⚡ electrical engineering

RTS Smoother-Guided Learning of Physics-Based Neural Differential Models

This paper proposes a two-stage hybrid neural-physics framework that alternates between Rauch--Tung--Striebel (RTS) smoother-based state estimation and neural network parameter optimization to learn missing components of ordinary differential equations from partial state observations while preserving interpretable mechanistic structure.

Original authors: Ahmet Demirkaya, Georgios Stratis, Tales Imbiriba, Zachary D. Danziger, Deniz Erdogmus

Published 2026-07-17

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Ahmet Demirkaya, Georgios Stratis, Tales Imbiriba, Zachary D. Danziger, Deniz Erdogmus

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but you only have a few blurry snapshots of the suspect's movements, and the rest of the story is missing. In the world of science, this is a common problem when studying how things change over time, like how a heart beats, how a chemical reaction bubbles, or how a neuron fires. Scientists use mathematical recipes called "Ordinary Differential Equations" (ODEs) to describe these changes. Think of an ODE as a set of instructions that says, "If the system is in this state right now, here is exactly where it will be a split second later." Usually, these recipes are perfect, but in the real world, we often don't know the whole recipe. Some ingredients are missing, and we can't see every part of the system; we might only see the temperature but not the pressure, or the voltage but not the hidden switches inside a cell. This makes it incredibly hard to predict the future or understand the past because the missing pieces could be the most important ones.

This paper tackles that exact headache. The authors propose a clever new way to fill in the blanks. Instead of guessing the missing parts with a black-box computer program that might get confused by noise, they use a "time-traveling" math tool called a Rauch–Tung–Striebel (RTS) smoother. Imagine you are watching a movie, but the projector is flickering and skipping frames. A normal viewer (or a standard computer filter) tries to guess what happens next based only on what they just saw. But an RTS smoother is like a detective who watches the entire movie from start to finish, then rewinds and fills in the missing frames by looking at the clues from the future as well as the past. By using this "all-seeing" view to guess the hidden states during training, the authors teach a hybrid model—part known physics, part artificial intelligence—to learn the missing rules of the game. They found that this method is much better at reconstructing the hidden parts of the system and predicting the future, even when the data is noisy or incomplete, compared to other popular methods that rely on guessing the hidden states one step at a time.

The Detective's Dilemma: Missing Pieces and Noisy Clues

In many fields, from biology to engineering, scientists use math to model how systems evolve. But often, the math is incomplete. We might know that a car moves forward, but we don't know exactly how the engine reacts to the gas pedal. Or we might see a neuron's electrical spike but miss the tiny internal switches that cause it. To make matters worse, our sensors are imperfect. We get noisy measurements, like trying to hear a whisper in a windstorm, and we often can't measure everything at once.

The paper suggests a hybrid approach: keep the parts of the math we know for sure (the "mechanistic" parts) and use a neural network (a type of AI) to learn the parts we don't know. The tricky part is training this AI. Usually, to teach the AI, you need to know the "ground truth"—the actual, hidden path the system took. But if we can't measure the hidden parts, we don't have that truth. Other methods try to guess the hidden path using a "recurrent" neural network, which is like a student trying to memorize a story by reading it one sentence at a time. If the story is messy or the student is tired (noisy data), they might get lost.

The Time-Traveling Solution

The authors of this paper decided to take a different route. Instead of guessing the hidden path one step at a time, they used a tool called the RTS smoother. Think of the RTS smoother as a super-smart editor who reads the entire story of the system's measurements from beginning to end before writing a single word of the "hidden" story. Because it looks at the whole timeline, it can smooth out the noise and figure out the most likely path the hidden variables took, even if they were never directly measured.

The process works in two alternating steps, like a dance:

The Guess: First, they freeze the AI's brain and use the RTS smoother to guess the hidden states (the "latent" states) based on the noisy measurements and the known parts of the physics. This gives them a "smoothed" trajectory that looks like the ground truth.
The Lesson: Next, they treat this smoothed trajectory as if it were the real truth. They use it to teach the neural network how to fill in the missing physics equations.
Repeat: They switch back and forth between guessing the path and teaching the AI until the model gets really good at it.

Why This Matters: The Long Game

One of the biggest problems with other methods is that they are great at predicting the next second but terrible at predicting the next hour. If you make a tiny mistake in step one, that mistake grows and grows, like a snowball rolling down a hill, until the prediction is completely wrong. The authors realized that training the AI to only look one step ahead wasn't enough.

So, they added a "long-horizon" test. They didn't just ask the AI, "Can you predict the next second?" They asked, "Can you predict the next 50 seconds?" If the AI drifts away from the truth over time, they penalize it. This forces the AI to learn the real underlying rules of the system, not just a shortcut that works for a split second.

The Results: A Clearer Picture

The authors tested their method on five different systems, ranging from a simple swinging pendulum to complex biological models like a yeast cell's energy cycle and a neuron's electrical firing. In every case, they compared their method against other popular AI techniques.

The Pendulum: When only the position of a swinging weight was measured (but not its speed), their method perfectly reconstructed the hidden speed and kept the swing going correctly.
The Neuron: For a model of a brain cell, they successfully reconstructed a hidden "gate" variable that controls when the cell fires, even though it was never measured.
The Chemical Reaction: In a complex chemical system, they recovered the hidden chemical concentrations and kept the oscillating pattern stable.

The results were impressive. In simulations, their method reduced the error in guessing the hidden states by a huge margin compared to other methods. For example, in the yeast model, the error dropped from about 0.19 (for other methods) to just 0.0004. They also found that their model stayed accurate over long periods, while others drifted off course.

Robustness: Handling the Noise

Perhaps the most exciting finding was how well the method handled bad data. The authors tested their system with increasingly noisy measurements, simulating a situation where the "windstorm" got louder and louder. Even when the signal was very weak (a signal-to-noise ratio of just 5), the RTS smoother was still able to reconstruct a coherent path for the hidden variables. This suggests that the method is very robust and doesn't break down easily when the data is messy, which is exactly what happens in the real world.

The Bottom Line

This paper doesn't claim to have solved every mystery in science. It admits that if the data is extremely sparse or if the noise isn't what we think it is, the method might still get confused. However, it offers a powerful new tool for scientists who are trying to understand complex systems with missing pieces. By using a "time-traveling" smoother to guide the learning process, they can build hybrid models that are not only accurate but also interpretable—we know which parts are based on known physics and which parts are learned by the AI.

The key takeaway is that sometimes, to understand the future, you need to look at the whole picture, not just the next step. By combining the reliability of classical math with the flexibility of AI, and using a clever "all-seeing" training trick, this method helps us build better digital twins of the real world, even when we can't see everything.

Technical Summary: RTS Smoother-Guided Learning of Physics-Based Neural Differential Models

Problem Statement

Ordinary Differential Equations (ODEs) are standard tools for modeling dynamical systems in fields such as physics, biology, and engineering. However, practical applications face two primary challenges:

Unknown or Incomplete Dynamics: Mechanistic models are often only partially specified, with missing terms, simplified feedback mechanisms, or subject-specific effects not captured by closed-form equations.
Partial and Noisy Measurements: Typically, only a subset of state variables is observed, often indirectly and with noise, making full-state reconstruction difficult.

Existing hybrid approaches that combine mechanistic models with neural networks often rely on learned neural encoders or recurrent architectures (e.g., Latent ODEs, GRU-ODE-Bayes) to infer latent states. The authors argue that under strong noise, irregular sampling, or severe partial state measurement, these learned inference mechanisms can be brittle, and their latent representations may lack physical interpretability or stability.

Methodology

The paper proposes a hybrid neural-physics framework where known ODE components remain explicit, while missing components are represented by a neural network. The core innovation is an iterative two-stage algorithm that alternates between state estimation and parameter learning, utilizing a Rauch–Tung–Striebel (RTS) smoother rather than a learned encoder.

1. Problem Formulation

The system is modeled as a stochastic state-space model:

Dynamics: $\dot{x}(t) = f_\theta(x(t)) + w(t)$ , where $f_\theta$ is a partially known model with parameters $\theta$ .
Measurements: $y(t_k) = h(x(t_k)) + \epsilon_k$ .
The goal is to minimize the negative joint log-likelihood of the parameters $\theta$ given noisy, partial observations. Since the latent states $x$ are unobserved, direct optimization is intractable.

2. The RTS-Guided Training Loop

The authors replace the unknown latent trajectory with a smoothed estimate $\hat{x}$ produced by an RTS smoother, treating this estimate as a "pseudo-ground truth" for training. The process alternates between:

Stage 1: State Estimation (RTS Smoothing):
With fixed neural network parameters $\theta$ , the RTS smoother processes the full measurement sequence to estimate the latent state trajectories $\hat{x}_{0:T}$ and their associated covariances $P_{0:T}$ . This step leverages the known mechanistic structure and noise models to infer states consistent with the entire observation history, rather than just past data (as in filtering).
Stage 2: Parameter Update (Backpropagation):
Treating the smoothed trajectories $\hat{x}$ and covariances $P$ as fixed, the neural network parameters $\theta$ are updated via gradient descent to minimize a surrogate loss function.

3. Multi-Horizon Training Objective

The training loss is a weighted sum of errors over multiple prediction horizons ( $n \in \mathcal{H}$ ):
$\mathcal{L} = \frac{1}{T_{tot}} \sum_{i,n} \left( \| e^{(y)}_{k+n} \|^2_{R^{-1}} + \| e^{(x)}_{k+n} \|^2_{P^{-1}_{k+n}} \right)$

Observation Error: Weighted by the inverse measurement noise covariance ( $R^{-1}$ ).
Latent State Error: Weighted by the inverse smoother covariance ( $P^{-1}$ ). This acts as a confidence weight; if the smoother is uncertain (large $P$ ), the error is downweighted.
Multi-Horizon: The authors argue that single-step ( $n=1$ ) training is insufficient for long-horizon stability, especially in partially observed systems where errors in unmeasured variables may only manifest after several integration steps. Therefore, the loss includes rollout errors over longer time scales.

4. Hybrid Architecture

The dynamics model $f_\theta$ is partitioned into:

$x_{phys}$ : States with known analytical dynamics ( $\dot{x}_{phys} = f_{phys}$ ).
$x_{unk}$ : States with unknown dynamics, replaced entirely by a neural network ( $\dot{x}_{unk} = f_{unk}$ ).
This preserves the interpretability of known physics while learning the missing terms.

Key Contributions

RTS-Smoother-Guided Learning: The paper introduces a method to train hybrid neural-ODEs by using RTS smoothing to generate stable, dynamics-consistent pseudo-targets for latent states, decoupling state estimation from the learning of the neural dynamics.
Separation of Estimation and Learning: Unlike methods that learn a joint encoder-decoder or require online filtering at test time, this approach distills the smoothing information into a standalone hybrid ODE. The final deployed model requires no online state estimator.
Uncertainty-Aware Weighting: The training loss utilizes the smoother's posterior covariance ( $P_k$ ) to weight state-transition errors, ensuring that the neural network learns more from high-confidence state estimates.
Multi-Horizon Objective: The inclusion of longer rollout horizons in the loss function addresses the issue of error accumulation in long-term predictions, which single-step training often fails to prevent.

Experimental Results

The method was evaluated on five benchmark systems with varying complexity (linear, nonlinear, stiff) and partial observability:

Harmonic Oscillator: 1D unmeasured state.
Hodgkin–Huxley Neuron: Stiff, multi-state biophysical model with unmeasured gating variables.
Retinal Circulation: Physiological model with indirect measurements.
Brusselator: Nonlinear reaction network with oscillatory behavior.
Yeast Glycolysis: High-dimensional biochemical oscillator.

Performance Metrics:

Latent State RMSE: The proposed method achieved the lowest error on four of the five systems, with significant improvements over NeuralODE, GRU-ODE-Bayes, and CKF-based baselines (e.g., a drop from $1.92 \times 10^{-1}$ to $4.08 \times 10^{-4}$ on Yeast Glycolysis).
Rollout Hausdorff Distance: The method achieved the lowest distance on all five systems, indicating superior long-horizon trajectory alignment with ground truth.
Robustness: Experiments showed the method remains stable even under high measurement noise (low SNR), whereas encoder-based methods tend to become brittle.
Ablation Studies: Removing RTS guidance, latent-state loss, or covariance weighting all resulted in performance degradation, confirming the necessity of all three components.

Significance and Claims

The authors claim that their approach offers a robust alternative to purely data-driven latent inference, particularly in scenarios with sparse or noisy data. By leveraging classical Kalman smoothing, the method:

Retains Interpretability: The known mechanistic parts of the model remain explicit.
Improves Stability: The smoother provides coherent trajectories that prevent the "brittleness" often seen in learned encoders under partial observability.
Enables Deployment: The final output is a compact, physics-based neural differential equation that can be simulated forward without the computational overhead of running a filter or smoother during inference.

The paper concludes that while the method assumes known measurement functions and Gaussian noise, it effectively bridges the gap between classical state estimation and modern neural differential equation learning, providing a reliable framework for discovering missing dynamics in partially observed systems.