In many scientific domains, we can only measure a fraction of what we care about. A cardiologist can see the velocity of blood flow through an MRI but cannot directly measure the pressure deep inside an artery. A neuroscientist can record the voltage of a neuron but cannot see the microscopic "gates" in ion channels opening and closing. Traditionally, these hidden variables were treated as "missing data" or required invasive sensors to capture.
A growing body of research demonstrates that we don't always need more sensors. Instead, we can use the mathematical laws of the universe to "see" the unobservable. By embedding physical constraints—like conservation of mass, momentum, or energy—directly into machine learning models, researchers are now reconstructing latent physical states from partial, noisy, and indirect observations.
The Principle of Physical Coupling
The core intuition behind this research is coupling. In a physical system, variables are rarely independent; they are linked by differential equations. If you know the value of variable $A$ and you know the mathematical rule that relates $A$ to $B$, you can often uniquely determine $B$.
As George Em Karniadakis and colleagues explain in their review of Physics-informed machine learning, this allows models to infer latent quantities like pressure and velocity fields directly from visual observations, such as the movement of smoke or temperature gradients. The physics acts as a "bridge" that carries information from the observed world to the hidden one.
For example, in fluid dynamics, the Navier-Stokes equations link velocity and pressure. Maziar Raissi and George E. Karniadakis demonstrated in Physics Informed Deep Learning (Part II) that a neural network constrained by these equations can accurately reconstruct a full pressure field using only velocity measurements. The network isn't just "guessing" the pressure; it is finding the only pressure distribution that is mathematically consistent with the observed flow.
Resolving Unidentifiability
Without physics, reconstructing hidden states is often an "ill-posed" problem—there are infinitely many ways a hidden variable could behave that would still result in the same observed data. This is known as unidentifiability.
Paul Ghanem and Ahmet Demirkaya address this in Learning Physics Informed Neural ODEs with Partial Measurements. By enforcing structural constraints within a Neural Ordinary Differential Equation (NODE), they show that known physical relationships can resolve this ambiguity. Even when the underlying dynamics of the hidden states are unknown, the requirement that they must "fit" into the broader physical architecture allows the model to reconstruct latent trajectories that would otherwise be mathematically invisible.
Cross-Modal Inference
This approach enables a form of "cross-modal" sensing, where one type of sensor effectively acts as another.
- Electricity to Magnetism: Marc Harkonen and Markus Lange-Hegermann developed EPGP, a Gaussian Process framework that uses the coupled nature of Maxwell’s equations to infer magnetic field components from electric field measurements.
- Temperature to Flow: Nagahiro Ohashi and Beomjin Kwon used MSPINNs to reconstruct full velocity and pressure fields from sparse, noisy temperature data.
- Acceleration to Displacement: In structural engineering, Calvin Alvares and Souvik Chakraborty showed that governing equations discovered from acceleration-only measurements can accurately predict unobserved displacement and velocity.
Embedding Physics into the Architecture
How is this actually implemented? There are three primary strategies emerging across the literature:
- Physics-Informed Loss Functions: The most common method, popularized by PINNs, involves adding a "physics loss" term to the training objective. The network is penalized if its predictions violate a differential equation. This is used by Nicola Toschi to infer hidden ion channel gating from raw neuronal spikes.
- Differentiable Physics Backbones: Instead of just a loss term, the model itself is built around a physical solver. Dapeng Feng and colleagues used this in hydrologic simulations, where a model trained only on streamflow could accurately estimate unobserved evapotranspiration because the internal architecture maintained a strict water balance.
- Structured Latent Spaces: Some models bake physics into the "bottleneck" of an autoencoder. Ruochen Jiao and Qi Zhu developed LK-SDE, which forces the latent space of a Variational Autoencoder to follow vehicle kinematics. This allows the model to "discover" unobserved variables like steering and slip angles from simple trajectory data.
Beyond Prediction: Discovering Mechanisms
Perhaps the most exciting facet of this direction is its ability to reveal "missing physics." In climate modeling, Ali J. Ramadhan and John C. Marshall used Neural Differential Equations to capture unobserved physical fluxes that traditional models omit. Similarly, Brandon C Reyes and Alexandre M. Tartakovsky used PINNs to discover the hidden viscosity of non-Newtonian fluids using only velocity data.
By treating physical laws not as rigid barriers, but as informative priors, these researchers are turning machine learning into a tool for scientific discovery—allowing us to reconstruct the hidden machinery of the world from the sparse clues it leaves behind.