When we observe the physical world—whether it is a pendulum swinging, water flowing through a catchment, or stars pulsing in the night sky—we are looking at the complex outputs of underlying mathematical rules. Traditionally, scientists have had to manually discover these rules, identifying constants like mass, gravity, or friction through meticulous experimentation and explicit parameter calibration.
A rapidly growing body of research takes a different approach: unsupervised latent physical parameter discovery. Instead of hardcoding physical equations, researchers train deep learning models on raw, unlabeled observational data (such as videos, time series, or sensor logs). Remarkably, to solve generic tasks like next-state prediction or data reconstruction, these models autonomously organize their internal latent spaces to mirror fundamental physical constants.
The core intuition behind why this works is simple: compression requires capturing invariants. To perfectly predict the future or compress a physical system, a network must separate what changes from what stays the same.
The Core Mechanism: Separating State from Parameter
Every dynamical system can be described by two types of variables:
- Dynamic States ($x_t$): Quantities that change at every time step (e.g., position, velocity).
- Static Parameters ($\theta$): Intrinsic properties that remain constant across a specific system or environment (e.g., mass, gravity, material stiffness).
When an unsupervised model is exposed to data from multiple environments, it faces an information bottleneck. If it tries to memorize every trajectory individually, it fails to generalize. To minimize error efficiently, the network is forced to disentangle the time-varying states from the time-invariant parameters.
Raw Observations (Videos, Sensor Logs) ──► [ Encoder ] ──► Latent Space Disentanglement:
├──► Dynamic States (Position, Velocity)
└──► Static Parameters (Mass, Stiffness, Gravity)
This exact phenomenon is demonstrated by Abitha Thankaraj et al. in Context is Everything: Implicit Identification for Dynamics Adaptation. By training a robot to predict its next state across different environments, the authors found that the learned latent vectors naturally clustered according to physical properties like surface material or object weight, despite receiving no explicit supervision or labels for those properties.
To make this separation explicit, architectural innovations like those in Modulated Neural ODEs introduce time-invariant "modulator" variables. These modulators capture hidden physical constants (such as frequency or mass) across varying trajectories, allowing the core differential equations to focus purely on time-dependent dynamics. Similarly, David Zheng et al. developed the Perception-Prediction Network in Unsupervised Learning of Latent Physical Properties Using Perception-Prediction Networks. They showed that when a model is optimized solely to predict future motion, a principal component analysis of its latent space reveals a near-perfect correlation with human-interpretable constants like mass and the coefficient of restitution.
Inductive Biases: Guiding the Latent Space
While pure prediction can surface physical parameters, researchers often introduce structural or geometric constraints—known as inductive biases—to force the latent space into a physically meaningful alignment.
Geometric and Manifold Constraints
Physical systems often move along continuous, low-dimensional surfaces (manifolds) within a high-dimensional observation space. In a foundational paper, Joshua B. Tenenbaum demonstrated in Mapping a Manifold of Perceptual Observations that non-linear dimensionality reduction (Isomap) can recover physical parameters like camera angles or object orientation directly from raw pixel data.
Modern iterations expand this to graph structures. In Latent Graph Inference using Product Manifolds, Haitz Sáez de Ocáriz Borde et al. showed that embedding data into complex product geometries allows inferred latent graphs to organize themselves automatically based on physical variables like velocity.
Smoothness and Invariance
The physical world is bound by continuity; objects cannot teleport. By enforcing a physical smoothness constraint on continuous hidden states, the framework proposed by David A. Nix et al. in Maximum-Likelihood Continuity Mapping (MALCOM): An Alternative to HMMs successfully reconstructed underlying physical dynamics (like articulatory tongue and jaw movements) directly from raw acoustic signals.
Another powerful constraint is property-specific bottlenecks. In Learning Conditional Invariance Through Cycle Consistency, Maxim Samarin et al. utilized cycle consistency to force a model to become invariant to specific redundant features. This constraint forces the remaining latent space to organize itself strictly according to natural generative factors, such as polar angles or radii.
Evidence Across Scientific Domains
The power of this approach is its universality. Because the underlying mechanism relies on the mathematical structure of compression and prediction, it functions across vastly different scales of reality.
Hydrology and Earth Systems
In macro-scale environmental modeling, Daniel Klotz et al. demonstrated in Rainfall–runoff modelling using Long Short-Term Memory (LSTM) networks that the internal cell states of an LSTM network autonomously track physical phenomena like snow accumulation and melting. The model was never given physical equations; it discovered these concepts purely by learning to predict river discharge from rainfall data.
At a landscape level, A. Bassi et al. showed in Learning landscape features from streamflow with autoencoders that latent features extracted from streamflow data align tightly with established physical metrics like the baseflow index and aridity, effectively discovering "landscape fingerprints" without manual feature engineering.
Quantum Mechanics and Particle Physics
At the subatomic scale, unsupervised models can discover abstract mathematical properties of matter. In Decoding conformal field theories: from supervised to unsupervised learning, Alireza Seif et al. showed that unsupervised autoencoders can automatically extract the "central charge"—a fundamental measure of a conformal field theory's degrees of freedom—directly from raw energy spectra.
Similarly, Mohamad Ali Marashli et al. in Identifying quantum phase transitions with minimal prior knowledge by unsupervised learning utilized non-linear dimensionality reduction on an autoencoder's bottleneck to show that the compressed latent space naturally clusters quantum states into distinct physical phases without any prior theoretical labels.
Astrophysics
On a cosmic scale, wide-field surveys generate massive amounts of periodic data. In Searching for Anomalies in the ZTF Catalog of Periodic Variable Stars, Ho-Sang Chan et al. proved that unsupervised models automatically organize complex periodic stellar data into clusters that align perfectly with fundamental physical parameters of stars, providing an automated path to astronomical discovery.
Why It Works: The Mathematical Imperative
Ultimately, latent physical parameter discovery works because the laws of physics are the most mathematically efficient compression algorithms that exist. When a neural network is forced to compress spatiotemporal data—whether through high-ratio masking, next-state prediction, or generative modeling—it discovers that the shortest path to minimal loss is to reinvent the constants of nature. By isolating these time-invariant properties into distinct latent dimensions, deep learning models do not just simulate physics; they independently discover its foundations.