In the last few years, physics-informed machine learning has been dominated by the success of architectures that "bake in" the fundamental laws of conservation. Models like Hamiltonian Neural Networks and Lagrangian Neural Networks by Miles Cranmer and colleagues have shown that by constraining a network to learn an energy function rather than a raw trajectory, we can achieve remarkable data efficiency and long-term stability.
However, these models share a common "ideal world" bias: they assume energy is perfectly preserved. In the real world, energy is rarely constant. Friction, air resistance, heat dissipation, and external driving forces are the norm, not the exception. A new research direction is emerging to bridge this gap, extending energy-based priors to handle the "messy" physics of dissipative and non-conservative systems.
The Friction Problem in Symplectic Manifolds
The core difficulty in modeling real-world physics is that standard Hamiltonian and Lagrangian frameworks operate on symplectic manifolds, where the volume of phase space is conserved. As Yaofeng Desmond Zhong and Amit Chakraborty note in their work on Symplectic ODE-Net, these architectures struggle to account for dissipative forces like friction because such forces fundamentally break the symmetry of the system.
When a robot arm moves, it loses energy to joint friction; when a drone flies, it battles air drag. If a model is strictly constrained to conserve energy, it will either fail to fit the data or "hallucinate" ghost forces to explain why the energy is disappearing. To solve this, researchers are moving beyond pure conservation to architectures that explicitly model how energy leaves or enters the system.
Augmenting the Hamiltonian with Dissipation
The most direct way to handle energy loss is to augment the governing equations with a learnable dissipation term. Tomoharu Iwata and Hongyu Zhang demonstrate this by incorporating a learned dissipation matrix alongside the Hamiltonian framework. This allows the model to represent both the conservative "swing" of a pendulum and the "decay" caused by friction within a single unified representation.
Similarly, Mingquan Feng and Junchi Yan suggest using a Rayleigh dissipation function. By adding this term to the Euler-Lagrange equations, the network can learn to disentangle the underlying potential energy from the non-conservative forces that damp the motion. This approach is further refined in Forced Variational Integrator Networks by Aaron J. Havens, which uses the Lagrange-d’Alembert principle to account for both friction and active motor inputs.
Geometric Generalization: Contact and Metriplectic Systems
While adding a "friction term" works for simple mechanical systems, more complex phenomena—like biological development or chemical reactions—require a deeper geometric rethink.
One promising path is Contact Hamiltonian mechanics. Unlike standard Hamiltonians, contact geometry treats energy as an explicit state variable that can fluctuate. Andrea Testa, Leonel Rozo, and Søren Hauberg leverage this in their Non-Conservative Generalized Schrödinger Bridge framework. By incorporating a damping factor directly into the geometric structure, they can model stochastic processes where energy is not conserved, such as systems with growth or decay.
Another sophisticated approach involves Metriplectic systems. As Anthony Gruber and Nathaniel Trask explain, these systems use a "bracket-based" dynamics that combines a Poisson bracket (for conservation) with a metric bracket (for dissipation and entropy production). This allows a single model to respect both the First and Second Laws of Thermodynamics simultaneously, making it ideal for modeling non-equilibrium processes like heat loss in fluids.
Extracting Physical Truth from Dissipation
A major goal of this research is not just to simulate physics, but to discover it. If a model can accurately separate conservative forces from non-conservative ones, it can identify "new physics" or hidden variables.
Bohan Wang and Max Tegmark developed the Neural New-Physics Detector (NNPhD), which decomposes force fields into their conservative and non-conservative parts. By watching a system's trajectory, the model can detect when energy conservation is violated, potentially signaling the presence of friction or even a hidden external body.
To ensure these learned models are physically meaningful, Aiqing Zhu and Qianxiao Li introduced the UIDD framework. Their work focuses on identifiability—ensuring that the neural network extracts a unique, physically consistent energy landscape and dissipation profile from the data, rather than just finding one of many possible mathematical fits.
Why It Works: The Power of Structured Inductive Biases
The shift from "purely conservative" to "structured dissipative" modeling works because it retains the best of both worlds. By keeping the Hamiltonian or Lagrangian core, these models maintain the stability and interpretability of classical physics. By adding flexible, learnable components for dissipation, they gain the ability to generalize to real-world hardware and non-equilibrium environments.
As we move from simulating pendulums in a vacuum to controlling soft robots and modeling biological systems, the ability to learn how energy is lost will be just as important as knowing how it is saved.