Imagine trying to map the blood pressure inside a patient's heart using only an external ultrasound video, or trying to calculate the exact momentum of a chaotic pendulum when you can only see its position through a noisy camera. In traditional machine learning, this is a nightmare of partial observability. If a neural network only sees variable $A$, it has no inherent reason to know anything about an unobserved, hidden variable $B$.
However, the physical world is not a collection of independent variables. It is governed by strict, immutable laws—such as the conservation of mass, momentum, and energy. This research direction focuses on a powerful paradigm: by embedding these physical laws directly into machine learning frameworks, we can accurately reconstruct unobserved, hidden physical states from sparse, partial observations.
Mathematical Coupling as an Information Bridge
Why does this work? The core intuition lies in mathematical coupling. Physical equations (such as partial differential equations) dictate that the behavior of an observable variable is directly tied to the state of an unobserved one. The observed data acts as a window into the hidden dynamics.
Consider fluid dynamics. Standard imaging might capture the movement of smoke or dye, but not the underlying velocity or pressure fields. As Karniadakis et al. point out, physics-informed models can infer these latent fields directly from visual observations because the fluid must obey the Navier-Stokes equations. Similarly, Chen et al. demonstrate that embedding discretized Navier-Stokes equations allows a network to resolve hidden velocity and pressure fields from simple passive observations.
[ Observed Data (e.g., Velocity) ]
│
▼ (Coupled via Physical Laws / PDEs)
[ Hidden State (e.g., Absolute Pressure) ]
This coupling means that the observed data constraints the search space of the hidden equations. For instance, in cardiology, Kissas et al. show that absolute arterial blood pressure can be derived from non-invasive 4D flow MRI velocity data. Because velocity and pressure are mathematically coupled within momentum equations, treating the physical equations as a regularizer forces the network to find the unique, physically consistent pressure field that satisfies the observed flow.
Resolving Unidentifiability via Inductive Biases
When data is sparse or incomplete, purely data-driven models suffer from unidentifiability—there are infinitely many mathematical trajectories that could perfectly fit the observed data points, yet most of them are physically impossible.
Physical constraints act as a severe regularizer, carving away these unphysical solutions. Ghanem et al. demonstrate that known physical relationships between measured and unmeasured variables can resolve model unidentifiability even when the exact underlying dynamics of the hidden states are completely unknown.
This is clearly illustrated in conservative systems. If a model only observes the position of a system, it shouldn't inherently know its momentum. However, by enforcing Hamiltonian dynamics, Ilin et al. show that a model can accurately reconstruct unobserved momentum trajectories from noisy position data because the system must conserve energy. The physical inductive bias provides the exact mathematical structure needed to fill the gaps.
This concept traces its roots back to early qualitative reasoning work, where Kenneth D. Forbus argued that the status of unmeasured components in a physical system can be accurately deduced by ensuring global consistency across known physical laws and observed trends.
Embedding Physics into Architecture and Priors
While adding a physics-based penalty to a loss function (like in standard Physics-Informed Neural Networks) is effective, a deeper approach embeds these constraints directly into the model's architecture or probabilistic priors. This ensures physical consistency by construction.
For example, Bergman et al. designed a grid-structured model for chemical reactors that mimics the topology of numerical PDE solvers, embedding conservation laws directly into the network to track unmeasurable states like catalyst degradation.
In a probabilistic framework, Tronarp handles partially observed systems by treating unobserved variables as latent states within a Gauss-Markov process, where the physics are baked directly into the prior. This allows the model to infer hidden components and estimate parameters where traditional gradient matching fails.
Similarly, Bacsa et al. introduce Symplectic Encoders that force Variational Autoencoders to obey Hamiltonian constraints, enabling the reconstruction of full system dynamics and hidden degrees of freedom from partial, noisy sensor data.
Transforming Sensing and Diagnostics
The practical implication of this paradigm is revolutionary: it turns software into a "virtual sensor," replacing invasive, expensive, or impossible physical measurements with algorithmic inference.
- Medicine: Alaa et al. use physics-informed self-supervised learning to estimate hidden physiological parameters like heart valve resistance from non-invasive ultrasound videos, eliminating the need for risky catheterization.
- Robotics: Clark et al. treat unmeasurable joint reaction forces as latent variables, using musculoskeletal simulations to infer internal stresses from simple wearable sensors like accelerometers.
- Infrastructure: Ban et al. combine traffic flow theory with mobile GPS traces to extract high-resolution urban infrastructure states, filling in the massive gaps where empirical data is missing.
By leveraging the mathematical rules that govern our universe, these models prove that we do not always need to see a hidden state to know exactly how it behaves.
Key Foundational Work
Measurement Interpretation in Qualitative Process Theory (1983) — contributed 2 concept elements to this direction.
Towards fine-grained urban traffic knowledge extraction using mobile sensing (2012) — contributed 1 concept elements to this direction.