A Hybrid ABM-PDE Framework for Real-World Infectious Disease Simulations
This paper introduces a computationally efficient hybrid ABM-PDE framework that couples agent-based and partial differential equation models to simulate infectious disease spread in the Berlin-Brandenburg region, achieving faster runtimes and comparable accuracy to full-scale agent-based simulations while maintaining consistency at the model interface.
Original authors:Kristina Kehrer, Tim O. F. Conrad
Original authors: Kristina Kehrer, Tim O. F. Conrad
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
Imagine you are trying to predict how a rumor (or a virus) spreads through a massive city and its surrounding countryside. You have two main ways to do this:
The "Crowd" Method (Agent-Based Model): You create a digital twin for every single person. You track exactly where they go, who they bump into, and what they do. It's incredibly accurate, but if you have millions of people, your computer has to do millions of calculations every second. It's like trying to count every grain of sand on a beach by picking them up one by one.
The "Cloud" Method (PDE Model): Instead of tracking individuals, you look at the "fog" of people. You treat the population like a fluid or a gas spreading across a map. It's much faster to calculate, but it loses the details of individual behavior. It's like looking at the beach from a helicopter and just estimating the total volume of sand.
The Problem: The authors of this paper looked at the Berlin-Brandenburg region in Germany. Berlin is a dense, crowded city, while Brandenburg is a sprawling, rural area.
If you use the "Crowd" method for the whole region, your computer might crash or take too long to give you an answer.
If you use the "Cloud" method for the whole region, you might miss important details about how people actually move in the city.
The Solution: A Hybrid Framework The authors built a hybrid model that uses the best of both worlds. Think of it as a smart traffic system that switches between a detailed GPS for individual cars and a general flow map for the highway.
Berlin (The City): They used the "Cloud" method (PDE). Since the city is so dense and people move in predictable patterns, treating the population as a spreading fluid is fast and accurate enough.
Brandenburg (The Countryside): They used the "Crowd" method (ABM). Here, people are more spread out, and their specific movements matter more. They used real mobile phone data to create digital agents that actually walk, drive, and visit specific places like homes, schools, and shops.
How They Talk to Each Other (The Coupling) The magic happens at the border between the city and the countryside.
City to Country: When a digital person walks out of the "Cloud" (Berlin) and into the "Crowd" (Brandenburg), the system takes them out of the fluid map and turns them back into a specific, individual agent with a real schedule.
Country to City: When a specific agent walks from the countryside into the city, the system removes them from the individual list and adds their "weight" back into the fluid map.
This ensures that the total number of people stays correct and that the infection spreads naturally across the whole region without the computer getting overwhelmed.
What They Found The researchers tested this system using real data from the 2020 pandemic in Germany.
Speed: The hybrid model was significantly faster than trying to simulate every single person in the entire region. It allowed them to run many more simulations to get stable results.
Accuracy: Despite being faster, it was just as accurate as the slow, full "Crowd" model. In fact, for short timeframes, it was even slightly more accurate.
Realism: By using real mobile phone data for the countryside part, the model captured realistic movement patterns, like people commuting to work or going to the grocery store.
The Bottom Line This paper proves that you don't always have to choose between "super detailed but slow" and "fast but vague." By splitting the problem into a city (fluid) and a countryside (individuals) and letting them swap people at the border, you can get a fast, accurate simulation of how diseases spread in the real world. This helps scientists run scenarios quickly to understand how infections might move, without needing a supercomputer for every single test.
Technical Summary: A Hybrid ABM-PDE Framework for Real-World Infectious Disease Simulations
Problem Statement Simulating the spatial spread of infectious diseases presents a fundamental trade-off between computational efficiency and modeling fidelity. Agent-Based Models (ABMs) offer high-resolution individual dynamics and spatial heterogeneity but are computationally expensive, particularly for large populations. Conversely, Partial Differential Equation (PDE) models provide efficient spatial representations but often lack the granularity to capture individual behaviors and specific mobility patterns. Real-world regions, such as the Berlin-Brandenburg area in Germany, exhibit mixed characteristics—dense urban centers alongside rural, heterogeneous landscapes—making a single modeling paradigm insufficient. The challenge is to develop a framework that leverages the strengths of both approaches while maintaining consistency across the interface and reducing the overall computational runtime required for stable epidemiological results.
Methodology The authors propose a hybrid modeling framework that couples an ABM with a PDE model, partitioning the simulation domain into non-overlapping subregions based on local dynamics:
PDE Component (Berlin): The urban area of Berlin is modeled using a stochastic reaction-diffusion PDE system derived from a motion-based ABM. This component captures population-level dynamics and health state transitions across seven compartments: Susceptible (S), Exposed (E), Infectious (I), Symptomatic ($SY$), Hospitalized ($H, HC$), Critical (C), and Recovered (R). The model incorporates a landscape potential derived from agent movement data to guide population density.
ABM Component (Brandenburg): The surrounding rural area is modeled using a high-resolution ABM driven by real-world mobile phone trajectory data. Agents move between facilities (home, work, school, leisure) based on time-stamped events. Infection transmission occurs only within facilities where susceptible and infectious agents coexist; commuting is treated as isolated.
Coupling Mechanism: The two domains are dynamically coupled at each time step.
ABM to PDE: When an agent crosses the boundary into the PDE domain, they are removed from the ABM, and their health state is converted into a density contribution for the corresponding PDE compartment.
PDE to ABM: When density leaves the PDE domain, new agents are instantiated in the ABM. Their health states are deterministically assigned based on the current PDE compartment density ratios, and their trajectories are drawn from pre-existing mobile phone data plans.
Numerical Stability: To prevent instabilities at the boundary, density adjustments are distributed proportionally over the domain rather than applied to single grid nodes. Zero Neumann boundary conditions are imposed to avoid artificial fluxes.
Data and Implementation: The model utilizes mobile phone data representing 25% and 100% of the population in the Berlin-Brandenburg region. Behavioral changes (e.g., lockdowns) are simulated via activity change rates that modulate infection probabilities. The PDE is solved using the finite element method (Kaskade7 software) on a triangular mesh, while the ABM is a streamlined C++ implementation of EpiSim.
Key Contributions
Hybrid Architecture: The paper introduces a specific coupling mechanism for infectious disease modeling that allows for the seamless exchange of individuals between agent-based and density-based representations, preserving total population mass and health state distributions.
Real-World Validation: Unlike previous hybrid applications limited to cellular biology, this framework is applied to human-to-human disease transmission using real-world mobility and infection data from the Robert Koch Institute (RKI) for the Berlin-Brandenburg region.
Parameter Identification Strategy: The authors detail a method for fitting infection rates (βconst) separately for the PDE and ABM components, accounting for the non-linear scaling of pairwise interactions in the PDE versus the facility-based contact intensity in the ABM.
Scalability Analysis: The study evaluates the framework across two population scales (25% and 100% samples), demonstrating that parameters fitted on smaller samples can be effectively transferred to larger populations with appropriate scaling.
Results
Computational Efficiency: The hybrid model significantly reduces the overall simulation runtime compared to a full-ABM. This reduction is defined as the product of the number of runs required for stable results and the duration of a single run.
Accuracy: The hybrid model achieves smaller errors compared to the full-ABM reference across both 25% and 100% population samples.
Short-Term Dynamics: The hybrid approach demonstrated superior accuracy over short simulation horizons (approximately two months) when fitting was performed for separate time intervals (pre- and post-school closures).
Parameter Fitting: A grid search identified optimal infection rate parameters (βconst) for different time intervals. For the 25% sample, the optimal parameters were 4.5×102 (March 2–15) and 1.6×102 (March 16–April 28). The full-ABM required different scaling factors for its calibration parameter.
Significance and Claims The authors claim that the proposed ABM-PDE framework provides a robust and computationally efficient alternative to full-scale agent-based simulations. By reducing the computational burden, the framework enables faster parameter fitting and makes large-scale, realistic epidemic modeling and scenario analysis more feasible. The paper emphasizes that while the hybrid model sacrifices some individual-level tracking (e.g., agent IDs are lost when entering the PDE domain), it successfully captures core epidemiological dynamics. The work serves as a proof of concept for coupling different modeling paradigms to address the structural heterogeneity of real-world regions, with the long-term objective of extending this to a comprehensive ABM-PDE-ODE coupled model. The implementation and data are made available to support further research in this domain.