Stochastic reaction-diffusion simulations meet deep learning.
deepRD is a Python package for simulating molecular dynamics and stochastic reaction-diffusion processes, and for learning reduced models of them with PyTorch. Detailed molecular or particle simulations produce training data; deep generative models learn the unresolved noise and interaction terms; and reduced Langevin integrators sample those terms on the fly to reproduce the full dynamics at a fraction of the cost (stochastic closure).
The core of deepRD's data-driven modeling lives in deepRD.noiseSampler:
cvaeSampler: a conditional variational autoencoder that learns the distribution of the noise term conditioned on the particle state (e.g. velocity and previous noise values), for bistable and dimer systems.diffusionSampler: a conditional diffusion model with adaptive layer norm (AdaLN) blocks, supporting several conditioning schemes on velocities, relative positions and noise history.binning: a non-parametric baseline that samples from binned trajectory data.
Training code is in deepRD/noiseSampler/training/. Models run on GPU automatically when CUDA is available. Trained samplers plug directly into the data-driven integrators langevinNoiseSampler and langevinInteractionSampler.
- Diffusion integrators: Langevin (BAOAB, ABOBA, semi-implicit Euler), overdamped Langevin, and Smoluchowski dynamics with a reactive boundary and particle reservoir
- Reaction integrators: Gillespie SSA and tau-leaping
- Reaction models: genetic feedback, Schlögl, reservoir
- Potentials: harmonic, bistable, Gaussian, and pair bistable (dimer) potentials
- Tools: trajectory, particle and analysis utilities, plus plotting helpers
git clone https://github.com/maojrs/deepRD.git
cd deepRD
pip install -e .
pip install numpy scipy matplotlib joblib torch jupyterimport deepRD
from deepRD.diffusionIntegrators import langevin
particle = deepRD.particle([0, 0, 0], velocity=[0, 0, 0], mass=1)
particleList = deepRD.particleList([particle])
# dt, stride, tfinal, Gamma (friction), kBT, boxsize
integrator = langevin(0.01, 1, 10, 10.0, 1, 1)
t, X, V = integrator.propagate(particleList)from deepRD.reactionModels import geneticFeedback
from deepRD.reactionIntegrators import gillespie
model = geneticFeedback(G=1, Gstar=0, M=50, P=300)
t, X = gillespie(stride=1, tfinal=10).propagate(model)deepRD/
├── deepRD/
│ ├── noiseSampler/ # PyTorch CVAE and diffusion samplers, binning, training code
│ ├── diffusionIntegrators/ # Langevin, Smoluchowski and data-driven integrators
│ ├── reactionIntegrators/ # Gillespie and tau-leap
│ ├── reactionModels/ # Genetic feedback, Schlögl, reservoir
│ ├── potentials/ # External and pair potentials
│ └── tools/ # Trajectory, particle and analysis utilities
├── notebooks/ # Examples and analyses
└── scripts/ # Data generation and benchmarks
The notebooks/ folder contains worked examples:
stochasticClosureCVAE/: benchmarks of reduced models driven by the deep generative samplersstochasticClosure/: the binning-based closure pipeline, from building the sampling model to comparing reduced and reference dynamicsopenSystems/: Smoluchowski dynamics coupled to a particle reservoirlangevin.ipynb,geneticFeedbackModel.ipynband others for the basic integrators
Matching batch scripts for data generation and benchmarking, including first-passage-time analyses, are in scripts/.