Traditional deep learning models for physical systems act as rigid specialists. They are typically trained to simulate a single, specific partial differential equation (PDE)—such as fluid flow under a fixed viscosity—and fail catastrophically when presented with a new equation or a novel combination of physical forces. This lack of generalization stems from treating complex spatiotemporal trajectories as monolithic entities to be memorized.
However, the physical universe is inherently compositional. Complex phenomena are built from a small dictionary of fundamental mathematical terms: advection, diffusion, reaction, and external forces. Instead of training a single neural network to model an entire multi-physics system, a powerful new paradigm focuses on compositional physical modeling. By isolating individual physical mechanisms into modular, reusable neural operators, researchers can assemble them on the fly to simulate entirely unseen, complex physical systems.
Deconstructing Physics into Atomic Operators
The core mechanism behind this approach is the decomposition of complex governing equations into isolated, manageable sub-units. Rather than learning a direct mapping from initial states to final states across a complex multi-physics field, models are trained on isolated physical phenomena.
For instance, Louis Serrano et al. and Li Sun et al. demonstrate that neural operators can be trained as a "dictionary of physical atoms," where individual models represent pure advection or pure diffusion within a shared latent space. Similarly, Alberto Bietti et al. show that pretraining neural surrogates on isolated mechanisms allows them to internalize the modular building blocks of physics, enabling compositional generalization when these mechanisms interact later.
This decomposition can also happen structurally within the network architecture itself. In Composing Partial Differential Equations with Physics-Aware Neural Networks, Matthias Karlbauer et al. introduce specialized neural modules that mirror numerical simulation components, separating flux kernels (for spatial transport) from state kernels (for local reactions). By isolating these terms, the model learns the underlying physical rules instead of memorizing trajectories, allowing it to generalize to entirely new boundary and initial conditions.
Monolithic Modeling (Standard AI Surrogates):
[Initial State] ───► [ Massive Black-Box Neural Network ] ───► [Predicted Trajectory]
(Fails on unseen physics/coefficients)
Compositional Modeling (Operator Splitting):
┌───► [ Advection Atom ] ───┐
[Initial State] ───► ├───► [ Diffusion Atom ] ───┼─► [Splitting/Coupling] ─► [Accurate Prediction]
└───► [ Reaction Atom ] ───┘
Beyond mechanism-specific splitting, decoupling can occur across structural and mathematical dimensions:
- Linear vs. Non-linear Decoupling: Vignesh Gopakumar et al. separate linear terms (handled via fixed finite-difference kernels) from non-linear terms (handled via specialized neural networks), achieving shift invariance across varying physical coefficients like viscosity without retraining.
- Spatio-temporal Decoupling: Chuizheng Meng et al. separate universal spatial derivative operators from task-specific temporal evolution functions, allowing spatial calculus to be reused across different phenomena. Similarly, Hrishikesh Viswanath et al. decouple spatial representation from the temporal evolution operator, allowing independent optimization or replacement of the physics solver.
- Response-Force Decoupling: Cristian Guarnizo et al. decouple the latent force kernel from the physical response function in Gaussian processes, allowing various ordinary differential equations to be incorporated modularly without changing the underlying approximation strategy.
Recomposition via Operator Splitting
How do we combine these isolated atomic operators to solve a complex, unseen equation at test time? The answer lies in classical operator splitting techniques (such as Lie or Strang splitting) adapted for neural network architectures.
In Test-time Generalization for Physics through Neural Operator Splitting, Louis Serrano et al. show that complex PDEs like Navier-Stokes can be solved zero-shot by sequentially applying pretrained atomic operators. Instead of retraining or fine-tuning via gradient descent—which is often unstable and data-hungry—the system simply sequences the advection operator, the diffusion operator, and other terms over small time steps. Because the underlying physics dictates that these operators approximate the full evolution when interleaved, the composition yields highly accurate simulations of unseen multi-physics equations.
This "plug-and-play" capability is echoed in the work of Zhiping Mao et al., who assemble pretrained DeepONets as modular components to predict coupled hypersonic flows and non-equilibrium chemistry. By treating neural operators as independent blocks, different physical domains can be coupled dynamically without rebuilding the model from scratch.
Preserving Physical Invariants and Conservation Laws
A major risk of modular composition is that connecting independently trained neural networks might introduce compounding errors that violate fundamental physical laws, such as the conservation of energy or mass. To prevent this, researchers embed strict physical frameworks directly into the composition mechanism.
One prominent strategy uses Port-Hamiltonian systems to govern the interfaces between modules. Jacob H Seidman et al. combine Gaussian Processes with Port-Hamiltonian structures, ensuring that when multiple probabilistic models are coupled to represent complex interconnected systems, the aggregate system remains physically consistent and retains properties like passivity and energy conservation.
For multi-physics thermodynamic systems, Muhammad Zakwan et al. leverage Irreversible Port-Hamiltonian Systems (IPHS) to interconnect smaller, physically consistent sub-units. This modularity allows the simulation to scale to larger systems while guaranteeing by design that thermodynamic properties and entropy production bounds are preserved across all boundaries.
Similarly, Suresh Bishnoi et al. enforce Hamiltonian mechanics within a modular graph neural network, allowing distinct learned physical behaviors (like gravity and spring forces) to be combined safely to simulate hybrid environments.
This structural grounding is equally effective in macro-scale engineering. Nathalie Cauchi demonstrates this by representing complex building automation systems as collections of interconnected thermal resistance-capacitance circuits. On an industrial scale, Yongjun Yin et al. decompose complex paper-drying processes into functional modules (evaporation, air mixing, steam heating) grounded in Fick's law and the Stefan equation, ensuring the entire coupled system remains interpretable and physically valid.
Why It Works: The Power of Inductive Modularity
Compositional physical modeling succeeds because it shifts the learning burden from trajectory memorization to operator identification. When a model is forced to learn a specific multi-physics simulation holistically, it overfits to the precise interaction balance of that training set.
By forcing the architecture to decouple forces—whether through mathematical terms (Kazem Meidani et al.), structural kernels (Matthias Karlbauer et al.), or interconnected physical zones (Nathalie Cauchi)—the neural components capture the invariant laws of nature. At test time, the complexity of the universe is handled not by a larger neural network, but by the combinatorial richness of how these atomic components are arranged.
Key Foundational Work
Fast Kernel Approximations for Latent Force Models and Convolved Multiple-Output Gaussian processes (2018) — contributed 1 concept elements to this direction.
Data-driven Identification of 2D Partial Differential Equations using extracted physical features (2020) — contributed 1 concept elements to this direction.
Benchmarks for stochastic models from building automation systems (2018) — contributed 1 concept elements to this direction.