Neural Ordinary Differential Equations for Continuous Physical Dynamics

This direction is merged into another direction: neural ordinary differential equations (continuous-time latent dynamics). Its last version stays readable and citable here.

Version 1 of 118 papers18 conceptsfrozen 2026-10-04

Most machine learning models treat time as a sequence of discrete snapshots. To predict the future state of a system, a standard neural network learns a mapping from one time step to the next: $x_{t+1} = f(x_t)$. However, the physical world—from the flow of air over a wing to the spread of a virus—operates in continuous time.

A growing body of research suggests that instead of learning discrete transitions, we should use neural networks to parameterize the derivative of the system's state. This approach, known as Neural Ordinary Differential Equations (Neural ODEs), allows models to learn the underlying "vector field" that governs how a system evolves.

Learning the Rate of Change

The core intuition behind Neural ODEs is that the neural network $f(x, t, \theta)$ does not predict the next state directly. Instead, it defines the slope: $\frac{dx}{dt} = f(x, t, \theta)$. To find the state of the system at any future time, we simply integrate this learned derivative using a numerical solver.

This shift from "jumps" to "flows" offers several fundamental advantages:

  1. Irregular Sampling: Because the model defines a continuous path, it can handle data points collected at any time interval, a feature Rongfan Li and colleagues used to improve land deformation predictions from irregularly monitored sensors.
  2. Physical Consistency: Physical laws like heat conduction or fluid flow are naturally expressed as differential equations. By using Neural ODEs, researchers like Maximilian Kannapinn can create "digital twins" of 3D printing processes that respect the continuous nature of thermal dynamics.
  3. Parameter Efficiency: By learning the governing laws rather than specific transitions, models can be significantly smaller. Yogesh Verma demonstrated this with ClimODE, which achieves state-of-the-art weather forecasting with 100x fewer parameters than traditional black-box models.

Hybrid Physics and "Patching" Equations

We often have a partial understanding of a system's physics but lack the full picture. Neural ODEs excel at "patching" these gaps. In mechanical engineering, Meike Wohlleben and Sebastian Peitz have explored hybrid models for rubber-metal bushings. They use simplified physics equations to handle the known dynamics and a neural network to learn the "residual" errors—the complex, non-linear behaviors that are too difficult to derive by hand.

Similarly, in fluid dynamics, Romit Maulik and Daniel Livescu use Neural ODEs to solve the "closure problem." When simulating complex fluids, scientists often use Reduced-Order Models (ROMs) to save time, but these models lose fine-grained details. Neural ODEs can learn to fill in these missing dynamics in the latent space, maintaining accuracy without the massive computational cost of full-scale simulations.

Zero-Shot Adaptation via Basis Functions

One of the most powerful extensions of this idea is the ability to adapt to new environments without retraining. If we view different environments (like a drone flying in wind vs. calm air) as variations of the same underlying physics, we can learn a "basis" of dynamics.

Tyler Ingebrand and Ufuk Topcu developed a framework where a system learns a set of fundamental Neural ODE basis functions. When the system encounters a new environment, it doesn't need to update its weights. Instead, it performs a fast, gradient-free calculation to find the right linear combination of those basis functions. This allows robots to adapt to unpredictable terrains—like moving from mud to ice—in real-time, as shown in work by William Ward.

Beyond Physics: Biological and Social Flows

The utility of continuous-time modeling extends beyond traditional physics. Mahta Ramezanian Panahi and Irina Rish have highlighted how Neural ODEs can bridge the gap between biophysical brain models and data-driven AI, capturing the continuous temporal patterns of neural activity.

Even social phenomena can be modeled as flows. Matías Núñez and Rafael A. Barrio demonstrated that by training Neural ODEs on social media data, they could forecast COVID-19 outbreaks months in advance. By learning the "vector field" of how symptoms are reported online, the model can simulate "what-if" scenarios for public health interventions, treating the spread of a virus as a continuous dynamical system rather than a series of disconnected events.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_neural_ordinary_differential_equations_f_bc5a84d9,
  title = {Neural Ordinary Differential Equations for Continuous Physical Dynamics},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/neural-ordinary-differential-equations-for-continuous-physical-dynamics/rdr_bc5a84d9923a431d8729b57e506fcc53},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_neural_ordinary_differential_equations_f_bc5a84d9_v1,
  title = {Neural Ordinary Differential Equations for Continuous Physical Dynamics},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/neural-ordinary-differential-equations-for-continuous-physical-dynamics/rdr_bc5a84d9923a431d8729b57e506fcc53/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}