Neural Constitutive Models (NCMs)

• Neural Constitutive Models (NCMs) are neural network representations of constitutive laws that map kinematic inputs to stress, flux, or stiffness outputs while honoring physical principles like objectivity and symmetry.
• They employ diverse formulations—including energy-based, incremental, corrective, and nonlocal operator approaches—to capture complex material behaviors in hyperelasticity, turbulence closure, and biomechanics.
• NCM architectures enforce constitutive admissibility through invariant, equivariant, and convexity-preserving designs, and are trained on synthetic, experimental, or full-field data to address challenges in extrapolation and computational cost.

Searching arXiv for recent and foundational papers on Neural Constitutive Models to ground the article in current literature. Neural Constitutive Models (NCMs) are neural-network-based representations of constitutive laws or closure relations in continuum mechanics and computational physics. In recent work, they appear as learned maps from kinematic or resolved fields to stresses, fluxes, stiffnesses, conductivity tensors, or closure variables, and increasingly as neural strain-energy functions or neural operators that are constructed to satisfy objectivity, symmetry, incompressibility, polyconvexity, permutation invariance, or frame indifference by design rather than by post hoc correction (Linka et al., 2022, Han et al., 2022, Knipper et al., 22 May 2026). This usage spans hyperelasticity, turbulence closure, elasticity imaging, multiscale surrogate modeling, stochastic anisotropic transport, and small-data constitutive evolution.

1. Conceptual scope and constitutive role

In continuum mechanics, a constitutive model specifies how stresses, fluxes, or internal variables depend on kinematic or state variables. The forms represented in the recent NCM literature include classical hyperelastic relations such as $$\boldsymbol{\sigma} = \mathcal{F}(\boldsymbol{\varepsilon},\dot{\boldsymbol{\varepsilon}},T,\ldots)$$, RANS-style closure relations in which Reynolds stress or a scalar surrogate depends on mean-flow information, and spatially varying mappings in which the constitutive response depends on both deformation and position (Han et al., 2022, Hoerig et al., 2018).

A useful way to organize the field is by the mathematical object being learned. Some NCMs learn a local constitutive map, some learn a strain-energy density and recover stress by differentiation, some learn a nonlocal region-to-point operator, and some learn an effective corrective term on top of a nominal constitutive law. This suggests that “NCM” is not a single architecture class but a broader designation for neural representations of constitutive response embedded in the structure of continuum modeling.

In the literature represented here, major directions include local invariant networks, equivariant or invariant tensor networks, nonlocal graph or point-cloud models, neural operators, and physics-informed approaches that calibrate constitutive parameterizations from full-field data (Han et al., 2022). A related but distinct branch treats constitutive tensors such as conductivity as stochastic SPD objects and learns constitutive-to-response maps with geometry-aware preprocessing on the SPD manifold (Schuttert et al., 16 Jun 2025).

2. Mathematical formulations

A central distinction among NCMs concerns whether the network predicts stress-like quantities directly or instead predicts a scalar potential. In hyperelasticity, many of the most structured models are energy-based. The network outputs a Helmholtz free energy density $\Psi$, typically as a function of invariants of $\mathbf{C}=\mathbf{F}^\top\mathbf{F}$ or of principal stretches, and stresses follow from differentiation. For incompressible isotropic hyperelasticity this is often written as $\Psi=\Psi(I_1,I_2)$, with first Piola–Kirchhoff or second Piola–Kirchhoff stresses obtained through the chain rule (Hou et al., 2023, Linka et al., 2022). This representation makes differentiability, stress consistency, and FE integration natural, and in several papers tangent operators are obtained analytically or by automatic differentiation (Parvez et al., 2024, Knipper et al., 22 May 2026).

A second formulation is incremental or evolutionary. NICE, for example, treats the constitutive law as a material-point initial value problem,

$$\dot{\mathcal{X}} = \mathbf{f}_\theta(\mathcal{X}, \dot{\boldsymbol{\varepsilon}}),$$

with state $$\mathcal{X}=\{\rho,\boldsymbol{\varepsilon}^e,\mathbf{z}\}$$ and a neural internal energy potential providing stresses and dissipation (Masi et al., 2023). This places NCMs close to classical internal-variable constitutive theory rather than to static regression.

A third formulation is corrective. In "Neural Material" (Wang et al., 2018), the neural network does not replace the nominal law; it learns a residual stress correction in the principal frame,

$$\Delta \hat{\mathbf{P}} = N_\theta(\hat{\mathbf{F}},\hat{\dot{\mathbf{F}}}),$$

and the total stress is the nominal co-rotational response plus the neural correction. This is a constitutive model in the sense that the stress response is partially parameterized by neural weights, but it remains anchored to a hand-crafted baseline.

A fourth formulation is nonlocal operator learning. In turbulence and transport closures, the constitutive response is explicitly region-to-point. The VCNN-e architecture learns

$$R(\mathbf{x}_0)=\mathcal{N}\big(Q(\Omega(\mathbf{x}_0))\big),$$

where $Q$ collects vector-cloud features over a region of influence around $\mathbf{x}_0$ (Han et al., 2022). Earlier nonlocal work expresses the same idea through a Green’s-function-inspired convolutional structure in which a learned kernel acts on a learned production field (Zhou et al., 2020), while a frame-independent vector-cloud network extends the region-to-point map to arbitrary grids and unordered point clouds (Zhou et al., 2021).

3. Architectural principles and enforced structure

A persistent theme in NCM research is that architecture is used to encode constitutive admissibility. CANNs exemplify this strategy. They are reverse-engineered from continuum mechanics: inputs are invariant combinations such as $\Psi$0 and $\Psi$1, outputs are free energies rather than stresses, activation functions are selected to encode admissible constitutive subforms, and connectivity is restricted to support polyconvexity and interpretability (Linka et al., 2022). In the hyperelastic model-discovery setting, this can be extended from fixed invariants to trainable generalized invariants

$\Psi$2

so that the network learns both the energy function and the exponents $\Psi$3 that define the invariant basis (Martonová et al., 16 Aug 2025).

Objectivity and symmetry are often enforced at the representation level. In isotropic hyperelastic models, invariant inputs guarantee frame indifference. In transversely isotropic atrial models, the free energy is written as $\Psi$4 with two orthogonal fiber families, and the discovered sparse laws remain directly interpretable in terms of isotropic and fiber contributions (Peirlinck et al., 3 Apr 2025). For stochastic anisotropic conductivity, CMNN introduces a preprocessing layer that maps SPD tensors from the manifold $\Psi$5 to a tangent-space representation before standard neural processing, precisely to preserve SPD geometry and scale–orientation structure (Schuttert et al., 16 Jun 2025).

Nonlocal tensorial models require stronger equivariance. VCNN-e achieves invariance to translation and permutation and equivariance to rotation for tensor-valued outputs by constructing the output tensor as

$\Psi$6

with invariant summaries driving the learned diagonal weights and scalar isotropic term (Han et al., 2022). The earlier vector-cloud formulation achieves translation, rotation, and permutation invariance for scalar outputs by operating on relative coordinates, scalar features, and permutation-invariant aggregates over the cloud (Zhou et al., 2021).

Convexity and polyconvexity are another major design axis. A physics-preserving ICNN for isotropic fibrous materials uses shifted invariants $\Psi$7 and nonnegative-weight Softplus layers to obtain a convex energy in invariant space, while stresses and stiffnesses are generated by automatic differentiation from that potential (Parvez et al., 2024). FE-MAD similarly distinguishes a grey-box CANN, described as polyconvex and highly flexible, from a white-box CANN with phenomenologically interpretable strain-energy terms (Knipper et al., 22 May 2026).

4. Training regimes and identification strategies

NCMs are trained under markedly different data regimes. One regime uses synthetic or high-fidelity simulation data to learn constitutive surrogates. VCNN-e is trained first to emulate a Reynolds stress transport model and then to learn an effective closure from DNS of periodic hills (Han et al., 2022). The fibrous-material ICNN is trained on strain energy, stress, and stiffness data generated by discrete fiber-network simulations, effectively replacing an FE$\Psi$8/multiscale constitutive update by a learned single-scale hyperelastic law (Parvez et al., 2024).

A second regime uses experimental constitutive datasets directly. The brain-cortex study compares a constitutive ANN against multivariate regression on uniaxial tension, compression, and simple shear data. It shows that neural networks can automatically identify accurate constitutive models from admissible estimators, but also that regularization and cross-validation are decisive, and that regression can yield more compact laws when the candidate library is fixed (Hou et al., 2023). The generalized-invariants framework further turns model discovery into joint optimization over network weights and invariant exponents, rather than over weights alone (Martonová et al., 16 Aug 2025).

A third regime uses full-field structural data rather than homogeneous stress–strain curves. FE-MAD embeds a constitutive neural network directly inside a JAX-FEM nonlinear solver and identifies parameters by minimizing a measurement-mismatch loss against open experimental datasets including full DIC of a perforated tensile specimen, a reduced-data stretch-profile scenario, and a heterogeneous matrix–inclusion system (Knipper et al., 22 May 2026). A related weak-form PINN framework calibrates classical hyperelastic models from full-field displacement data and global force–displacement curves without interpolating the data onto a computational grid (Hamel et al., 2022). Although that framework calibrates analytic constitutive models rather than a neural constitutive law, its weak-form inversion strategy is directly relevant to NCM deployment.

A fourth regime addresses sparse and incomplete observations. NICE replaces increment-based training by integrated trajectory matching for constitutive ODEs and is explicitly designed for “small data - sparse and partial material state observations” (Masi et al., 2023). “Neural Material” addresses sparse surface trajectories through reduced space-time optimization and least-squares stress recovery, then retrains a residual constitutive network from the recovered stress labels (Wang et al., 2018). These works challenge the misconception that NCMs necessarily require dense, full-state data.

5. Applications and reported capabilities

The application range is unusually broad. In turbulence closure, VCNN-e reports TKE error of approximately $\Psi$9 for interpolated periodic-hill geometries, with extrapolation errors of approximately $\mathbf{C}=\mathbf{F}^\top\mathbf{F}$0 for $\mathbf{C}=\mathbf{F}^\top\mathbf{F}$1 and $\mathbf{C}=\mathbf{F}^\top\mathbf{F}$2 for $\mathbf{C}=\mathbf{F}^\top\mathbf{F}$3; in the DNS-based setting, a purely local NCM gives $\mathbf{C}=\mathbf{F}^\top\mathbf{F}$4 error versus $\mathbf{C}=\mathbf{F}^\top\mathbf{F}$5 for the nonlocal VCNN-e, underscoring the benefit of nonlocal information in separated and reattaching flow (Han et al., 2022). The earlier nonlocal operator formulation and the frame-independent vector-cloud model show that the learned region size, kernel asymmetry, and point-cloud invariance can reproduce convection–diffusion–reaction structure on periodic hills and related transport problems (Zhou et al., 2020, Zhou et al., 2021).

In hyperelasticity and multiscale mechanics, the physics-preserving ICNN achieves normalized mean square errors of $\mathbf{C}=\mathbf{F}^\top\mathbf{F}$6 for strain energy density, $\mathbf{C}=\mathbf{F}^\top\mathbf{F}$7 averaged across stress components, and $\mathbf{C}=\mathbf{F}^\top\mathbf{F}$8 averaged across stiffness components, and maintains upward of $\mathbf{C}=\mathbf{F}^\top\mathbf{F}$9 accuracy in stress up to $\Psi=\Psi(I_1,I_2)$0 strain in a facet capsular ligament simulation while reducing computational cost by orders of magnitude relative to the multiscale reference (Parvez et al., 2024). FE-MAD extends this direction to identification from experiments and shows that constitutive laws for both phases in a heterogeneous matrix–inclusion system can be identified and generalized to twenty-two previously unseen samples (Knipper et al., 22 May 2026).

In biomechanics, constitutive ANNs and related CANNs have been used for human brain cortex and atrial tissue. For brain cortex, ANN-based discovery validates the applicability of artificial neural networks to automatically discover constitutive material models with proper regularization, but the study also reports that multi-mode loading is essential for generalization and that simpler regression models can outperform over-parameterized neural ones when the admissible library is fixed (Hou et al., 2023). In the generalized-invariants framework, rubber is fit with average $\Psi=\Psi(I_1,I_2)$1 using one learned positive exponent, while brain tissue reaches average $\Psi=\Psi(I_1,I_2)$2 with a one-term identity-only energy in a learned large negative invariant (Martonová et al., 16 Aug 2025). For atrial tissue, sparse four-term constitutive neural networks achieve overall $\Psi=\Psi(I_1,I_2)$3 for left atrium and $\Psi=\Psi(I_1,I_2)$4 for right atrium, with isotropic response dominated by $\Psi=\Psi(I_1,I_2)$5-based terms and anisotropy captured by $\Psi=\Psi(I_1,I_2)$6-type fiber invariants (Peirlinck et al., 3 Apr 2025).

In elasticity imaging, CaNNCMs learn both linear-elastic behavior and its spatial distribution. They reconstruct smooth and sharp Young’s modulus distributions with mean relative errors of about $\Psi=\Psi(I_1,I_2)$7–$\Psi=\Psi(I_1,I_2)$8 and $\Psi=\Psi(I_1,I_2)$9–$$\dot{\mathcal{X}} = \mathbf{f}_\theta(\mathcal{X}, \dot{\boldsymbol{\varepsilon}}),$$0 on one mesh, report approximately $$\dot{\mathcal{X}} = \mathbf{f}_\theta(\mathcal{X}, \dot{\boldsymbol{\varepsilon}}),$$1 error on a kidney phantom, and recover geometry under noisy conditions while exposing over-training effects in the spatial network (Hoerig et al., 2018).

In stochastic constitutive parameterization, CMNN shows that geometry-preserving preprocessing such as logarithmic maps for scale information significantly improves learning performance over conventional MLPs when the constitutive object is an SPD conductivity tensor rather than a scalar parameter vector (Schuttert et al., 16 Jun 2025).

6. Limitations, misconceptions, and current directions

A common misconception is that NCMs are simply black-box stress regressors. The recent literature is more heterogeneous. Many successful models are not direct stress maps at all: they are energy-based hyperelastic laws, neural ODE evolutions, corrective residual laws, or nonlocal operators with explicit invariant or equivariant structure (Linka et al., 2022, Masi et al., 2023, Wang et al., 2018, Han et al., 2022). Another misconception is that physics enters only through loss penalties. In several important architectures, symmetry, objectivity, incompressibility, convexity, or manifold structure are enforced directly by the network design (Parvez et al., 2024, Schuttert et al., 16 Jun 2025).

The main limitations are equally recurrent. Extrapolation remains difficult when the evaluation regime lies outside the training distribution; VCNN-e exhibits larger errors for extreme hill shapes and identifies separation-dominated regions as the main difficulty (Han et al., 2022). Hyperparameter sensitivity and model non-uniqueness remain substantial in constitutive discovery; the brain-cortex comparison stresses the need for detailed cross-validations of regularization parameters and shows that different selection criteria can lead to different but similarly accurate constitutive laws (Hou et al., 2023). Some high-fidelity NCMs depend on derivative data, especially stiffness data, that are readily available in simulation but rarely in experiments (Parvez et al., 2024). Orientation handling on manifolds remains less mature than scale handling in SPD-based constitutive learning (Schuttert et al., 16 Jun 2025).

Computational cost is a further barrier. Stress and stiffness evaluations dominate FE runtimes when the constitutive update is neural and differentiable. COMMET addresses this by redesigning FE assembly around batched and vectorized constitutive evaluations, compute-graph-optimized derivatives, and MPI parallelism, reporting speed-ups exceeding three orders of magnitude relative to traditional non-vectorized automatic-differentiation-based implementations (Alheit et al., 1 Oct 2025). This suggests that the practical future of NCMs depends not only on constitutive architecture and identification strategy, but also on constitutive-update acceleration and solver-level integration.

Taken together, these developments indicate that NCMs are evolving from isolated surrogates into a layered methodology: architecture-level enforcement of constitutive principles, training strategies matched to available data, and solver frameworks capable of handling the resulting stress and tangent evaluations at scale.