acceptodds
Under review as a conference paper at ICLR 2027

Lagrangian and Hamiltonian Neural Networks With a Dissipative System

Abstract

We investigate the applicability of Largrangian and Hamiltonian Neural Network models to a dissipative system that has explicit time dependence in its Lagrangian, Hamiltonian, and total energy. To do so we consider these neural network models for simulated systems of a harmonic one-dimensional, one-component oscillator with damping, as well as without damping for comparison. We find that both the Lagrangian and Hamiltonian approaches are able to predict the empirical physical behavior of the damped oscillator systems and to effectively "learn" to varying degrees the underlying Lagrangians and Hamiltonians, as has previously been shown to be the case with undamped oscillator systems. These investigations elucidate important properties of Lagrangian and Hamiltonian mechanics, including properties that are not manifest when considering systems without explicit time dependence.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Lagrangian NN, Hamiltonian NN, One-dimensional Oscillator, Dissipative System

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_KJpQOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper applies Lagrangian neural networks (LNNs) and Hamiltonian neural networks (HNNs) to a one-dimensional damped harmonic oscillator and examines their behavior through comparisons with an undamped oscillator. It conducts experiments under several conditions, including with and without noise, and discusses the non-uniqueness of Lagrangians and the distinction between the Hamiltonian and physical energy in a damped system.

Critical strengths

The paper has educational value as an implementation example of LNNs and HNNs using a simple damped oscillator. However, this alone does not sufficiently support the novelty or significance of the research contribution.

Critical weaknesses

The contribution beyond applying existing LNNs and HNNs to a particular damped oscillator is unclear.

The paper presents experiments on a one-dimensional damped harmonic oscillator, varying time inputs, momentum representations, loss functions, and the presence or absence of noise. However, it is unclear what research question these comparisons answer or what they newly establish. Although the paper demonstrates that the target system’s behavior can be approximated under individual configurations, it does not adequately explain the insights obtained by bringing these results together.

Prior LNN and HNN approaches already address dissipative systems, including references [12] and [13], and HNNs with time as an input are also considered by Desai et al. (2021). The paper does not sufficiently establish a new methodology or theoretical result beyond these approaches, or design insights that can be reused for other problems. The experiments are also limited to a single oscillator and include no direct comparison with existing methods, making it difficult to assess what additional value the various configurations provide over prior work. In its current form, the research significance beyond a case study presenting individual application results remains unclear.

Additional reference: Desai, S., Mattheakis, M., Sondak, D., Protopapas, P., and Roberts, S. (2021). Port-Hamiltonian Neural Networks for Learning Explicit Time-Dependent Dynamical Systems.

Questions for authors
  1. What new insights, methods, or theoretical results can be reused for other problems without relying on formulations or known parameters specific to this oscillator? Please clarify the contribution beyond applying existing LNNs and HNNs to an individual problem.

  2. Please explain how the initial phases and initial states were selected, and the rationale for the number of trajectories and the sampling intervals. How did you assess the effects of these choices on state-space coverage and predictive performance?

Reviewer_HFNJOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The papers re-uses physics inspired models (Lagrangian and Hamiltonian Neural Network) to model a toy one-dimensional damped harmonic oscillator. They observe that in the damped case the Hamiltonian and the Lagrangian have an explicit time dependence which they add to their model. They show empirical results of how their model fits the state of the damped harmonic oscillator, with and without the damping and with and without noise injection in the data and compare this with the Lagrangian and Hamiltonian Neural Network.

Critical strengths

The simple derivations are correctly laid out step by and step and the experiments are correctly designed.

Critical weaknesses

1. The problem and relevance to the ML community is not argued. The introduction refers to previous work (Cranmer et al. and Greydanus et al.) "in a similar, straightforward method" (l.096–097), and that prior work treated time-independent L and H (l.073–076). But it's not clear why extending things to the time-dependent domain is an interesting problem.

2. Empirical and theoretical results are weak. Empirical. The results are only in 1D linear systems that have analytic solutions. I would have hoped for results for various systems with and without dissipations, especially systems where Neural Networks are necessary to overcome the difficulty of analytical insight.
Theory. There is no theoretical contribution beyond textbook classical mechanics and the way to deal with damping in the Lagrangian and Hamiltonian system.

Other issues: The text is presented in an unusual way. The results sections are often a single sentence without describing the result. The neural network they use is never defined in the main text, and they describe implementation details (torch.autograd(), savgol_filter) as a stand in for the mathematics. Eq. 1 is stated without explanation/derivation.

Questions for authors

Clarification/Future work. Could you explain how your methodology with time-dependent Hamiltonian or Lagrangian Neural networks would improve the modeling of physical systems at scale? Is there a known gap where current method fails for these dissipative systems? Could you test your method on such systems?

Clarification. How does the approach compare with port-Hamiltonian or and classical Neural ODE ? Do you think your methods would outperform the use of these other formalisms?

Reviewer_ZKxDOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper applies standard Lagrangian NNs [1] and Hamiltonian NNs [2] to a 1D damped harmonic oscillator. Its textbook Lagrangian and Hamiltonian depend explicitly on time. Models are trained on 30–1,200 simulated trajectories, with optional Gaussian noise. Velocities and accelerations are obtained by finite differences. This paper includes several variants: 1) A Lagrangian NN (LNN) with and without time as an input; 2) Hamiltonian NN (HNN) with time and the generalized (canonical) momentum; 3) "Energy NN" with the linear momentum and no time input.

[1] Miles Cranmer, Sam Greydanus, Stephan Hoyer, Peter Battaglia, David Spergel, and Shirley Ho. Lagrangian neural networks. In ICLR 2020 Workshop on Integration of Deep Neural Models and Differential Equations, 2019.

[2] Samuel Greydanus, Misko Dzamba, and Jason Yosinski. Hamiltonian neural networks. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alche-Buc, E. Fox, and R. Garnett, editors, ´Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019.

Critical strengths
  • This paper uses a clear exposition of the physics.
  • The observation that the "energy nn" learns the non-dissipative energy rather than the Hamiltonian is interesting.
Critical weaknesses
  • Limited Novelty. Novelty. Both architectures are standard, the Caldirola–Kanai Lagrangian is textbook, and dissipative LNNs and HNNs have been studied [3]. The contribution is a demonstration on one toy system.
  • The toy system used in this work is narrow. One linear 1D system. There are no nonlinear or multi-dimensional systems.
  • The damping (d/m) in Eq. 21 is hand-coded, not learned.

[3] Andrew Sosanya and Sam Greydanus. Dissipative hamiltonian neural networks: Learning dissipative and conservative dynamics separately, 2022.

Questions for authors
  • Can you report quantitative error against baselines?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.