Simulation-trained neural networks boost forecasting accuracy across scientific fields

Researchers introduced Simulation-Grounded Neural Networks, which learn from mechanistic simulations and outperform traditional data-driven and physics-constrained models in tasks ranging from COVID-19 mortality to ecological forecasting.

A team of scientists unveiled Simulation-Grounded Neural Networks (SGNNs), a method that incorporates mechanistic theory by training neural networks on large corpora of synthetic simulations covering multiple model structures and realistic noise. By pretraining on these datasets, SGNNs develop a structural prior that captures underlying dynamics, enabling superior performance on forecasting problems compared with pure data-driven models and physics-constrained hybrids.

Evaluations across disciplines—including epidemiology, ecology, social science and chemistry—demonstrated that SGNNs delivered markedly higher predictive skill, notably tripling the accuracy of average CDC forecasts for COVID-19 mortality and handling high-dimensional ecological systems effectively. The approach also proved resilient to model misspecification, maintaining strong results even when trained on flawed assumptions.

Additionally, the researchers introduced a back-to-simulation attribution method that provides mechanistic interpretability by matching real-world dynamics to their closest simulated counterparts. Funding for the work came from an Insight Net cooperative agreement with the University of Michigan and a CDC Center for Forecasting and Outbreak Analytics grant.