Physics-Informed Neural Networks for Dissipative Dynamical Systems

190 concepts125 papers1 proposal

Early breakthroughs in physics-informed machine learning demonstrated that embedding classical mechanics directly into neural network architectures yields models with exceptional data efficiency and long-term stability. For instance, Hamiltonian Neural Networks introduced by Greydanus et al. and Deep Lagrangian Networks introduced by Lutter et al. force neural networks to respect fundamental conservation laws by learning a scalar energy function rather than predicting next states directly.

However, these idealized frameworks suffer from a critical flaw when applied to the real world: they assume perfect energy conservation. Real-world physical systems are almost always non-conservative. Friction, air resistance, thermal dissipation, and external control inputs constantly alter a system's total energy. When a rigidly conservative model encounters a system that loses energy, it experiences a mathematical tug-of-war, failing to reconcile its hard-coded conservation priors with the dissipative reality of the training data.

To bridge this gap, a major research direction focuses on extending physics-informed architectures to model non-conservative and dissipative dynamics without sacrificing the structural guarantees that make physics priors valuable.

The Additive Approach: Port-Hamiltonian and Forced Variational Systems

The most intuitive way to model a non-conservative system is to split its dynamics into two distinct parts: an idealized, energy-conserving core and an additive non-conservative term.

In classical mechanics, this is elegantly formalized by port-Hamiltonian systems. Instead of modeling a closed system where the state evolves purely according to the gradient of a Hamiltonian $H$, the port-Hamiltonian framework introduces explicit terms for internal dissipation and external interaction:

$$\dot{x} = [J(x) - R(x)] \nabla H(x) + g(x)u$$

Here, $J(x)$ is an anti-symmetric matrix representing the conservative internal physics, $u$ is an external control input scaled by $g(x)$, and $R(x)$ is a positive semi-definite dissipation matrix that captures energy loss.

Zhong et al. and Duong et al. leverage this structure by parameterizing $J$, $R$, and $H$ as separate neural networks. This works remarkably well because of the mathematical constraints placed on $R(x)$. By forcing $R(x)$ to be positive semi-definite (for example, by predicting its Cholesky decomposition), the network is structurally incapable of generating energy out of nowhere through friction. The model guarantees passivity—meaning the system can only dissipate energy internally, never create it—while still allowing the network the flexibility to learn complex, state-dependent friction forces from raw data.

A parallel strategy applies this additive philosophy to Lagrangian mechanics via the discrete Lagrange-d'Alembert principle. Rather than deriving motion strictly by minimizing an action integral, Havens et al. introduce Forced Variational Integrator Networks. This architecture uses a neural network to learn a generalized non-conservative force field that performs "virtual work" on the system, effectively disentangling the underlying conservative inertia from external damping and actuation.

The Geometric Approach: Contact Manifolds and Metriplectic Brackets

While additive terms successfully patch conservative equations, they treat dissipation as an external disturbance rather than a fundamental property of the underlying space. A deeper, more unified approach reformulates the geometry of the state space itself to natively accommodate energy loss.

Standard Hamiltonian mechanics operates on even-dimensional symplectic manifolds, where phase space volume is strictly conserved (Liouville's theorem). Because friction inherently contracts phase space volume, it cannot be cleanly integrated into a symplectic framework. To resolve this, researchers have turned to contact geometry, which operates on odd-dimensional manifolds.

By introducing an extra coordinate that explicitly tracks the system's energy action, contact geometry allows the system's total energy to fluctuate dynamically. Testa et al. utilize this framework to create Geometric Contact Flows, embedding dissipation directly into the geometric structure of the network. Similarly, contact Hamiltonian mechanics has been adapted to stochastic settings to model non-conservative bridges where energy levels fluctuate due to cellular growth or environmental friction, as demonstrated in the development of Contact Wasserstein Geodesics.

For complex thermodynamic systems where energy is conserved globally but degraded locally into heat, a more comprehensive algebraic structure is required. The authors of Machine learning structure preserving brackets for forecasting irreversible processes address this by embedding metriplectic algebraic structures (the GENERIC formalism) into neural ordinary differential equations.

A metriplectic system defines the evolution of a dynamical system using two separate brackets and two scalar functions (Total Energy $E$ and Total Entropy $S$):

$$\dot{x} = \{x, E\} + [x, S]$$

The first term is a Poisson bracket that governs the reversible, conservative dynamics, while the second term is a symmetric, positive semi-definite metric bracket that governs irreversible dissipation. By enforcing structural degeneracy conditions between these brackets (i.e., the energy function produces zero change through the dissipative bracket, and the entropy function produces zero change through the conservative bracket), the architecture guarantees by construction that total energy is perfectly conserved while total entropy strictly increases over time.

Why It Works: Preserving Stability Without Rigidity

The core reason these frameworks outperform standard black-box neural networks is that they preserve the global stability properties of physics while shedding the unrealistic constraint of perfect energy conservation.

If a standard neural network is trained on a pendulum slowing down due to air resistance, it may accurately predict the trajectory for a few cycles. However, because it lacks an intrinsic concept of energy bounds, small compounding errors can easily cause the predicted trajectory to explode into unphysical, self-amplifying oscillations over long time horizons.

By utilizing port-Hamiltonian matrices, contact geometry, or metriplectic brackets, the network is bound by a mathematical safety rail. It can learn highly complex, non-linear, and messy dissipative behaviors—such as joint friction, cable stretch, or aerodynamic drag—but it must do so within a geometric framework that forces the system's total energy to behave in a physically plausible manner. The result is a class of models that retains the data efficiency and long-term stability of conservative networks while successfully generalizing to real-world, non-idealized physical environments.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 133 papers
  2. Version 1 · 2026-10-04 · 206 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_non_conservative_force_modeling_c5aca2be,
  title = {non-conservative force modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-conservative-force-modeling/rdr_c5aca2be17934d17b79a0d099cb0523b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_non_conservative_force_modeling_c5aca2be_v2,
  title = {non-conservative force modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-conservative-force-modeling/rdr_c5aca2be17934d17b79a0d099cb0523b/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake