| name | rts-neural-physics-ode-learning |
| description | Hybrid neural-physics framework for learning unknown components of ODEs using Rauch-Tung-Striebel smoother and neural networks |
| metadata | {"arxiv_id":"2607.15180v1","authors":["Ahmet Demirkaya","Georgios Stratis","Tales Imbiriba","Zachary D. Danziger","Deniz Erdogmus"],"published":"2026-07-16","categories":["cs.LG","eess.SY"],"keywords":["Rauch-Tung-Striebel smoother","neural differential equations","partial state observation","hybrid modeling","system identification"]} |
| license | Complete terms in LICENSE.txt |
RTS Smoother-Guided Learning of Physics-Based Neural Differential Models
This skill implements the hybrid neural-physics framework from arXiv:2607.15180v1 for learning unknown components of ordinary differential equations (ODEs) when only partial state measurements are available.
Overview
When modeling dynamical systems, we often know some components of the dynamics from first principles but have unknown components that need to be learned from data. This skill provides a methodology for combining known physics with neural networks to identify missing ODE components using partial state measurements.
The approach alternates between:
- State estimation using a Rauch-Tung-Striebel (RTS) smoother (assuming known parameters)
- Parameter estimation using backpropagation on neural networks (assuming known states)
When to Use This Skill
Use this skill when you need to:
- Model dynamical systems with partially known dynamics
- Learn unknown ODE components from partial state measurements
- Combine mechanistic models with data-driven components
- Work with systems where only a subset of state variables are measurable
- Apply system identification to physical, biological, or physiological systems
Core Methodology
Two-Stage Alternating Procedure
Stage 1: State Estimation (RTS Smoother)
- Treat model parameters as known constants
- Use available measurements to estimate latent states via Rauch-Tung-Striebel smoother
- Produces smoothed state trajectories that minimize estimation error
Stage 2: Parameter Estimation (Backpropagation)
- Treat smoothed trajectories from Stage 1 as known
- Use these trajectories to train neural network parameters via backpropagation
- Minimize prediction error between model outputs and measurements
Iteration: Repeat stages until convergence (predetermined criterion met)
Mathematical Formulation
For a system:
dx/dt = f_known(x, θ_known) + f_nn(x, θ_nn)
y = h(x) + v
Where:
f_known: Known physics-based component
f_nn: Neural network component to be learned
θ_known: Known parameters
θ_nn: Neural network parameters to learn
y: Measured outputs (partial state observation)
v: Measurement noise
Implementation Steps
- Define Known Components: Identify and implement the known portions of your ODE based on first principles
- Neural Network Architecture: Design a neural network to approximate the unknown dynamics
- RTS Smoother Implementation: Implement the Rauch-Tung-Striebel smoother for state estimation
- Alternating Optimization:
- Fix neural network weights, run RTS smoother to estimate states
- Fix state trajectories, update neural network weights via backpropagation
- Repeat until convergence criteria met
- Validation: Test on held-out data and assess generalization to unseen conditions
Key Advantages
- Interpretability: Preserves known physical structure while learning unknown components
- Data Efficiency: Leverages partial measurements effectively through smoothing
- Generalization: Learned neural network components can extrapolate to new conditions
- Stability: RTS smoother provides optimal state estimates under Gaussian assumptions
Practical Considerations
Network Architecture
- Choose appropriate architecture (MLP, LSTM, etc.) based on system dynamics
- Ensure sufficient capacity to capture unknown dynamics without overfitting
- Consider physics-informed constraints if applicable
Convergence Criteria
- Monitoring validation loss on held-out data
- Checking parameter stability between iterations
- Setting maximum iterations to prevent infinite loops
Initialization
- Proper initialization of neural network weights is crucial
- Consider pretraining on available data if possible
- Initialize physical parameters based on prior knowledge when available
Validation Approach
- Synthetic Data Testing: Validate on systems with known ground truth
- Cross-Validation: Use temporal cross-validation for time series data
- Prediction Horizon: Test both short-term and long-horizon predictions
- Robustness: Evaluate sensitivity to noise levels and sampling rates
Extensions and Variations
- Multiple Unknown Components: Extend to learn several distinct unknown functions
- Time-Varying Parameters: Allow learned components to vary with time or operating conditions
- Uncertainty Quantification: Incorporate Bayesian approaches for uncertainty estimates
- Hybrid Architectures: Combine multiple neural network types for different subsystems
Common Pitfalls
- Over-reliance on Neural Network: Ensure known physics component is sufficiently accurate
- Poor Observability: Verify that measured states provide sufficient information for estimation
- Local Minima: Use multiple random initializations to avoid poor local solutions
- Overfitting: Regularize neural network and validate on unseen data
- Numerical Stability: Ensure proper scaling of states and parameters
References
- arXiv:2607.15180v1 - RTS Smoother-Guided Learning of Physics-Based Neural Differential Models
- Rauch, H.E., Tung, F., & Striebel, C.T. (1965). Maximum likelihood estimates of linear dynamic systems.
- Chen, T.Q., Rubanova, Y., Bettencourt, J., & Duvenaud, D.K. (2018). Neural Ordinary Differential Equations.
Related Skills
- neural-ode-based-modeling: For pure neural ODE approaches
- physics-informed-neural-networks: For PINNs applied to PDEs
- system-identification-methods: For traditional system identification techniques