Structuring Latent Spaces with Differentiable Physical Models

42 concepts34 papers1 proposal

Standard autoencoders are highly effective at compressing high-dimensional data into a low-dimensional bottleneck. However, the resulting latent spaces are typically abstract, unconstrained, and opaque. Because the network is free to organize its internal representations arbitrarily, human experts cannot easily interpret or validate the model's reasoning.

To bridge the gap between black-box optimization and scientific rigor, a powerful paradigm has emerged: physics-informed latent representations. Instead of allowing a neural network to invent arbitrary statistical features, researchers force the latent space to map directly to explicit physical parameters—such as mass, velocity, temperature, or the coefficients of a differential equation.

Understanding why this approach works requires looking at how physical constraints alter the flow of gradients during training.

Inverting Physics via Model-Based Decoders

The most direct way to ground a latent space is to replace the traditional neural network decoder with a non-trainable, deterministic, or differentiable physical forward model.

In a standard autoencoder, both the encoder and decoder learn simultaneously. If you replace the decoder with a fixed mathematical function derived from the laws of physics, the encoder is forced to adapt to the rules of that specific physical process. To minimize reconstruction loss, the encoder must learn to act as an inversion mechanism, mapping raw inputs directly to the exact physical parameters the decoder expects.

Raw Input (x) ---> [ Learnable Encoder ] ---> Physical Latent Parameters (z) ---> [ Fixed Physical Decoder ] ---> Reconstructed Input (x̂)

As demonstrated by Wong et al. and Weißer et al., this architectural constraint forces the latent space to represent explicit real-world variables, such as signal angles, noise levels, or material properties, rather than arbitrary statistical patterns. Basargin and Alonso-González show that by backpropagating reconstruction errors directly through established scientific equations, the encoder successfully learns to predict physically meaningful latent parameters entirely without labeled ground-truth data.

This mechanism can be adapted across diverse domains by swapping the physical decoder:

  • Hyperspectral Imaging: Welner integrated a linear physical mixing model into a neural decoder, forcing the latent space to extract pure material signatures from raw imagery.
  • Signal Processing: Baur et al. parameterized a decoder with domain-specific transforms, such as oversampled Fourier matrices, enabling the unsupervised discovery of signal source counts by forcing the latent space to capture spatial power spectral density.

Governing Latent Dynamics with Differential Equations

Physical systems evolve over time according to strict mathematical laws. Instead of allowing a latent space to transition arbitrarily from one state to the next, we can embed differential equations directly into the latent dynamics.

de Bezenac et al. established that structuring neural networks around the analytical solutions of differential equations produces interpretable latent representations, such as explicit velocity fields, which domain experts can visually validate. Building on this, THOME and Le Guen applied partial differential equations (PDEs) directly to learned latent features rather than raw pixels. This abstraction allows the system to ignore pixel-level noise and focus entirely on the fundamental physical laws governing the scene.

This structural grounding can take several forms depending on the underlying physics of the system:

  • Modal Analysis: Lai et al. mapped sparse sensor observations to a latent space where each dimension corresponds to a specific physical vibration mode, ensuring the dynamics are governed by structural properties.
  • Hamiltonian Mechanics: Bacsa et al. embedded Hamiltonian constraints directly into a variational autoencoder (VAE) architecture, forcing the latent space to be a linear transformation of the physical phase space (position and momentum coordinates).
  • Mechanistic Physiology: Wang et al. structured a hybrid VAE's latent space to represent the coefficients and initial states of a governing physiological differential equation, extracting rate constants and baseline levels directly from noisy observations.
  • Equivalent Circuits: Farsang and Grosu derived recurrent neural architectures from the principles of electrical equivalent circuits, creating a direct link that aligns latent neural activity with observable physical trajectories.

Disentangling Intrinsic Physics from Environmental Factors

A major challenge in physical modeling is separating the intrinsic properties of an object from extrinsic noise or unmodeled environmental factors. A physics-constrained latent space allows for explicit disentanglement by allocating specific dimensions to known physical constants.

For instance, Aleo et al. integrated a physics-based layer into a VAE to explicitly account for extrinsic factors like redshift and interstellar extinction. This architecture enables the neural network to isolate and learn the intrinsic physical properties of astronomical transients rather than just their observed appearance.

Supervising or partitioning the latent space ensures that known physical constants are cleanly separated from abstract features:

                  +---> Dimension 1: Mass (Supervised/Constrained)
                  |
Latent Space (z) -+---> Dimension 2: Friction (Supervised/Constrained)
                  |
                  +---> Dimensions 3-N: Unmodeled Residual Features (Abstract)

Fotiadis et al. demonstrated that supervising specific dimensions of a latent space with ground-truth parameters like mass or length results in a highly linear and interpretable representation of the system's physics. Similarly, Kit et al. forced specific dimensions to map directly to known control parameters, allowing a portion of the latent space to function as a verifiable white-box model while the remaining dimensions captured complex, unmodeled features.

This explicit structuring also provides a sharper boundary for anomaly detection. Lin constrained an autoencoder's latent space to represent known physical parameters like temperature or chemical composition, creating a physics-informed bottleneck that learns a highly structured representation of the normal class, making it easier to isolate anomalous signals.

Why It Works: The Power of Differentiable Mapping

The unifying reason these architectures succeed is that they preserve the end-to-end differentiability of deep learning while enforcing rigid scientific boundaries. As Chen et al. noted, using differentiable architectures demonstrates that abstract latent spaces can be successfully mapped back to concrete physical properties, allowing researchers to deduce the physical meaning of discovered coordinates through latent traversal (Liu et al.).

By replacing abstract feature vectors with explicit parameters like object mass, velocity, and friction coefficients, models achieve far better explainability and generalization to unseen force magnitudes (Kandukuri et al.). Incorporating explicit physical variables alongside abstract learned features ensures that surrogate models remain compatible with existing simulation frameworks (Xie and Tang).

Whether constraining latent variables to represent probability distributions over physical states (Arcieri et al.), grounding them in biological properties like synaptic weights (Bellec et al.), constraining them via historical observations (Meng et al.), grounding them in degradation physics (Wang et al.), or pre-structuring data with fluid mechanics equations (Halder et al.), the underlying principle remains identical: forcing deep learning to speak the language of physics inside its hidden layers bridges the gap between statistical correlation and physical reality.

Key Foundational Work

Deep Learning for Physical Processes: Incorporating Prior Scientific Knowledge (2018) — contributed 1 concept elements to this direction.
Training Auto-encoder-based Optimizers for Terahertz Image Reconstruction (2019) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_physics_informed_latent_representation_79c5a4c0,
  title = {physics-informed latent representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/physics-informed-latent-representation/rdr_79c5a4c0f554484e9952b45d43a63b21},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_physics_informed_latent_representation_79c5a4c0_v1,
  title = {physics-informed latent representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/physics-informed-latent-representation/rdr_79c5a4c0f554484e9952b45d43a63b21/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake