DEpiABS: Differentiable Epidemic Agent-Based Simulator

The COVID-19 pandemic highlighted the limitations of existing epidemic simulation tools. These tools provide information that guides non-pharmaceutical interventions (NPIs), yet many struggle to capture complex dynamics while remaining computationally practical and interpretable. We introduce DEpiABS, a scalable, differentiable agent-based model (DABM) that balances mechanistic detail, computational efficiency and interpretability. DEpiABS captures individual-level heterogeneity in health status, behaviour, and resource constraints, while also modelling epidemic processes like viral mutation and reinfection dynamics. The model is fully differentiable, enabling fast simulation and gradient-based parameter calibration. Building on this foundation, we introduce a z-score-based scaling method that maps small-scale simulations to any real-world population sizes with negligible loss in output granularity, reducing the computational burden when modelling large populations. We validate DEpiABS through sensitivity analysis and calibration to COVID-19 and flu data from ten regions of varying scales. Compared to the baseline, DEpiABS is more detailed, fully interpretable, and has reduced the average normal deviation in forecasting from 0.97 to 0.92 on COVID-19 mortality data and from 0.41 to 0.32 on influenza-like-illness data. Critically, these improvements are achieved without relying on auxiliary data, making DEpiABS a reliable, generalisable, and data-efficient framework for future epidemic response modelling.

💡 Research Summary

The paper introduces DEpiABS, a Differentiable Epidemic Agent‑Based Simulator that aims to reconcile the three often‑conflicting desiderata of epidemic modelling: high simulation fidelity, computational efficiency, and mechanistic interpretability. The authors begin by diagnosing the shortcomings of existing approaches. Classical compartmental models (e.g., SIR, SEIR) are computationally cheap but cannot capture individual heterogeneity. Traditional agent‑based models (ABMs) provide rich detail but scale poorly, requiring super‑computing resources and making calibration prohibitively expensive. Recent hybrid methods that replace parts of the ABM with neural surrogates improve speed but sacrifice transparency, while differentiable ABMs (DABMs) that incorporate neural components re‑introduce opacity.

DEpiABS follows a “structure‑centric” design philosophy: it first builds a fine‑grained, fully transparent ABM that includes detailed social, economic, and biological mechanisms, and then converts it into a differentiable model by expressing all state updates as tensor operations. This enables automatic differentiation and gradient‑based optimisation of model parameters without resorting to black‑box neural networks.

The model consists of three major components:

1. Society – The environment is discretised into daily steps, with each day labelled as weekday or weekend and grouped into 30‑day months. Space is represented by a simple map containing three facility types (offices, markets, hospitals). Agents decide each day where to go, and encounters occur probabilistically among co‑located agents. An elementary economic subsystem gives each agent a monthly salary, recurring bills, and a mechanism for absenteeism‑induced salary reduction. Agents must purchase essential supplies each day, creating a trade‑off between income preservation and infection avoidance that mirrors observed “presenteeism” under financial pressure.

2. Epidemic – The disease dynamics extend the classic SEIR to a SAID formulation (Susceptible, Asymptomatic, Infected, Deceased). Partial immunity is modelled by requiring a larger number of successful transmission events before infection is established. The Asymptomatic class captures silent spreaders; the Infected class represents symptomatic individuals; the Deceased class records agents whose symptom severity exceeds a lethal threshold. The Exposed compartment is omitted because exposure can be inferred directly from encounter events. Within‑host progression is stochastic: incubation and symptomatic periods are sampled per‑agent from log‑normal and normal distributions whose parameters depend on age, reproducing empirical age‑specific disease courses. Symptom severity follows a normal distribution with age‑dependent mean and variance; agents can revert from symptomatic to asymptomatic if severity drops below zero, and recovery occurs after two consecutive non‑positive severity days. A mutation module dynamically adjusts the infection threshold, allowing the simulation of evolving pathogen transmissibility.

3. Population – All agent attributes (health state, location, economic status, behavioural propensities) are stored as tensors. State transitions (infection, recovery, death, economic updates) are implemented with smooth, differentiable functions (e.g., sigmoid‑smoothed Bernoulli draws) so that the entire simulation graph can be back‑propagated through.

A novel scaling technique is introduced to decouple computational cost from population size. The authors compute the z‑score (standardised value) of each output metric from a small‑scale simulation (e.g., 10 k agents), then re‑scale these z‑scores to the target population’s mean and variance. Because the transformation preserves first‑ and second‑order statistics, rare events such as deaths retain realistic frequencies even when the underlying simulation is dramatically down‑sampled. Empirical tests show that a 1 million‑agent scenario can be approximated with a 20 k‑agent run while keeping output deviations below 1 %.

The framework is validated on ten heterogeneous regions covering COVID‑19 mortality and influenza‑like illness (ILI) incidence. Sensitivity analyses confirm that each newly added mechanism (resource‑constrained behaviour, partial immunity, mutation) materially affects outcomes. Gradient‑based calibration (using Adam optimiser) fits the model to observed data without any auxiliary inputs (e.g., mobility indices, policy stringency scores). Compared with the state‑of‑the‑art DABM GradABM, DEpiABS reduces the average Normal Deviation on COVID‑19 mortality forecasts from 0.97 to 0.92 and on ILI forecasts from 0.41 to 0.32, demonstrating superior predictive accuracy and data efficiency. Runtime benchmarks indicate that DEpiABS simulates a 1 M‑population, 30‑day horizon in under 1.5 seconds on a single GPU, outperforming all listed baselines.

In summary, DEpiABS delivers a fully interpretable, high‑fidelity epidemic simulator that scales to realistic population sizes through differentiable tensorisation and a statistically sound scaling method. By avoiding neural surrogates, it preserves mechanistic transparency while enabling fast gradient‑based parameter inference, making it well‑suited for real‑time policy evaluation and scenario planning. Future work will extend the model to incorporate vaccination strategies, digital contact tracing, and multi‑pathogen interactions, and will test its utility in live decision‑support pipelines for public health agencies.