A Neural-Network Framework to Learn History-Dependent Constitutive Laws and Identifiability of Internal Variables

This paper presents a causal and energetic neural-network framework for learning history-dependent constitutive laws that ensures thermodynamic consistency, stability, and solution existence while demonstrating that learned internal variables are unique up to a linear transform, achieving a 2% relative error in predicting the response of a polycrystalline magnesium unit cell.

The Gist

The Big Picture: Teaching a Computer to "Feel" Materials

Imagine you are trying to predict how a piece of metal will bend, stretch, or squish when you push on it. In engineering, we usually use math formulas (called constitutive laws) to describe this behavior.

However, metals are tricky. They don't just react to the push you are giving them right now; they remember every push and pull they've ever experienced. This is called history-dependence. If you stretch a piece of metal, let it go, and stretch it again, it behaves differently the second time because of its "memory."

Traditionally, scientists have to guess the right math formulas to describe this memory. But for complex materials (like the magnesium metal in this study), guessing the right formula is incredibly hard.

The Solution: The authors built a special type of Artificial Intelligence (AI)—specifically a Neural Network—that can learn these complex "memory" rules directly from data, without needing a human to guess the formula first.

The Problem: AI Can Be "Unphysical"

If you just let a standard AI learn from data, it might get very good at predicting the past, but it could make up crazy physics for the future. For example, it might predict that if you squeeze a metal block hard enough, it disappears into a single point without any resistance. In the real world, that's impossible; matter resists being crushed into nothingness.

Standard AI also doesn't naturally understand the Second Law of Thermodynamics (which basically says energy is lost as heat when things rub against each other) or stability (the material shouldn't suddenly explode or behave erratically).

The Solution: The "Physics-First" AI Framework

The authors created a new framework that forces the AI to obey the laws of physics by design, not just by luck. Think of it like building a car engine where the pistons are physically locked to the wheels; the car cannot drive backward if the wheels are moving forward.

Here is how they did it:

1. The "Internal Variables" (The Hidden Memory):
   Since the AI can't see the microscopic changes inside the metal (like tiny defects moving around), the authors introduced invisible "memory slots" called internal variables.

   • Analogy: Imagine a sponge. When you squeeze it, water moves inside. You can't see the water moving, but the sponge's shape changes because of it. The "internal variables" are the AI's way of tracking where that "water" (the microscopic changes) is, even though it's hidden.
   • The Discovery: The paper proves that while the AI might invent different "memory slots" depending on how it starts learning, those slots are always just a linear transformation of each other.
   • Simple translation: If one AI decides to call its memory "Slot A" and another calls it "Slot B," they are actually describing the exact same thing, just using a different coordinate system (like measuring distance in inches vs. centimeters). They are mathematically equivalent.

2. The "Energy Potentials" (The Rules of the Game):
   The AI learns two main things:

   • Stored Energy: How much energy is saved up when you stretch the material (like a spring).
   • Dissipation: How much energy is lost as heat (like friction).
     The authors built the AI so it must follow the rule that energy loss is always positive (you can't get energy back for free) and that the material gets infinitely hard to compress as it gets smaller (it can't be crushed to a point).

3. The "Growth Functions" (The Safety Net):
   To ensure the AI doesn't predict impossible scenarios (like infinite compression), they added special mathematical "guardrails."

   • Analogy: Imagine a video game character who can run fast, but if they try to walk off the edge of the map, a giant invisible wall pushes them back. These guardrails ensure that if you try to stretch or squeeze the material beyond the data the AI has seen, it still behaves realistically (getting harder and harder to deform) rather than breaking the laws of physics.

The Experiment: Polycrystalline Magnesium

The team tested this framework on magnesium, a metal used in cars and planes. Magnesium is made of many tiny crystals (grains) stuck together, making its behavior very complex.

• The Setup: They generated data by simulating the microscopic behavior of a tiny cube of this magnesium.
• The Training: They fed this data to their "physics-aware" AI.
• The Result: The AI learned to predict how the whole block of magnesium would behave with only 2% error. This is incredibly accurate.
• The Speed: Because the AI is a fast computer program, it can predict this behavior much faster than the slow, complex microscopic simulations it was trained on.

Key Takeaways

• Accuracy: The AI learned the complex "memory" of the metal with 2% error.
• Physics Compliance: The AI respects the laws of thermodynamics and material stability. It won't predict that a metal can be crushed into a dot.
• Unique Memory: Even though the AI creates "hidden" variables to track memory, the paper proves these variables are unique up to a simple math change (like switching units). This means the AI isn't just hallucinating random numbers; it's finding a real, consistent structure.
• Objectivity: The model works correctly even if you look at the material from a different angle (rotation), which is a crucial requirement for real-world engineering.

Summary

The authors built a smart, physics-savvy AI that can learn how complex metals behave over time. It's like teaching a student not just the answers to math problems, but the fundamental rules of arithmetic so they can solve any problem correctly, even ones they've never seen before. The result is a fast, accurate, and physically realistic model for predicting how materials like magnesium will react under stress.

Technical Summary

Technical Summary: A Neural-Network Framework to Learn History-Dependent Constitutive Laws and Identifiability of Internal Variables

Problem Statement
The identification of constitutive laws is critical for modeling material behavior, particularly when bridging the gap between microscopic physics and macroscopic engineering applications. Traditional approaches often rely on fitting experimental data to mathematical models or using surrogate models to reduce the computational cost of multiscale simulations (e.g., FE2). However, standard data-driven models, particularly those based on recurrent neural networks (RNNs) that map strain history directly to stress, often lack rigorous adherence to fundamental physical principles. Specifically, they may fail to satisfy:

1. The second law of thermodynamics (dissipation).
2. Material stability under extreme strains.
3. The mathematical conditions (such as polyconvexity) required to guarantee the existence of solutions for the governing non-linear momentum balance equations.

Furthermore, in inelastic materials, internal variables are often postulated a priori or learned from data without a clear theoretical characterization of their uniqueness. In data-driven settings, it is unclear whether learned internal variables are unique or merely different representations of the same physical state.

Methodology
The authors propose a causal, energetic neural network (NN) framework based on the theory of internal variables to learn history-dependent constitutive laws. The approach reformulates the non-Markovian stress-strain relationship into a Markovian system by introducing a set of internal variables $\xi(t)$.

• Energetic Formulation: The stress $P$ and the evolution of internal variables are derived from an energy density potential $W(F, \xi)$ and a dissipation potential $\bar{D}(\dot{\xi})$. The framework enforces the second law of thermodynamics by ensuring $\bar{D}$ is convex with a minimum at zero.
• Explicit Dynamics via Legendre Transform: To avoid solving implicit ordinary differential equations (ODEs) during training, the authors utilize the Legendre transform of the dissipation potential. This converts the implicit evolution equation into an explicit form, allowing the network to learn the dual dissipation potential $D$ directly.
• Neural Network Architecture:
  • Input Convex Neural Networks (ICNNs): The potentials $W$ and $D$ are parameterized using shallow ICNN architectures. The weights in the final layer are constrained to be positive to ensure the convexity of the potentials, which is necessary for thermodynamic consistency and solution existence.
  • Polyconvexity and Objectivity: The energy potential $W$ is decomposed into hydrostatic and deviatoric components. The hydrostatic part is modeled to ensure polyconvexity (a condition for existence theorems in nonlinear elasticity). To satisfy objectivity (frame indifference), the model is trained on symmetric deformation gradients (Right Cauchy-Green tensor space) and extended to non-symmetric gradients via polar decomposition, rather than relying on soft constraints in the loss function.
  • Material Stability: To prevent unphysical behavior under extreme strains (e.g., compressing a material to a point with finite energy), specific growth functions ($G_1, G_2$) are added to the potentials. These functions remain nearly constant within the data support but enforce polynomial growth outside it, ensuring the stress blows up appropriately as the determinant of the deformation gradient approaches zero or strain becomes large.
• Training Strategy: The framework is trained on data generated from a fine-scale crystal plasticity model of a polycrystalline magnesium unit cell. The training involves minimizing the error between predicted and true stress components, with separate loss terms for the hydrostatic and deviatoric parts to handle the disparity in their magnitudes.

Key Contributions

1. Physically Consistent Framework: The authors introduce a NN framework that simultaneously enforces the second law of thermodynamics, material stability, objectivity, and polyconvexity. Unlike previous works that impose these constraints partially or via soft penalties, this framework integrates them into the architecture and energy definition.
2. Objective Polyconvex Extension: The paper presents a method to achieve both polyconvexity and objectivity by training on symmetric tensors and extending the definition via polar decomposition, avoiding the poor predictive accuracy associated with invariant-based models or soft-constraint loss terms.
3. Identifiability of Internal Variables: The work provides a theoretical proof that internal variables learned from data are unique up to a linear transformation. This addresses the non-uniqueness issue often reported in data-driven constitutive modeling, characterizing the equivalence class of surrogate models.
4. High-Fidelity Surrogate Modeling: The framework is successfully deployed to learn the Taylor-averaged response of a polycrystalline magnesium microstructure, achieving high accuracy without compromising physical consistency.

Results

• Prediction Accuracy: The trained polyconvex neural network achieves a 2% relative error in predicting the Taylor-averaged stress response of the magnesium polycrystal. The use of energy decomposition (hydrostatic vs. deviatoric) was found to be crucial, reducing error from 15% (without decomposition) to 2%.
• Internal Variable Dimensionality: Empirical analysis indicates that six internal variables are optimal for this specific homogenization problem, balancing accuracy and computational cost.
• Material Stability: Visualizations confirm that the growth functions successfully enforce physical limits. Without these functions, the model predicts finite stress at zero volume (det $F=0$), which is physically unstable. With the functions, the stress correctly approaches negative infinity as the material is compressed.
• Identifiability Verification: The authors trained two independent NN models with different initializations on the same dataset. The resulting internal variable trajectories were found to be linear transformations of each other, with a median linear fit error of approximately 3%, empirically validating the theoretical uniqueness claim.

Significance and Scope
The paper claims that this framework offers a robust, interpretable, and mathematically sound approach to learning constitutive laws for history-dependent materials. By ensuring consistency with thermodynamics and existence theory, the resulting models are suitable for use as surrogates in multiscale finite element solvers (e.g., Abaqus), potentially replacing expensive FE2 calculations. The authors note that while the method was demonstrated on viscoelastic/viscoplastic magnesium, the framework is general enough to capture any history-dependent response. The work concludes by identifying future directions, such as learning heterogeneous, microstructure-dependent potentials and spatially non-local responses, but does not propose specific experimental validations beyond the numerical homogenization case study presented.