reactor-twin 0.4.0

Physics-constrained Neural DEs for chemical reactor digital twins

ReactorTwin

Physics-Constrained Neural Differential Equations for Chemical Reactor Digital Twins

---

Overview

ReactorTwin is a framework for building digital twins of chemical reactors using Neural Differential Equations with hard physics constraints. It combines:

• 5 Neural DE variants: Standard, Latent, Augmented, SDE, CDE
• 4 reactor types: CSTR, Batch, Semi-batch, PFR (plug flow with Method of Lines)
• 7 hard conservation laws: Mass, energy, thermodynamics, stoichiometry, port-Hamiltonian, GENERIC, positivity
• 6 kinetics models: Arrhenius, Michaelis-Menten, Power Law, Langmuir-Hinshelwood, Reversible, Monod
• Digital twin features: EKF state estimation, 4-level fault detection, MPC control, online adaptation, meta-learning
• 10-page Streamlit dashboard for interactive simulation and analysis

Key differentiator: Architectural projection onto constraint manifolds ensures physical laws are satisfied exactly, not approximately.

---

Quick Start

Installation

# From PyPI
pip install reactor-twin

# With all optional dependencies
pip install reactor-twin[all]

# Or from source
git clone https://github.com/ktubhyam/reactor-twin.git
cd reactor-twin
pip install -e ".[dev]"

30-Second Example

from reactor_twin import CSTRReactor, ArrheniusKinetics, NeuralODE
import torch

# Define CSTR with A -> B kinetics
reactor = CSTRReactor(
    name="my_cstr",
    num_species=2,
    params={"V": 100, "F": 10, "C_feed": [1.0, 0.0], "T_feed": 350},
    kinetics=ArrheniusKinetics(
        name="A_to_B",
        num_reactions=1,
        params={"k0": [1e10], "Ea": [50000], "stoich": [[-1, 1]]},
    ),
)

# Train Neural ODE to learn dynamics
model = NeuralODE(state_dim=2, solver="dopri5", adjoint=True)
predictions = model(z0=torch.randn(32, 2), t_span=torch.linspace(0, 10, 50))

# Predictions are 1000x faster than scipy integration!

Run the full quickstart: python examples/00_quickstart.py

---

Features

Neural DE Variants

Variant	Use Case	Status
Neural ODE	Standard continuous-time dynamics	✅ Complete
Latent Neural ODE	High-dimensional systems (encoder/decoder)	✅ Complete
Augmented Neural ODE	Topology-breaking expressivity	✅ Complete
Neural SDE	Uncertainty quantification (stochastic)	✅ Complete
Neural CDE	Irregular sensor data	✅ Complete

Reactor Types

Reactor	Description	Status
CSTR	Continuous stirred-tank reactor	✅ Complete
Batch	Time-varying volume for gas-phase reactions	✅ Complete
Semi-batch	Continuous feed + batch (pharmaceutical)	✅ Complete
PFR	Plug flow with Method of Lines discretization	✅ Complete

Kinetics Models

Model	Use Case	Status
Arrhenius	Standard temperature-dependent reactions	✅ Complete
Michaelis-Menten	Enzyme-catalyzed reactions	✅ Complete
Power Law	Empirical rate expressions	✅ Complete
Langmuir-Hinshelwood	Heterogeneous catalysis	✅ Complete
Reversible	Equilibrium-limited reactions	✅ Complete
Monod	Microbial growth kinetics	✅ Complete

Physics Constraints

Constraint	Hard Mode	Soft Mode	Status
Positivity	Softplus/ReLU projection	Penalty	✅ Complete
Mass Balance	Stoichiometric projection	Penalty	✅ Complete
Energy Balance	Soft mode only	Penalty	✅ Complete
Thermodynamics	Entropy/Gibbs/equilibrium	Penalty	✅ Complete
Stoichiometry	Predict rates not species	Penalty	✅ Complete
Port-Hamiltonian	Structure-preserving	N/A	✅ Complete
GENERIC	Reversible-irreversible coupling	Penalty	✅ Complete

Digital Twin Features

Feature	Description	Status
State Estimation	EKF with Neural ODE fusion, autograd Jacobians	✅ Complete
Fault Detection	4-level: SPC, residual, isolation, ML classification	✅ Complete
MPC Control	Gradient-based with constraint handling	✅ Complete
Online Adaptation	Replay buffer + Elastic Weight Consolidation	✅ Complete
Meta-Learning	Reptile for cross-reactor transfer	✅ Complete

Benchmark Systems

System	Type	Status
Exothermic A→B	Non-isothermal CSTR	✅ Complete
Van de Vusse	Series-parallel CSTR	✅ Complete
Bioreactor	Monod growth kinetics	✅ Complete
Consecutive A→B→C	Selectivity optimization	✅ Complete
Parallel A→B, A→C	Yield optimization	✅ Complete

---

Architecture

Plugin-Based Design

ReactorTwin uses a registry system for extensibility without modifying library code:

from reactor_twin import REACTOR_REGISTRY, AbstractReactor

# Register custom reactor
@REACTOR_REGISTRY.register("my_reactor")
class MyCustomReactor(AbstractReactor):
    def ode_rhs(self, t, y, u):
        # Your reactor equations
        return dydt

# Use anywhere
reactor = REACTOR_REGISTRY.get("my_reactor")(...)

Project Structure

reactor-twin/
├── src/reactor_twin/
│   ├── core/           # Neural DE Engine (Neural ODE, Latent, SDE, CDE)
│   ├── physics/        # 7 hard/soft constraints
│   ├── reactors/       # 4 reactor types + 6 kinetics models + 5 benchmarks
│   ├── training/       # Training engine + losses + data generation
│   ├── digital_twin/   # EKF, fault detection, MPC, online adaptation, meta-learning
│   └── dashboard/      # 10-page Streamlit app
├── tests/              # Test suite
├── examples/           # Runnable examples
└── pyproject.toml      # Package configuration

See the documentation for complete architecture details.

---

Dashboard

Launch the interactive Streamlit dashboard:

pip install -e ".[dashboard]"
reactor-twin-dashboard
# Or: streamlit run src/reactor_twin/dashboard/app.py

10 pages: Reactor Simulator, Phase Portraits, Bifurcation Diagrams, RTD Analysis, Parameter Sweeps, Sensitivity Analysis, Pareto Optimization, Fault Monitoring, Model Validation, Latent Space Explorer.

---

Benchmarks

Accuracy (vs ground truth):

• CSTR exothermic: RMSE < 5% (target), < 1% (stretch)
• PFR 1D: RMSE < 8% (target), < 3% (stretch)

Physics Compliance:

• Mass balance error: < 10^-8 (hard), < 10^-3 (soft)
• Positivity violations: 0 (hard), < 0.1% (soft)

Speed:

• Single trajectory: < 5 ms (100x faster than scipy)
• Parameter sweep (10K conditions): < 5 s (1000x faster)
• MPC optimization: < 100 ms (real-time capable)

---

Development

Setup

# Install development dependencies
pip install -e ".[dev]"

# Install pre-commit hooks
pre-commit install

Testing

# Run tests
pytest

# With coverage
pytest --cov=reactor_twin --cov-report=html

Linting & Formatting

# Lint with ruff
ruff check src/

# Format with ruff
ruff format src/

# Type check with mypy
mypy src/

---

Roadmap

• Phase 1 (Weeks 1-2): Core Neural ODE + CSTR benchmark ✅
• Phase 2 (Weeks 3-4): 7 physics constraints + training infrastructure ✅
• Phase 3 (Weeks 5-6): Advanced DEs (Latent/Augmented/SDE/CDE) ✅
• Phase 4 (Weeks 7-8): Batch, Semi-batch, PFR + 4 kinetics + 3 benchmarks ✅
• Phase 5 (Weeks 9-10): Digital twin features + 10-page dashboard ✅
• Phase 6 (Weeks 11-12): Tests, examples, docs, PyPI release ✅

See ROADMAP.md for details.

---

Citation

If you use ReactorTwin in your research, please cite:

@software{reactortwin2026,
  author = {Karthikeyan, Tubhyam},
  title = {ReactorTwin: Physics-Constrained Neural Differential Equations for Chemical Reactor Digital Twins},
  year = {2026},
  url = {https://github.com/ktubhyam/reactor-twin}
}

---

License

MIT License - see LICENSE for details.

---

Contact

Tubhyam Karthikeyan Email: [e-mail] GitHub: @ktubhyam Website: tubhyam.dev

---

Acknowledgments

Built on:

• torchdiffeq - Neural ODE solver
• PyTorch - Deep learning framework
• Inspired by DeepXDE (plugin architecture) and Wang et al. 2025 (foundation models for reactors)