Physics Informed Machine Learning with Unknown Governing Equations

101 concepts92 papers1 proposal

Physics-informed machine learning (PIML) has emerged as a powerful paradigm for modeling physical systems by embedding domain knowledge directly into neural network architectures. Typically, this integration relies on a strict mathematical formulation: a known partial differential equation (PDE) or ordinary differential equation (ODE) is used as a optimization constraint, penalizing the network whenever its predictions violate the underlying laws of nature.

However, this reliance introduces a severe bottleneck. In many real-world applications—ranging from complex industrial chemical processes to turbulent fluid dynamics and biological systems—the exact governing equations are either partially characterized, too complex to model analytically, or entirely unknown. As highlighted by Yusuke Tanaka et al. in Energy-consistent Neural Operators for Hamiltonian and Dissipative Partial Differential Equations, requiring an a priori exact mathematical description limits the utility of physics-informed models in poorly understood domains. When a model depends on a rigid, pre-defined mathematical template, any unmodeled latent dynamics or incorrect assumptions can cause the physics regularization to fail.

To overcome this limitation, a major research direction focuses on mitigating governing equation dependency. Rather than demanding explicit differential equations, modern frameworks learn physical dynamics by leveraging structural skeletons, learning pseudo-differential operators directly from data, or discovering the underlying equations entirely from scratch.

Bypassing Equations with Structural Skeletons and Invariants

If the exact differential equations governing a system are unknown, how can we inject physical intuition? The key insight is to shift focus from specific functional forms to universal physical principles, such as conservation laws, symmetries, and energy consistency. Even when we cannot write down the exact PDE for a system, we often know that mass must be conserved or that energy must dissipate over time.

One prominent strategy is to embed a general physical "skeleton" rather than a rigid equation. For instance, Pengwei Liu et al. developed PAPM: A Physics-aware Proxy Model for Process Systems, which integrates the general structure of conservation laws into a neural network. This architectural constraint ensures physical plausibility in complex industrial processes without requiring an exact mathematical description of every internal interaction.

Similarly, instead of forcing adherence to an explicit PDE constraint, networks can be designed to natively respect energy conservation or dissipation. The work by Yusuke Tanaka et al. demonstrates that a model can simultaneously learn an underlying energy functional and the solution operator from raw data. By enforcing energy consistency architecturally, their Energy-consistent Neural Operator (ENO) achieves high accuracy in data-scarce scenarios without ever requiring an explicit PDE formulation.

Taking this concept further, Takashi Matsubara et al. introduced FINDE: Neural Differential Equations for Finding and Preserving Invariant Quantities. FINDE automatically discovers and preserves "first integrals"—invariant quantities like total energy or momentum—directly from dynamical systems data. This eliminates the traditional requirement where a user must specify a Hamiltonian or Lagrangian structure before training begins.

Traditional PINN:
[Data] ───> [Neural Network] ───> [Compute Residual against EXACT PDE Template] ───> [Loss]

Equation-Agnostic PIML:
[Data] ───> [Neural Network with Invariant Layers] ───> [Conserve Energy/Mass Natively] ───> [Loss]

Learning Operators and Pseudo-Physics from Data

When governing equations are completely absent, another strategy is to let the neural network learn the differential operators themselves directly from observational data. Traditional physics-informed operator learning often requires a pre-calculated discretized operator based on the linear part of an equation. However, as noted by Peiyan Hu et al. in Deep Latent Regularity Network for Modeling Stochastic Partial Differential Equations, generating this operator becomes computationally or mathematically impossible when the underlying physics are unknown.

To bypass this limitation, Keyan Chen et al. proposed Pseudo-Physics-Informed Neural Operators: Enhancing Operator Learning from Limited Data (PPI-NO). Instead of relying on an explicit PDE to regularize training, PPI-NO trains a surrogate neural network to act as a local differential operator learned directly from limited data, providing its own physics-based regularization.

In a similar vein, Han Wan et al. developed a spectral-inspired operator network (SINO) in Spectral-inspired Operator Learning with Limited Data and Unknown Physics. SINO leverages the ability of spectral methods to capture global interactions and learns derivative operators in the frequency domain directly from a handful of trajectories. This allows the model to achieve strong out-of-distribution generalization without any prior PDE information.

Two-Step Discovery and Regularization

When the ultimate goal is still to utilize a standard physics-informed regularizer, researchers have turned to a two-step approach: discovering the governing equations from data first, and then enforcing them.

The authors of Physics-Informed Neural Networks with Unknown Partial Differential Equations: an Application in Multivariate Time Series demonstrate a framework that extracts PDEs from raw multivariate time series data using surrogate models and sparse regression. Once the physics are "discovered," they are immediately used to regularize a standard PINN, successfully injecting physics-based inductive biases into domains where the mathematical laws were previously unformulated.

Similarly, Wang Zhang et al. introduced ConCerNet: A Contrastive Learning Based Framework for Automated Conservation Law Discovery and Trustworthy Dynamical System Prediction. ConCerNet uses contrastive learning to automatically isolate hidden conservation laws from raw trajectory data and then employs a specialized projection layer to force the network's predictions to strictly obey these discovered laws.

Managing Imperfect and Parametric Priors

In many scenarios, we are not completely in the dark; we might possess an imperfect physical model or a parametric template. For instance, Florian Hofherr explores this in Neural Implicit Representations for Physical Parameter Inference from a Single Video, demonstrating how neural ordinary differential equations can infer specific physical parameters (like friction or length) given a parametric template, though the system still relies on that initial structural template.

To balance data-driven learning with potentially flawed or incomplete physical models, Agniva Chowdhury et al. proposed treating physics equations as generalized regularizers in Deep Learning with Physics Priors as Generalized Regularizers. By applying structural risk minimization, their framework automatically balances the influence of the data and the imperfect physics prior, preventing the model from being misled by incomplete governing equations.

Ultimately, the shift away from rigid governing equation dependency allows physics-informed machine learning to transition from a tool for accelerating known physics simulations to a flexible framework capable of modeling, understanding, and discovering the dynamics of complex, real-world systems.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_governing_equation_dependency_32ee56f7,
  title = {governing equation dependency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/governing-equation-dependency/rdr_32ee56f719ca4c919b074a6443a021b2},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_governing_equation_dependency_32ee56f7_v1,
  title = {governing equation dependency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/governing-equation-dependency/rdr_32ee56f719ca4c919b074a6443a021b2/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake