Emergent Physical Representations in Latent Spaces

This direction is split into other directions: latent physical parameter discovery. Its last version stays readable and citable here.

Version 1 of 1245 papers271 conceptsfrozen 2026-10-04

When a neural network is tasked with predicting the future of a physical system or compressing a complex observation, it often does something remarkable: it "reinvents" the laws of physics. Without being told about gravity, mass, or velocity, models trained on raw data—from video frames of falling blocks to the hit patterns in particle accelerators—naturally organize their internal representations to mirror the underlying physical reality.

This research direction explores how unsupervised and self-supervised learning objectives act as a crucible, forcing models to discover fundamental physical principles as the most efficient way to solve a task.

The Bottleneck as a Discovery Tool

The most intuitive way to understand this is through the "bottleneck" of an autoencoder. To compress high-dimensional data (like a 1024x1024 image of a galaxy) into a small latent vector, the model must discard everything except the most essential information.

In astronomy, Ho-Sang Chan and V. Ashley Villar demonstrated in Searching for Anomalies in the ZTF Catalog of Periodic Variable Stars that Variational Autoencoders (VAEs) naturally cluster stars into known physical categories. The model doesn't see "labels"; it sees patterns in light curves that are only explainable by physical properties like stellar period or temperature. Similarly, Cristiano Fanelli showed in DeepRICH that a VAE can organize raw detector hits into a latent space where different particle species are spatially separated, effectively rediscovering particle identity from raw geometry.

Dynamics and the Necessity of Physics

If you want to predict the next frame of a video, you have to understand how things move. This requirement makes "future prediction" a powerful unsupervised task for learning physics.

In Generalizable Features From Unsupervised Learning, Aaron Courville and Yoshua Bengio found that models trained to imagine how a tower of blocks falls develop an internal grasp of gravity and structural support. Because these principles are universal, the model can generalize to towers it has never seen before.

This concept is taken further by Maximilian Karl in Deep Variational Bayes Filters. By forcing a latent space to follow strict Markovian transition rules, the model is pressured to find the "true" coordinates of the system. If a ball is moving in a circle, the model might autonomously discover polar coordinates because they make the transition math simpler than Cartesian coordinates. In this view, physical laws are not just constraints we impose; they are the "path of least resistance" for a model trying to minimize prediction error.

Multimodal Alignment and Physical Symmetry

Another powerful mechanism for discovering physics is cross-modal alignment. If you have two different ways of looking at the same object—say, an image of a galaxy and its light spectrum—the only thing they truly share is the underlying physics of the galaxy itself.

Francois Lanusse and colleagues demonstrated this with AstroCLIP. By using contrastive learning to align images and spectra, the model creates a shared latent space organized by physical properties like redshift and stellar mass. Because the model must find a representation that works for both "views," it discards view-specific noise and hones in on the invariant physical truth.

A similar effect occurs in materials science. In Multimodal Learning for Crystalline Materials, Charlotte Loh and Peter Y. Lu showed that aligning crystal structures with their electronic properties causes the latent space to cluster materials by symmetry and conductivity. The model "learns" chemistry by trying to bridge the gap between how a material looks and how it behaves.

From Latent Vectors to Symbolic Laws

The ultimate goal of this research is not just to have a "smart" latent space, but to extract human-readable science from it. Researchers are now bridging the gap between the continuous vectors of deep learning and the discrete equations of classical physics.

In Integration of Neural Network-Based Symbolic Regression, Li Jing and Marin Soljacic showed that neural encoders can discover latent variables that correspond exactly to physical constants. These variables can then be fed into "Equation Learner" (EQL) networks to formulate symbolic governing laws.

This suggests a new pipeline for scientific discovery:

  1. Observe: Collect raw, unlabeled data.
  2. Compress/Predict: Use self-supervised learning to force the model to find an efficient latent representation.
  3. Interpret: Analyze the latent space to find clusters or gradients that correlate with physical reality.
  4. Symbolize: Use symbolic regression to turn those latent relationships into mathematical equations.

By treating the neural network as a "feature discoverer" rather than just a classifier, researchers like Taoli Cheng have shown that we can "pull back the curtain" on the black box. In Interpretability Study on Deep Learning for Jet Physics, Cheng demonstrated that networks trained on particle data naturally reinvent established observables like jet mass, proving that the model's internal logic is often just a rediscovered version of our own physical laws.

Papers in this version

Showing 60 of 245 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_emergent_physical_representations_in_lat_ae7a3017,
  title = {Emergent Physical Representations in Latent Spaces},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/emergent-physical-representations-in-latent-spaces/rdr_ae7a3017fcec403b98b5f456378cd426},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_emergent_physical_representations_in_lat_ae7a3017_v1,
  title = {Emergent Physical Representations in Latent Spaces},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/emergent-physical-representations-in-latent-spaces/rdr_ae7a3017fcec403b98b5f456378cd426/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}