Learning Physics-Constrained Constitutive Models from Data

43 concepts33 papers1 proposal

To simulate how a bridge deforms under traffic, how a metal sheet stamps into a car panel, or how biological tissue reacts to a surgical tool, engineers rely on constitutive models. These are the mathematical equations that bridge the gap between stress (the internal forces within a material) and strain (how much it deforms).

For decades, these equations were hand-crafted by human experts. A researcher would observe material behavior in a lab, use physical intuition to propose a mathematical framework (such as a linear spring or an idealized plastic flow rule), and then fit parameters to the experimental data. However, this manual approach has hit a rigid bottleneck.

The Rigidity of Hand-Crafted Material Laws

Modern engineering frequently utilizes materials with highly intricate microstructures, such as fiber-reinforced composites, 3D-printed lattices, or complex geological formations. Traditional analytical models often lack the mathematical flexibility required to describe these systems.

As Mark Messner points out in A Data-Driven Method for Modeling Creep-Fatigue Stress-Strain Behavior Using Neural ODEs, conventional physics-based models possess rigid mathematical structures that cannot capture subtle, non-standard material behaviors like initial loading overshoots. When a material's true physical response deviates from the idealized assumptions of the modeler, these rigid frameworks break down. For instance, Xiang Ding et al. demonstrated in Unexpected viscoelastic deformation of tight sandstone: Insights and predictions from the fractional Maxwell model that standard models struggle to accurately represent the complex, multi-parameter viscoelastic deformations found even in hard materials like sandstone.

Furthermore, calibrating these hand-crafted models is notoriously slow and labor-intensive. As Daniel N. Blaschke et al. note in Machine learning based approach to predict ductile damage model parameters for polycrystalline metals, traditional parameter tuning relies heavily on human expertise to guide iterative, computationally heavy inverse searches, often requiring dozens of CPU-hours to calibrate a single material system.

The Data-Driven Shift and the Black-Box Trap

To overcome these limitations, the field of computational mechanics is undergoing a paradigm shift toward data-driven representations, as reviewed by Francisco J. Montáns et al. in Data-driven modeling and learning in science and engineering. Instead of forcing a material to conform to a simplistic, pre-defined analytical equation, researchers are leveraging machine learning to learn the material laws directly from experimental data.

Yet, blindly applying standard "black-box" neural networks to solid mechanics introduces a dangerous flaw: standard neural networks do not understand physics. A vanilla neural network might perfectly fit a specific training dataset, but it can easily predict physically impossible behaviors when generalizing to new stress states—such as violating the conservation of energy or predicting that a material will spontaneously generate stress without any applied strain.

To build reliable models, machine learning frameworks must be constrained by the immutable laws of continuum mechanics. The goal is to create a system that possesses the infinite flexibility of neural networks but remains strictly bound by physical reality.

Hardcoding Physics into Neural Networks

The most direct way to ensure physical consistency is to embed thermodynamic and mechanical constraints directly into the neural network's mathematical architecture.

Consider hyperelastic materials, where the stress-strain relationship is derived from a strain energy density function. In Neural networks meet hyperelasticity: A guide to enforcing physics, Dominik K. Klein et al. introduce Physics-Augmented Neural Networks (PANNs). Instead of letting a network predict stress arbitrarily, the authors design the network architecture to satisfy fundamental principles like objectivity (invariance under material rotation) and polyconvexity (which ensures material stability). Because these constraints are built into the network's mathematical structure, the model is physically "uncorruptible" and can accurately extrapolate even when trained on sparse data. The authors expanded this concept in Parametrized polyconvex hyperelasticity with physics-augmented neural networks to capture complex deformations in materials with intricate microstructures, such as fiber-reinforced composites.

Another approach is to embed neural networks directly inside differentiable physics engines. In Learning Neural Constitutive Laws from Motion Observations for Generalizable PDE Dynamics, Pingchuan Ma et al. present NCLaw, a framework that lets a neural network learn the messy, unknown constitutive behaviors of materials like sand or rubber while a differentiable simulator strictly enforces the conservation of mass and momentum. This hybrid approach achieves "one-shot generalization," learning a material's properties from a single moving object and successfully predicting chaotic collisions in entirely new geometries.

To make data-driven mechanics even more efficient, Bahador Bahmani developed a method outlined in Manifold embedding data-driven mechanics. Instead of performing slow database searches for the closest experimental data point during a simulation, the author uses invertible neural networks to project complex material states onto a continuous, flat Euclidean manifold, drastically accelerating simulations while maintaining high accuracy.

Automated Discovery of Human-Readable Laws

While embedding neural networks into mechanics solves the flexibility problem, engineers still face a practical hurdle: neural networks are black boxes containing millions of uninterpretable weights. For safety-critical applications, engineers prefer explicit, human-readable mathematical equations.

To bridge this gap, researchers are using machine learning not just to replace equations, but to discover them. In A Data-Driven Method for Modeling Creep-Fatigue Stress-Strain Behavior Using Neural ODEs, Mark Messner demonstrates that a Neural Ordinary Differential Equation (Neural ODE) can be trained to capture complex cyclic stress behaviors and subsequently "distilled" into explicit symbolic equations using symbolic regression. Similarly, Bahador Bahmani proposes a "divide and conquer" strategy in Physics-constrained symbolic model discovery for polyconvex incompressible hyperelastic materials, automatically extracting interpretable, physically valid symbolic equations for advanced synthetic and biological materials.

Taking this a step further, some researchers treat the derivation of material laws as a linguistic or strategic problem. Georgios Kissas et al. introduce a linguistic framework in The Language of Hyperelastic Materials, defining a formal mathematical grammar for material deformations. By pairing this grammar with deep generative models, the authors automatically discover optimal, physically valid constitutive equations directly from experimental data without human intervention.

In a parallel vein, Kun Wang 0005 introduces a "meta-modeling game" in Meta-modeling game for deriving theory-consistent, microstructure-based traction–separation laws via deep reinforcement learning. By treating the formulation of a mathematical model as a game—akin to chess or Go—a deep reinforcement learning agent learns to assemble theory-consistent constitutive laws for complex material interfaces, routinely outperforming human-designed alternatives.

Key Foundational Work

A PROJECTION APPROACH TO THE NUMERICAL ANALYSIS OF LIMIT LOAD PROBLEMS (2011) — contributed 2 concept elements to this direction.
Effect of Stress Ratio on Fatigue Threshold of Cracked Components. (1998) — 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_manual_constitutive_modeling_a2e2003e,
  title = {manual constitutive modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/manual-constitutive-modeling/rdr_a2e2003eda92499583b6fdc286892df9},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_manual_constitutive_modeling_a2e2003e_v1,
  title = {manual constitutive modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/manual-constitutive-modeling/rdr_a2e2003eda92499583b6fdc286892df9/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake