Physics-Inspired Graph Neural Networks

• Physics-inspired GNNs are graph neural network models that integrate physical laws, such as conservation and symmetry, directly into their architectures.
• They employ inductive biases like object-centric compositionality and physics-informed message-passing to enhance generalization, convergence, and interpretability.
• These models are applied in domains such as dynamical systems, high-energy physics, and combinatorial optimization, demonstrating robust scalability and precise system control.

Physics-inspired Graph Neural Networks (GNNs) constitute a specialized and rapidly maturing family of models that embed physical knowledge—including symmetries, conservation laws, and relational structure—directly into graph-based deep learning architectures. These frameworks span supervised, semi-supervised, and unsupervised pipelines, targeting domains where physical systems or constraints are naturally represented by graphs: dynamical systems, high-energy physics, combinatorial optimization, soft-sensing, and beyond. Core principles include the construction of object- and relation-centric inductive biases, explicit or implicit enforcement of physical laws (e.g., energy conservation, mass continuity), and the leveraging of message-passing schemes closely mirroring real-world physical interactions.

1. Inductive Biases and Architectural Principles

Physics-inspired GNNs are characterized by strong, physically-motivated inductive biases. These take several interrelated forms:

• Object- and relation-centric compositionality: Nodes correspond to discrete physical entities (e.g., bodies, grains, sensors, detector hits), and edges model interactions (e.g., forces, joints, couplings). Node-update and edge-update functions share weights across the graph, encoding the fundamental assumption of local, homogeneous, and compositional laws (Sanchez-Gonzalez et al., 2018, Thangamuthu et al., 2022, Cranmer et al., 2019).
• Symmetry invariance: GNN layers and input features are often constructed to respect physical symmetries, such as permutation invariance, rotation or Lorentz invariance, and conservation under various transformations. This is critical in domains like particle physics reconstruction, where Lorentz symmetry and object set-invariance are essential (Shlomi et al., 2020, Thais et al., 2022).
• Physics-inspired message-passing: Message functions may be designed such that their outputs correspond, up to an invertible linear transformation, to real physical quantities (e.g., force vectors in $D$-dimensional space), enabling identification and even symbolic recovery of underlying laws (Cranmer et al., 2019).
• Constrained dynamics: Physically-motivated GNNs often impose explicit holonomic or energetic constraints — such as enforcing mass conservation at nodes, or using Lagrangian/Hamiltonian formalisms for constrained mechanical systems — directly in their differentiable forward pass or loss function (Thangamuthu et al., 2022, Botta et al., 11 Dec 2025).

These inductive biases increase data efficiency, accelerate convergence, support out-of-domain generalization, and facilitate the extraction of human-interpretable, physically meaningful latent representations.

2. Physics-inspired GNN Designs for Dynamical Systems

The simulation and inference of dynamical systems are prominent focuses for physics-inspired GNNs. Key paradigms include:

• Learnable Physics Engines: Systems are modeled as graphs with nodes encoding per-object state (e.g., position, velocity) and edges parameterizing interactions (e.g., joint type, force). Graph blocks implement rounds of edge update, node update, and global update, with deep or recurrent stacking to enable accurate forward modeling, system identification, and differentiable control/planning (e.g., model-predictive control, value-gradient RL) (Sanchez-Gonzalez et al., 2018).
• Hamiltonian and Lagrangian GNNs: The GNN learns to output a Hamiltonian $H(q,p)$ or Lagrangian $L(q,\dot q)$ as a graph-structured function, and physical evolution is integrated using symplectic or Euler-Lagrange equations, often with explicit constraint layers for holonomic or energetic conservation (Thangamuthu et al., 2022). Decoupling kinetic and potential energies (separate GNN “heads” for $T$ and $V$) and embedding constraints (projection onto constraint manifolds) significantly sharpen long-term prediction, conservation properties, and zero-shot generalization to system sizes far exceeding the training domain.

Framework	Physical Quantity Modeled	Key Inductive Bias	Constraint Handling
GN Physics Engine	Transition model $\Delta n$	Node/edge sharing, graph	Optional recurrent ID phase
H/L-GNN (Hamiltonian)	$H(q,p)$, $L(q,\dot q)$	Explicit conservation	Constraint projectors

Empirically, such models match or surpass classical baselines on long-horizon rollout accuracy, energy/momentum conservation, and robust system identification across diverse mechanical systems (e.g., pendulums, acrobots, walkers, robotic arms), demonstrating "zero-shot" combinatorial generalization to new body counts and connectivity (Sanchez-Gonzalez et al., 2018, Thangamuthu et al., 2022).

3. Embedding Physical Laws for Discovery, Control, and Interpretability

Physics-inspired GNNs are leveraged not only for data fitting but also for discovering governing laws, symbolic model extraction, and model-based control:

• Law discovery and symbolic regression: By bottlenecking edge-messages to the true physical dimensionality and enforcing message aggregation corresponding to force superposition, GNNs trained on data can recover explicit analytic laws (e.g., Newtonian gravitation, Hookean elasticity) via symbolic regression on learned message functions (Cranmer et al., 2019).
• System identification: Recurrent GNNs can encode latent "static" graphs from observed trajectories, yielding parameters necessary for closed-form system prediction and control without needing direct access to all underlying system variables (Sanchez-Gonzalez et al., 2018).
• Trajectory optimization and RL: Differentiable GNN-based forward models serve as surrogates for physical simulators in model predictive control, value-gradient reinforcement learning, and policy optimization, showing improved convergence and sample efficiency versus model-free baselines (Sanchez-Gonzalez et al., 2018).

4. Integration of Physics in High-Energy and Materials Science Applications

High-energy physics (HEP) and materials science have driven significant innovations in physics-inspired GNNs through domain-specific inductive biases:

• Detector and particle graph construction: In HEP, measurement elements—such as hits, clusters, or reconstructed particles—are mapped to nodes, and physically-plausible connections are enforced by distance-based or geometry-based filters, mirroring causality and locality (e.g., connecting only adjacent layers or nearby calorimeter cells) (Ju et al., 2020, Thais et al., 2022, Shlomi et al., 2020).
• Edge classification and combinatorial reduction: Edge labels reflect physical relations (shared particle origin, common shower), with adjacency structure and MLP message-passing reducing the vast combinatorial search space (from $O(N^2)$ to near-linear scaling) (Ju et al., 2020).
• Symmetry and conservation encoding: Physics-inspired features and loss regularizers enforce Lorentz, permutation, or gauge invariance, and conservation constraints (e.g., ensuring predicted momentum/energy sum match global event quantities), improving calibration, resolution, and out-of-distribution generalization (Shlomi et al., 2020, Thais et al., 2022).
• Tensor-valued and rotated node features: In materials modeling, embedding per-grain tensors rotated to the laboratory frame (rather than quaternions or Euler angles) allows robust generalization to arbitrary loading directions without vastly inflating the training set, as exemplified by AnisoGNN (HU et al., 2024).

5. Physics-Inspired GNNs for Combinatorial Optimization and Statistical Mechanics

The mapping from statistical mechanics to combinatorial optimization underlies a class of “physics-inspired” GNN frameworks targeting NP-hard optimization on graphs:

• QUBO and spin-glass formalism: Binary- or multi-class problems (MaxCut, MIS, graph coloring) are encoded as quadratic unconstrained binary optimization (QUBO) or anti-ferromagnetic Potts models, whose energy becomes the differentiable loss for unsupervised GNN training (Schuetz et al., 2021, Schuetz et al., 2022, Colantonio et al., 2024).
• Continuous relaxations and binarization: Model outputs are continuous relaxations (probabilities) during training, binarized for final discrete assignments, with sophisticated binarization or fuzzy logic (e.g., Łukasiewicz t-norm) strategies mitigating phase-transition-induced collapse in dense graphs, as empirically demonstrated (Krutský et al., 18 Jul 2025).
• Statistical mechanical inductive biases: Losses, regularizers, and training schedules mirror physical setups (e.g., Gibbs measures, entropy maximization, noise annealing) to accelerate convergence, break permutation symmetry, and improve scalability across hard problem instances (Colantonio et al., 2024).

Physics-inspired GNNs in this class scale to graph instances containing up to $10^6$–$H(q,p)$0 nodes, matching or surpassing conventional solvers on both synthetic and real-world benchmarks (Schuetz et al., 2021, Schuetz et al., 2022, Colantonio et al., 2024, Krutský et al., 18 Jul 2025).

6. Specialized Physics-Informed Message-Passing for PDEs and Conservation Laws

Physics-informed GNNs extend to spatiotemporal PDEs and network-constrained flows by directly incorporating discrete physical operators and conservation principles:

• Physics-informed losses and graph augmentation: Loss function terms penalize residuals of governing equations (e.g., RBF-FD discretized Laplacians for PDEs, Kirchhoff-type constraints for network flows), enabling training and inference without exhaustive data labels (Xiang et al., 2022, Botta et al., 11 Dec 2025).
• Graph augmentation with physics-derived nodes: Augmenting the node set with physics-computed virtual sensors or process-specific variables (e.g., pressure/temperature-drop nodes in industrial networks) endows the learned model with robustness, interpretability, and improved extrapolation accuracy over naive data-driven GNNs (Niresi et al., 2024).
• Operator learning on irregular meshes: By combining meshless local finite difference schemes (e.g., RBF-FD) with GNN architectures, models circumvent topological regularity constraints and generalize to arbitrary unstructured domains (Xiang et al., 2022).

Physics-informed GNN surrogates in these settings offer orders-of-magnitude speed-up over traditional solvers, stable error properties under out-of-domain conditions, and strong architectural flexibility.

7. Advances, Limitations, and Outlook

Physics-inspired GNNs unify relational learning with strict enforcement or soft guidance by physical principles, anchoring their expressivity and generalization on foundational laws.

Advances:

• Demonstrated zero-shot generalization to unseen configurations, topologies, or parameter regimes;
• Strong gains in energy, momentum, and mass conservation over generic data-driven baselines;
• Seamless scalability from small benchmarks to industry-scale, real-world systems;
• Synthesis of statistical mechanics, combinatorial optimization, and deep learning for robust graph optimization.

Limitations:

• Best performance is achieved in highly regular, compositional systems; highly amorphous domains may require further architectural tuning;
• Rollout and extrapolation errors still accumulate over very long horizons or under strong noise;
• Physics constraints assumed to be known or easily computable, which can be challenging in poorly characterized domains;
• Existing approaches predominantly deterministic, with limited treatment of stochasticity, partial observability, and structure discovery.

Future directions include the development of stochastic or probabilistic physics-inspired GNNs, automated structure identification (learning adjacency), domain adaptation, integration with interpretable symbolic learning pipelines, and cross-disciplinary expansion (fluid mechanics, quantum systems, biomedical networks) (Thangamuthu et al., 2022, Botta et al., 11 Dec 2025, Sanchez-Gonzalez et al., 2018).

---

Physics-inspired GNNs constitute a mathematically principled, empirically validated, and highly extensible paradigm, driving both practical advances and theoretical insights at the confluence of graph deep learning and fundamental physical law. For further technical details, see (Sanchez-Gonzalez et al., 2018, Thangamuthu et al., 2022, Cranmer et al., 2019, Ju et al., 2020, HU et al., 2024, Xiang et al., 2022, Schuetz et al., 2021, Schuetz et al., 2022, Krutský et al., 18 Jul 2025, Botta et al., 11 Dec 2025, Niresi et al., 2024).