chain at the population sizes used; the non-negativity post-processing is an unquantified correction.
  • standard math Erlang-distributed transition times via linear chain coupling of subcompartments, Eqs. (3) and (8)
    LCT models assume that a chain of n identical exponential subcompartments yields an Erlang approximation to the target stay-time distribution; in Setup S2 the Erlang is matched only to mean and variance of the lognormal.
  • standard math Exactness of temporal Gillespie sampling for inhomogeneous Poisson processes (Vestergaard & Genois) and the master-equation description, Eq. (6)
    The dABM, SMM, and stochastic mobility schemes rely on the Markov property and well-specified state-dependent rates; error is controlled only by the discretization of movement, not analyzed here.
  • ad hoc to paper dABM movement kernel is the single-well potential F(x,y) = (x^4 + y^4)/2
    The diffusion-based ABM results (Fig. 5a-c, Fig. 6d), including the hybrid comparison and extinction statistics, depend on this chosen potential; no sensitivity analysis over movement kernels is provided.
  • domain assumption Cross-model comparability holds when mean stay times are matched (S1) or mean/variance matched (S2) plus rho harmonized to the ABM
    The inter-model comparison in Fig. 5e assumes that matching these moments and one contact parameter makes ODE/LCT/IDE/ABM implementations of the 'same' SECIR process comparable; this is stated in SI Section F.2 and is the basis for ranking models in the ensemble.

pith-pipeline@v1.3.0-alltime-deepseek · 43138 in / 18911 out tokens · 197978 ms · 2026-08-03T00:14:02.446079+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of MEmilio -- A high performance Modular EpideMIcs simuLatIOn software for multi-scale and comparative simulations of infectious disease dynamics." pith.science (2026). https://pith.science/paper/ASFQKI3L

@misc{pith2026260211381,
  author       = {Pith},
  title        = {Pith review of: MEmilio -- A high performance Modular EpideMIcs simuLatIOn software for multi-scale and comparative simulations of infectious disease dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ASFQKI3L}},
  note         = {Machine review of arXiv:2602.11381}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Epidemic and pandemic preparedness with rapid outbreak response rely on timely, trustworthy evidence. Mathematical models are crucial for supporting timely and reliable evidence generation for public health decision-making with models spanning approaches from compartmental and metapopulation models to detailed agent-based simulations. Yet, the accompanying software ecosystem remains fragmented across model types, spatial resolutions, and computational targets, making models harder to compare, extend, and deploy at scale. Here we present MEmilio, a modular, high-performance framework for epidemic simulation that harmonizes the specification and execution of diverse dynamic epidemiological models within a unified and harmonized architecture. MEmilio couples an efficient C++ simulation core with coherent model descriptions and a user-friendly Python interface, enabling workflows that run on laptops as well as high-performance computing systems. Standardized representations of space, demography, and mobility support straightforward adaptations in resolution and population size, facilitating systematic inter-model comparisons and ensemble studies. The framework integrates readily with established tools for uncertainty quantification and parameter inference, supporting a broad range of applications from scenario exploration to calibration. Finally, strict software-engineering practices, including extensive unit and continuous integration testing, promote robustness and minimize the risk of errors as the framework evolves. By unifying implementations across modeling paradigms, MEmilio aims to lower barriers to reuse and generalize models, enable principled comparisons of implicit assumptions, and accelerate the development of novel approaches that strengthen modeling-based outbreak preparedness.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

51 extracted references · 10 canonical work pages · 1 internal anchor

  1. [89]

    & Mulansky, M

    Ahnert, K. & Mulansky, M. Odeint – Solving Ordinary Differential Equations in C++.AIP Conference Proceedings1389, 1586–1589 (2011). URL https://doi.org/10.1063/1.3637934

    work page doi:10.1063/1.3637934 2011

  2. [90]

    Cash, J. R. & Karp, A. H. A variable order Runge-Kutta method for initial value problems with rapidly varying right-hand sides.ACM Transactions on Mathematical Software16, 201–222 (1990). URL https://doi.org/10.1145/79505.79507

    arXiv 1990

  3. [91]

    & Vecchio, A

    Messina, E., Pezzella, M. & Vecchio, A. A non-standard numerical scheme for an age-of-infection epidemic model.Journal of Computational Dynamics9, 239 (2022). URL https://doi.org/10.3934/ jcd.2021029

    2022

  4. [92]

    & Kühn, M

    Wendler, A., Plötzke, L., Tritzschak, H. & Kühn, M. J. A nonstandard numerical scheme for a novel secir integro-differential equation-based model allowing nonexponentially distributed stay times. Applied Mathematics and Computation509, 129636 (2026). URL https://doi.org/10.1016/j.amc. 2025.129636

    arXiv 2026

  5. [93]

    URL https://www.mpi-forum.org/docs/mpi-5.0/mpi50-report.pdf

    Message Passing Interface Forum.MPI: A Message-Passing Interface Standard Version 5.0(2025). URL https://www.mpi-forum.org/docs/mpi-5.0/mpi50-report.pdf

    2025

  6. [94]

    & Menon, R

    Dagum, L. & Menon, R. OpenMP: an industry standard API for shared-memory programming. Computational Science & Engineering, IEEE5, 46–55 (1998). URL https://doi.org/10.1109/99. 660313

    work page doi:10.1109/99 1998

  7. [95]

    T.Practical methods for optimal control and estimation using nonlinear programming (SIAM, 2010)

    Betts, J. T.Practical methods for optimal control and estimation using nonlinear programming (SIAM, 2010). URL https://doi.org/10.1137/1.9780898718577.bm

    work page doi:10.1137/1.9780898718577.bm 2010

  8. [96]

    & Lawley, M

    Lin, F., Muthuraman, K. & Lawley, M. An optimal control theory approach to non- pharmaceutical interventions.BMC Infectious Diseases10, 32 (2010). URL https://doi.org/10. 1186/1471-2334-10-32

    2010

  9. [97]

    Tsay, C., Lejarza, F., Stadtherr, M. A. & Baldea, M. Modeling, state estimation, and optimal control for the us covid-19 outbreak.Scientific Reports10, 10711 (2020). URL https://doi.org/10.1038/ s41598-020-67459-8

    2020

  10. [98]

    & Mitsos, A

    Sass, S. & Mitsos, A. Obscured by terminology: Hidden parallels in direct methods for open-loop optimal control.Journal of Process Control154, 103513 (2025). URL https://doi.org/10.1016/j. jprocont.2025.103513

    arXiv 2025

  11. [99]

    & Biegler, L

    Wächter, A. & Biegler, L. T. On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear programming.Mathematical programming106, 25–57 (2006). URL https: //doi.org/10.1007/s10107-004-0559-y

    doi:10.1007/s10107-004-0559-y 2006

  12. [100]

    Software, Environments, and Tools (Society for Industrial and Applied Mathematics, 2011)

    Naumann, U.The Art of Differentiating Computer Programs. Software, Environments, and Tools (Society for Industrial and Applied Mathematics, 2011). URL https://doi.org/10.1137/1. 9781611972078

    doi:10.1137/1 2011

  13. [101]

    & Radev, S

    Arruda, J., Bracher, N., Köthe, U., Hasenauer, J. & Radev, S. T. Diffusion models in simulation- based inference: A tutorial review.arXiv(2025). URL https://doi.org/10.48550/arXiv.2512.20685

    Pith/arXiv arXiv doi:10.48550/arxiv.2512.20685 2025

  14. [102]

    T.et al.BayesFlow: Amortized Bayesian workflows with neural networks.Journal of Open Source Software8, 5702 (2023)

    Radev, S. T.et al.BayesFlow: Amortized Bayesian workflows with neural networks.Journal of Open Source Software8, 5702 (2023). URL https://doi.org/10.21105/joss.05702

    doi:10.21105/joss.05702 2023

  15. [103]

    & Hasenauer, J

    Schälte, Y., Klinger, E., Alamoudi, E. & Hasenauer, J. pyABC: Efficient and robust easy-to- use approximateBayesian computation.Journal of Open Source Software7, 4304 (2022). URL https://doi.org/10.21105/joss.04304

    doi:10.21105/joss.04304 2022

  16. [104]

    & Holenstein, R

    Andrieu, C., Doucet, A. & Holenstein, R. Particle Markov chain Monte Carlo methods.Journal of the Royal Statistical Society: Series B (Statistical Methodology)72, 269–342 (2010). URL https: //doi.org/10.1111/j.1467-9868.2009.00736.x

    arXiv 2010

  17. [105]

    & Papaspiliopoulos, O.An Introduction to sequential Monte Carlo

    Chopin, N. & Papaspiliopoulos, O.An Introduction to sequential Monte Carlo. Springer Series in Statistics (Springer International Publishing, 2020). URL https://doi.org/10.1007/ 978-3-030-47845-2

    2020

  18. [106]

    & Roberts, G

    Andrieu, C. & Roberts, G. O. The pseudo-marginal approach for efficient Monte Carlo computations.The Annals of Statistics37(2009). URL https://doi.org/10.1214/07-AOS574

    doi:10.1214/07-aos574 2009

  19. [107]

    Assessment of Simulation-based Inference Methods for Stochastic Compartmental Models in Epidemiological Research

    Wieland, V., Wassmuth, N., Contento, L., Kühn, M. & Hasenauer, J. Assessment of simulation- based inference methods for stochastic compartmental models.arXiv(2025). URL https://doi.org/ 10.48550/arXiv.2512.02528

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2512.02528 2025

  20. [108]

    URL https://doi.org/10.1016/j.epidem.2011.04.002

    Edlund, S.et al.Comparing three basic models for seasonal influenza.Epidemics3, 135–142 (2011). URL https://doi.org/10.1016/j.epidem.2011.04.002

    work page doi:10.1016/j.epidem.2011.04.002 2011

  21. [109]

    Buchholz, U.et al.GrippeWeb - Daten des Wochenberichts [Data set]

  22. [110]

    F., Boelle, P.-Y., Valleron, A.-J

    Legrand, J., Grais, R. F., Boelle, P.-Y., Valleron, A.-J. & Flahault, A. Understanding the dynamics of Ebola epidemics.Epidemiology & Infection135, 610–621 (2007). URL https://doi.org/10.1017/ s0950268806007217. 45

    2007

  23. [111]

    URL https://doi.org/10

    Rico, A.et al.Epidemiology of epidemic Ebola virus disease in Conakry and surrounding prefec- tures, Guinea, 2014–2015.Emerging infectious diseases22, 178 (2016). URL https://doi.org/10. 3201/eid2202.151304

    arXiv 2014

  24. [112]

    URL https://plotdigitizer.com

    Plotdigitizer: Version 3.1.6 (2026). URL https://plotdigitizer.com

    2026

  25. [113]

    Kühn, M. J.et al.Assessment of effective mitigation and prediction of the spread of SARS-CoV- 2 in Germany using demographic information and spatial resolution.Mathematical Biosciences 108648 (2021). URL https://doi.org/10.1016/j.mbs.2021.108648

    arXiv 2021

  26. [114]

    T., Mertens, U

    Radev, S. T., Mertens, U. K., Voss, A., Ardizzone, L. & Köthe, U. BayesFlow: Learning com- plex stochastic models with invertible neural networks.IEEE transactions on neural networks and learning systems33, 1452–1466 (2020). URL https://doi.org/10.1109/TNNLS.2020.3042395

    arXiv 2020

  27. [115]

    Pendlerverflechtungen der sozialversicherungspflichtig Beschäftigten nach Kreisen - Deutschland (Jahreszahlen) (2022)

    BMAS. Pendlerverflechtungen der sozialversicherungspflichtig Beschäftigten nach Kreisen - Deutschland (Jahreszahlen) (2022). URL https://statistik.arbeitsagentur.de/SiteGlobals/Forms/ Suche/Einzelheftsuche_Formular.html?topic_f=beschaeftigung-sozbe-krpendd. Publication Title: www.bmas.de

    2022

  28. [116]

    G., Lepri, B

    Beneduce, C., Muñoz-Repiso, T. G., Lepri, B. & Luca, M. pyspainmobility: a python package to access and manage spanish open mobility data (2025). URL https://doi.org/10.48550/arXiv.2506. 13385

    doi:10.48550/arxiv.2506 2025

  29. [117]

    Intensivkapazitäten und COVID-19-Intensivbettenbelegung in Deutschland (2025)

    Robert Koch-Institut. Intensivkapazitäten und COVID-19-Intensivbettenbelegung in Deutschland (2025). URL https://doi.org/10.5281/zenodo.17046865

    work page doi:10.5281/zenodo.17046865 2025

  30. [118]

    SARS-CoV-2 Infektionen in Deutschland (2025)

    Robert Koch-Institut. SARS-CoV-2 Infektionen in Deutschland (2025). URL https://doi.org/10. 5281/zenodo.17046972

    2025

  31. [119]

    Datos abiertos de capacidad asistencial (2023)

    Ministerio de Sanidad. Datos abiertos de capacidad asistencial (2023). URL https: //www.sanidad.gob.es/areas/alertasEmergenciasSanitarias/alertasActuales/nCov/documentos/ Datos_Capacidad_Asistencial_Historico_14072023.csv

    2023

  32. [120]

    Evolución de enfermedad por el coronavirus (covid-19) (2023)

    Ministerio de Sanidad. Evolución de enfermedad por el coronavirus (covid-19) (2023). URL https: //datos.gob.es/en/catalogo/e05070101-evolucion-de-enfermedad-por-el-coronavirus-covid-19

    2023

  33. [121]

    Prem, K., Cook, A. R. & Jit, M. Projecting social contact matrices in 152 countries using contact surveys and demographic data.PLoS computational biology13, e1005697 (2017). URL https: //doi.org/10.1371/journal.pcbi.1005697

    doi:10.1371/journal.pcbi.1005697 2017

  34. [122]

    & Schmon, S

    Ward, D., Cannon, P., Beaumont, M., Fasiolo, M. & Schmon, S. Robust neural posterior estimation and statistical model criticism. In Koyejo, S.et al.(eds.)Advances in Neural Information Processing Systems, vol. 35, 33845–33859 (Curran Associates, Inc., 2022). URL https://proceedings.neurips. cc/paper_files/paper/2022/file/db0eac6747e3631eb91095cd76065611-P...

    2022

  35. [123]

    C.et al.Covasim: An agent-based model of COVID-19 dynamics and interventions.PLOS Computational Biology17, 1–32 (2021)

    Kerr, C. C.et al.Covasim: An agent-based model of COVID-19 dynamics and interventions.PLOS Computational Biology17, 1–32 (2021). URL https://doi.org/10.1371/journal.pcbi.1009149. Model repository: https://github.com/InstituteforDiseaseModeling/covasim

    doi:10.1371/journal.pcbi.1009149 2021

  36. [124]

    & Merler, S

    Fumanelli, L., Ajelli, M., Manfredi, P., Vespignani, A. & Merler, S. Inferring the Structure of Social ContactsfromDemographicDataintheAnalysisofInfectiousDiseasesSpread.PLoS Computational Biology8, e1002673 (2012). URL https://doi.org/10.1371/journal.pcbi.1002673

    doi:10.1371/journal.pcbi.1002673 2012

  37. [125]

    & Kühn, M

    Schmidt, A., Zunker, H., Heinlein, A. & Kühn, M. J. Graph neural network surrogates to leverage mechanistic expert knowledge towards reliable and immediate pandemic response.Scientific Reports (2026). URL https://doi.org/10.1038/s41598-026-39431-5

    work page doi:10.1038/s41598-026-39431-5 2026

  38. [126]

    URL https://www.tensorflow.org/

    Abadi, M.et al.TensorFlow: Large-scale machine learning on heterogeneous systems (2015). URL https://www.tensorflow.org/. Software available from tensorflow.org

    2015

  39. [127]

    & Alippi, C

    Grattarola, D. & Alippi, C. Graph Neural Networks in TensorFlow and Keras with Spektral [Application Notes].IEEE Computational Intelligence Magazine16, 99–106 (2021). URL https: //doi.org/10.1109/MCI.2020.3039072

    arXiv 2021

  40. [128]

    M., Grattarola, D., Livi, L

    Bianchi, F. M., Grattarola, D., Livi, L. & Alippi, C. Graph Neural Networks with Convolutional ARMA Filters.IEEE Transactions on Pattern Analysis and Machine IntelligencePP, 1–1 (2021). URL https://doi.org/10.1109/TPAMI.2021.3054830

    arXiv 2021

  41. [129]

    Pendlerverflechtungen der sozialversicherungspflichtig Beschäftigten nach Kreisen - Deutschland (Jahreszahlen) (2020)

    BMAS. Pendlerverflechtungen der sozialversicherungspflichtig Beschäftigten nach Kreisen - Deutschland (Jahreszahlen) (2020). URL https://statistik.arbeitsagentur.de/SiteGlobals/Forms/ Suche/Einzelheftsuche_Formular.html?topic_f=beschaeftigung-sozbe-krpendd. Publication Title: www.bmas.de

    2020

  42. [130]

    C.et al.Estimating infectiousness throughout SARS-CoV-2 infection course.Science 373, eabi5273 (2021)

    Jones, T. C.et al.Estimating infectiousness throughout SARS-CoV-2 infection course.Science 373, eabi5273 (2021). URL https://doi.org/10.1126/science.abi5273

    doi:10.1126/science.abi5273 2021

  43. [131]

    URLhttps://10.1038/s41564-022-01105-z

    Ke, R.et al.Daily longitudinal sampling of SARS-CoV-2 infection reveals substantial heterogeneity ininfectiousness.Nature Microbiology7,640–652(2022). URLhttps://10.1038/s41564-022-01105-z

    work page doi:10.1038/s41564-022-01105-z 2022

  44. [132]

    DESTATIS Zeitverwendungserhebung (ZVE) 2022. Available online: https://www.destatis.de/ DE/Themen/Gesellschaft-Umwelt/Einkommen-Konsum-Lebensbedingungen/Zeitverwendung/ 46 Tabellen/aktivitaeten-insgesamt-zve.html, accessed 2025-08-22

    2022

  45. [133]

    URL https://doi.org/10.1136/bmj.n1412

    Nyberg, T.et al.Risk of hospital admission for patients with SARS-CoV-2 variant b.1.1.7: cohort analysis.BMJ373, n1412 (2021). URL https://doi.org/10.1136/bmj.n1412

    work page doi:10.1136/bmj.n1412 2021

  46. [134]

    URL https: //doi.org/10.1038/s41598-022-23312-8

    Zali, A.et al.Mortality among hospitalized COVID-19 patients during surges of SARS-CoV-2 alpha (b.1.1.7) and delta (b.1.617.2) variants.Scientific Reports12, 18918 (2022). URL https: //doi.org/10.1038/s41598-022-23312-8

    work page doi:10.1038/s41598-022-23312-8 2022

  47. [135]

    Jülich Supercomputing Centre. JURECA: Data Centric and Booster Modules implementing the Modular Supercomputing Architecture at Jülich Supercomputing Centre.Journal of large-scale research facilities7, A182 (2021). URL https://doi.org/10.17815/jlsrf-7-182

    doi:10.17815/jlsrf-7-182 2021

  48. [136]

    Available online: https:// www-genesis.destatis.de/datenbank/online/table/12411-0005/search/s/MTI0MTEtMDAwNQ==, accessed 2025-08-22

    DESTATIS Bevölkerung: Deutschland, Stichtag, Altersjahre. Available online: https:// www-genesis.destatis.de/datenbank/online/table/12411-0005/search/s/MTI0MTEtMDAwNQ==, accessed 2025-08-22

    2025

  49. [137]

    DESTATIS Erwerbstätigenquoten 1991 bis 2024. Available online: https: //www.destatis.de/DE/Themen/Arbeit/Arbeitsmarkt/Erwerbstaetigkeit/Tabellen/ erwerbstaetigenquoten-gebietsstand-geschlecht-altergruppe-mikrozensus.html, accessed 2025-08-25

    1991

  50. [138]

    Available online: https://www-genesis.destatis.de/datenbank/ online/table/12421-0100/search/s/MTI0MjEtMDEwMA==, accessed 2025-08-25

    DESTATIS Vorausberechnete Privathaushalte: Deutschland, Jahre, Varianten der Haushaltsvo- rausberechnung, Haushaltsgröße. Available online: https://www-genesis.destatis.de/datenbank/ online/table/12421-0100/search/s/MTI0MjEtMDEwMA==, accessed 2025-08-25

    2025

  51. [139]

    Avail- able online: https://www-genesis.destatis.de/datenbank/online/table/21111-0002/search/s/ MjExMTEtMDAwMg==, accessed 2025-08-25

    DESTATIS Schüler: Deutschland, Schuljahr, Geschlecht, Schulart, Jahrgangsstufen. Avail- able online: https://www-genesis.destatis.de/datenbank/online/table/21111-0002/search/s/ MjExMTEtMDAwMg==, accessed 2025-08-25. 47

    2025