Traditional scientific computing relies on "white-box" models—mathematical equations like Partial Differential Equations (PDEs) that describe the world with perfect consistency but struggle with noisy, real-world data. Conversely, modern deep learning offers "black-box" models that excel at pattern recognition but often produce physically impossible results, such as fluid flowing through solid walls or energy appearing out of nowhere.
Physics-Informed Machine Learning (PIML) bridges this gap. By embedding governing equations directly into the learning process, researchers create models that are both data-efficient and physically grounded.
The Loss Function as a Physical Constraint
The most common way to "teach" physics to a neural network is through the loss function. In a standard network, the loss measures the distance between a prediction and a label. In a Physics-Informed Neural Network (PINN), as popularized by Maziar Raissi in Data-driven Discovery of Nonlinear Partial Differential Equations, the loss function includes an extra term: the residual of a physical equation.
Imagine training a network to predict the temperature in a room. Without physics, the network might predict a sharp temperature spike in a corner just because the training data was noisy. A PINN, however, evaluates the heat equation at that point. If the predicted temperature gradient violates the laws of thermodynamics, the "physics loss" increases, penalizing the network. This allows models to learn from sparse or noisy data because the physics acts as a global regularizer, filling in the gaps where data is missing.
Hardwiring Physics into Architecture
While loss-based constraints are "soft" (the network can still technically violate them), some researchers prefer "hard" constraints built directly into the network's architecture.
For example, Johannes Hendriks and colleagues in Linearly Constrained Neural Networks design networks where every possible output is guaranteed to satisfy linear operator constraints, such as being divergence-free. Similarly, S. Mohammad H. Hashemi demonstrated that physical laws can be converted into specific convolutional kernels. In Deep-learning PDEs with unlabeled data and hardwiring physics laws, the network doesn't just "try" to follow the physics; its internal structure makes it mathematically impossible to do otherwise.
Latent Dynamics and Conservation Laws
For complex systems like fluid dynamics, modeling every pixel is computationally expensive. Researchers like Kookjin Lee and Kevin T. Carlberg use autoencoders to compress high-dimensional states into a small "latent space."
In Deep Conservation, they introduce a framework that ensures these latent-space dynamics exactly satisfy physical conservation laws (like mass or momentum). By solving a constrained optimization problem at each time step, the model ensures that even as it simplifies the data, it never loses the fundamental "accounting" that makes physics reliable.
Discovering the Unknown
PIML isn't just for simulating known laws; it's also for discovering new ones. Yiping Lu and Zichao Long developed PDE-Net, which uses filters that correspond to differential operators. By training on observational data, the network "learns" which operators best describe the system's evolution, effectively writing the underlying PDE from scratch.
This "inverse modeling" is particularly powerful for identifying hidden forces. Priyabrata Saha and Saibal Mukhopadhyay showed in PhICNet that by embedding PDEs into recurrent neural networks, one can reconstruct "hidden" external forces—like an unmeasured wind gust pushing a drone—simply by seeing how the system deviates from the expected physics.
The Cost of Knowledge
The primary limitation across this field is the "requirement of prior knowledge." As noted in work ranging from HybridNet to PI-WNO, these models generally require the functional form of the governing equations to be known beforehand.
If the underlying physics are poorly understood—such as in complex biological processes or "black-box" industrial systems—the very constraints that make PIML powerful become a bottleneck. The field is currently moving toward more flexible "meta-learning" approaches, such as LeMON, which attempt to generalize across different families of equations, reducing the need for a perfectly defined mathematical model from the start.