Learning Composable Physics via Modular Energy Decomposition

26 concepts21 papers1 proposal

Standard deep learning treats physical systems as black boxes, mapping current states directly to future states. While highly expressive, these models fail in catastrophic ways: they violate conservation laws, require massive datasets, and completely break down if you change the system's size or configuration (such as adding a third pendulum to a double pendulum).

Instead of learning the dynamics directly, a powerful alternative is to use neural networks to learn the underlying scalar energy functions of individual components, and then let classical physics handle the rest. This is the core intuition behind modular energy decomposition. Because energy is a scalar quantity that naturally sums across independent subsystems, modeling physics through decomposed energy functions unlocks unprecedented composability, zero-shot scalability, and physical consistency.

The Disentanglement Intuition: Separating Kinetic and Potential Energy

When a standard neural network models a physical system, it mixes position and velocity into a single latent representation. This often leads to learning spurious cross-terms that violate fundamental physics.

In a comprehensive benchmark of physics-informed architectures, Abishek Thangamuthu et al. demonstrated in Unravelling the Performance of Physics-informed Graph Neural Networks for Dynamical Systems that explicitly decoupling kinetic and potential energy into separate neural network modules is the single most critical factor for preserving physical quantities like momentum. By preventing the network from blending position and velocity, the model remains stable over long-term simulations.

Standard Black-Box:  [Position, Velocity] ---> [Neural Network] ---> [Next State] (Violates Conservation)

Modular Energy:      [Velocity] ---> [Kinetic Net]   ---\
                                                         +---> [Total Energy] ---> [Physics Engine]
                     [Position] ---> [Potential Net] ---/

This structural separation allows models to perform unsupervised physical disambiguation. For instance, Michael Lutter et al. introduced Deep Lagrangian Networks (DeLaN) in Deep Lagrangian Networks: Using Physics as Model Prior for Deep Learning, which use distinct network heads to independently learn the mass matrix (kinetic energy) and potential energy vector. By constraining the network to satisfy the Euler-Lagrange equations, the model can autonomously disentangle individual forces—such as separating gravity from Coriolis effects—using only a single total torque signal as supervision.

In subsequent work, the authors leveraged this unsupervised decomposition in Deep Lagrangian Networks for end-to-end learning of energy-based control for under-actuated systems to apply energy-based control laws directly to under-actuated robotic systems without manual physics derivations. Similarly, Yaofeng Desmond Zhong et al. (SymODEN) in Symplectic ODE-Net: Learning Hamiltonian Dynamics with Control and Tohru Yoshioka et al. in Vision-Based Modeling and Control of Dynamical Systems Using Deep Learning showed that embedding this kinetic-potential split within Hamiltonian or port-Hamiltonian latent spaces acts as a powerful regularizer, enabling data-efficient learning directly from raw visual inputs while guaranteeing physically plausible trajectories.

The Power of Additivity: Zero-Shot Scalability via Graphs

The true magic of energy-based modeling lies in a simple mathematical property: additivity. The total energy of a system is simply the sum of the energies of its constituent parts. By combining this property with Graph Neural Networks (GNNs), researchers can train models on tiny systems and deploy them on massive ones without retraining.

For example, Ravinder Bhattoo et al. developed the Lagrangian Graph Neural Network (LGNN) in Learning Articulated Rigid Body Dynamics with Lagrangian Graph Neural Network, which assigns independent neural networks to learn local potential and kinetic energy functions across nodes and joints. By aggregating these local components to form a total system Lagrangian, a model trained on a simple 4-link chain can simulate a 100-link chain out-of-the-box.

This node-and-edge decomposition extends naturally to particle systems. In Learning the dynamics of particle-based systems with Lagrangian graph neural networks, Ravinder Bhattoo et al. mapped kinetic energy to nodes and potential energy to edges, allowing models trained on small particle groups to scale to systems two orders of magnitude larger. Furthermore, Suresh Bishnoi et al. demonstrated in Discovering symbolic laws directly from trajectories with hamiltonian graph neural networks that enforcing this Hamiltonian graph structure makes the learned energy landscape so clean that symbolic regression can directly extract exact mathematical laws (like the Lennard-Jones potential) from the trained network weights.

To generalize across topologies, Yupu Lu et al. (ModLaNets) in ModLaNets: Learning Generalisable Dynamics via Modularity and Physical Inductive Bias used independent modules for each element's energy, summing them to construct the global Lagrangian. This allowed a seamless transition from a double-pendulum to a quintuple-pendulum configuration without global retraining.

Snapping Subsystems Together Like LEGO Bricks

Beyond particles and chains, modular energy decomposition enables macro-scale engineering systems to be assembled compositionally.

Alex Beatson et al. introduced Composable Energy Surrogates (CES) in Learning Composable Energy Surrogates for PDE Order Reduction for simulating complex mechanical meta-materials. Instead of running expensive global Finite Element Analysis (FEA), they trained neural networks to predict the potential energy of individual modular components based on their boundary states. To simulate a massive, novel assembly, they simply sum these individual energy surrogates and perform global energy minimization, maintaining strict physical consistency across the entire structure.

[Component A Energy Net] ---\
[Component B Energy Net] ----+---> Summed Total Energy ---> Global Minimization ---> Simulation State
[Component C Energy Net] ---/

For multi-agent or multi-domain systems, this plug-and-play capability is invaluable. Armand Comas et al. modeled multi-agent interactions in Inferring Relational Potentials in Interacting Systems as a sum of independent relational potentials, allowing hybrid agent behaviors to be simulated by mixing and matching potentials trained on entirely different datasets.

In the domain of control engineering, Cyrus Neary et al. showed in Compositional Learning of Dynamical System Models Using Port-Hamiltonian Neural Networks that complex systems can be modeled by aggregating independently trained port-Hamiltonian submodels. By summing their energy functions while arranging their dissipation and control matrices block-diagonally, they predicted large-scale system behavior without requiring any training data from the fully assembled system.

Beyond Ideal Physics: Dissipation, Noise, and Baselines

Real-world systems are rarely perfectly conservative; they suffer from friction, noise, and complex boundary conditions. Modular decomposition handles this by separating the conservative energy core from non-conservative effects.

Kevin Course et al. formulated Generalized Hamiltonian Neural Networks (GHNNs) in Weak Form Generalized Hamiltonian Learning using a weak form loss function. They decomposed the system's state evolution into a skew-symmetric component (for conservative motion) and a symmetric component (for dissipative effects), both driven by the gradient of a single learned energy function. This allows the network to isolate energy-preserving forces from friction. Similarly, Razmik Arman Khosrovian et al. (PoDiNNs) used distinct networks in Poisson-Dirac Neural Networks for Modeling Coupled Dynamical Systems across Domains to separate energy storage from resistive dissipation across coupled electrical, hydraulic, and mechanical domains.

When dealing with noisy or multi-fidelity data, modularity serves as a filter. Carlo Lucibello et al. split a system's total energy in Reconstruction of Pairwise Interactions using Energy-Based Models into an explicit, interpretable pairwise interaction term and a flexible neural network term. The neural network "soaks up" complex higher-order noise, leaving the pairwise term clean and accurate.

In quantum chemistry, Bowen Deng demonstrated in Cross-functional transferability in foundation machine learning interatomic potentials that modeling total potential energy as a combination of a linear per-element baseline and a non-linear interaction term isolates arbitrary energy shifts from structural physics, boosting data efficiency 10-fold.

Finally, this decoupling strategy extends to operators and solvers. Steindór Sæmundsson et al. (Variational Integrator Networks) isolated the potential energy landscape into a standard neural network in Variational Integrator Networks for Physically Structured Embeddings while the broader architecture strictly enforced geometric integration logic.

Yusuke Tanaka et al. utilized a dual-network setup for PDEs in Energy-consistent Neural Operators for Hamiltonian and Dissipative Partial Differential Equations, where one network maps initial conditions to solutions and a second network approximates the energy density.

In industrial applications like smart grids, Julen Cestero et al. showed in Optimizing energy management of smart grid using reinforcement learning aided by surrogate models built using physics-informed neural networks that separating power generation components from global balance equations allows physics-informed surrogates to accelerate reinforcement learning training by 50% while strictly respecting physical constraints.

Key Foundational Work

Learning Composable Energy Surrogates for PDE Order Reduction (2020) — contributed 2 concept elements to this direction.
Variational Integrator Networks for Physically Structured Embeddings (2020) — contributed 1 concept elements to this direction.
Deep Lagrangian Networks: Using Physics as Model Prior for Deep Learning (2019) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_modular_energy_decomposition_229bbabf,
  title = {modular energy decomposition},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/modular-energy-decomposition/rdr_229bbabf42f9416399119f41aa1e4678},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_modular_energy_decomposition_229bbabf_v1,
  title = {modular energy decomposition},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/modular-energy-decomposition/rdr_229bbabf42f9416399119f41aa1e4678/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake