© 2025 The Regents of the University of Michigan
Carson Dudley — University of Michigan

---

Overview

Simulation-Grounded Neural Networks (SGNNs) are a general framework for training neural networks on mechanistic simulations to perform forecasting, inference, regression, and classification in scientific domains. SGNNs enable robust, interpretable, and data-efficient learning even in noisy, low-resource, or unobservable settings. The approach has been validated across epidemiology, chemistry, ecology, and social science.

SGNNs unify mechanistic reasoning and deep learning by treating simulations as flexible supervisory signals—rather than fixed priors or surrogates—enabling both zero-shot generalization and mechanistic interpretability via back-to-simulation attribution.

🚀 Try It Instantly

👉 Run the Colab Tutorial

No installation or coding required.

---

Repository Contents

File / Directory	Description
sgnn_tutorial.ipynb	End-to-end SGNN tutorial: stochastic multi-mechanism simulators (SIR, SEIR, SEAIR) + observation model + single-pass training and evaluation
chem_yield_zeroshot.ipynb	SGNN evaluation on chemical reaction yield prediction (zero-shot mode)
deaths.csv	Example real-world mortality dataset (COVID-19, for forecasting task)
dengue_br.csv	Dengue case data for out-of-domain generalization benchmarking
einn_evals.ipynb	PINN/EINN baseline evaluations for disease forecasting
generation_flu_death_hosp.ipynb	Simulation script for multi-wave infectious disease trajectories
generation_v5_human.ipynb	Unified disease simulator with realistic observational effects
sgnn_forecaster.ipynb	Script for training SGNN disease forecaster
hybrid_chem_model_best.ipynb	SGNN pretraining + fine-tuning for chem yield prediction
r0_est.py	SGNN-based estimation of $R_0$ from early outbreak curves
source_id.ipynb	SGNN training and evaluation for source inference in diffusion cascades
README.md	This file
LICENSE.txt	Licensing information (see below)
NOTICES.txt	Third-party notices and copyright
FigureCode.R	Code for generating schematic and main figures in manuscript
Appendix.R	Code for generating figure in appendix

---

Getting Started

Installation

This repo requires Python 3.10+ and PyTorch 2.0+ for model implementation. The repo also requires R 4.5.1 for MacOS.

Recommended starting point for new users:

sgnn_tutorial.ipynb — End-to-end SGNN workflow tutorial:

• Stochastic SIR, SEIR, and SEAIR simulators with day-by-day Poisson sampling.
• One-pass synthetic pretraining: generate enough data for a single epoch instead of looping over fixed samples.
• Observation model: under-reporting, reporting delays, and overdispersion for realistic training data.
• Lightweight CNN forecaster trained on case data.
• Validation on held-out synthetic data with forecast plots.

Running the Code

Clone the repo and run any of the notebooks (.ipynb) in Jupyter or VS Code. For example:

• Run generation_v5_human.ipynb to generate synthetic infectious disease data.
• Use source_id.ipynb to evaluate SGNNs on cascade source inference.
• Run hybrid_chem_model.ipynb to evaluate SGNN performance on chemical reactions.
• For reproduction number inference from early outbreak curves, run: python r0_est.py

Each notebook is self-contained and includes example runs on provided datasets.

---

Documentation

The SGNN framework, simulation engines, architectures, and evaluation protocols are fully documented in our paper:

Simulation as Supervision: Mechanistic Pretraining for Scientific Discovery
Carson Dudley, Reiden Magdaleno, Christopher Harding, Marisa Eisenberg (2025)
[Preprint]

---

References & Related Work

• Dudley et al., Simulation as Supervision: Mechanistic Pretraining for Scientific Discovery (2025) — (https://arxiv.org/abs/2507.08977)
• Dudley et al., Mantis: A simulation-Grounded Foundation Model for Disease Forecasting (2025) - (https://arxiv.org/abs/2508.12260)
• PINNs: Raissi et al. (2019), Journal of Computational Physics
• DEFSI: Wang et al. (2019), AAAI
• PFNs: Müller et al. (2022), ICLR

---

Contact

Carson Dudley
PhD Student University of Michigan
📧 cdud@umich.edu
🌐 carsondudley1.github.io

---

License

See LICENSE.txt.
© 2025 The Regents of the University of Michigan
Carson Dudley — University of Michigan

---

Notices

See NOTICES.txt for third-party software attributions and licenses.

---

Community & Roadmap

Community-building and roadmap planning are in progress. Future directions for SGNNs include:

• Domain-specific finetuning pipelines
• Active learning for simulation generation
• Real-time deployment interfaces
• Expanded simulator libraries across scientific fields
• Counterfactual/scenario predictions with SGNNs
• SGNNs + RL for optimal policies

If you're interested in collaborating or contributing to the SGNN ecosystem, please get in touch or watch this repository for updates.