Discovery of Hyperelastic Constitutive Laws from Experimental Data with EUCLID

This paper evaluates the EUCLID framework for the automated discovery of hyperelastic constitutive laws using experimental data from natural rubber specimens, comparing its performance against conventional parameter identification methods in terms of predictive accuracy, generalization to unseen geometries, and coverage of the material state space.

The Gist

Imagine you are a chef trying to recreate a secret family recipe for a perfect rubber band. You have the final dish (the rubber band) and you can taste it (measure how it stretches), but you don't know the exact list of ingredients or the specific amounts used.

This paper is about a new, smart way to figure out that recipe without guessing.

The Old Way: The "Guess-and-Check" Chef

Traditionally, scientists trying to understand how materials like rubber behave have to play a game of "Guess and Check."

1. Pick a Recipe: They start by guessing the type of recipe they think it might be (e.g., "It's probably a Mooney-Rivlin soup" or "It's an Ogden stew").
2. Taste and Adjust: They measure the rubber, tweak the numbers in their chosen recipe, and see if the math matches the real rubber.
3. The Problem: If the first guess is wrong, they have to start over with a completely different recipe. This is slow, tedious, and relies heavily on the scientist's intuition. It's like trying to bake a cake by randomly guessing whether you need salt, sugar, or sand.

The New Way: The "EUCLID" AI Chef

The authors of this paper tested a new tool called EUCLID (Efficient Unsupervised Constitutive Law Identification and Discovery). Think of EUCLID as a super-smart AI chef who doesn't need you to pick a recipe.

Instead, EUCLID has a giant pantry filled with thousands of possible ingredients (mathematical terms) and thousands of possible recipe structures.

1. The Tasting: You give EUCLID data from the rubber band (how much force you pull with, and how much it stretches).
2. The Discovery: EUCLID automatically sifts through its giant pantry. It tries to build a recipe using the fewest ingredients possible that still perfectly explains the rubber's behavior. It uses a mathematical "filter" (called sparse regression) to throw out the ingredients that aren't needed.
3. The Result: It hands you the exact recipe, discovering the "secret sauce" automatically, without you ever having to guess what the recipe should look like.

The Experiment: Simple vs. Complex Puzzles

To test if this AI chef was actually good, the researchers didn't just use simple rubber bands. They created a whole gym of challenges:

• The Simple Tests: They pulled on standard, dog-bone-shaped rubber strips (like a standard tug-of-war). This gives them "Global Data" (just the total force and total stretch).
• The Complex Tests: They cut holes, circles, and weird shapes into the rubber and pulled on them. This creates "Local Data." Imagine pulling on a rubber sheet with a hole in the middle; the material stretches wildly differently around the hole compared to the edges. This gives a much richer, more detailed map of how the material behaves.

They used a high-tech camera system (Digital Image Correlation) to watch the rubber stretch like a slow-motion movie, tracking every tiny speck of paint on the surface.

The Big Findings

Here is what they discovered, translated into everyday terms:

1. The AI Chef Wins: EUCLID found a recipe that was just as good as, or even better than, the best recipes humans had manually picked. It did this without needing a human to say, "I think it's an Ogden recipe."
2. Details Matter: Using the "Local Data" (the complex shapes with holes) helped the AI understand the rubber better than just the simple "Global Data." It's like trying to learn how a car engine works by just listening to the noise (Global) vs. taking the engine apart and looking at every piston (Local). The detailed view gave a more robust recipe.
3. One Recipe Fits All: The best part? The recipe EUCLID discovered worked perfectly on new shapes it had never seen before. If you gave it a weird, complex shape it hadn't been trained on, it could still predict exactly how that shape would stretch.
4. Speed and Simplicity: The old way of finding the "Ogden" recipe is mathematically messy and hard to solve (like trying to solve a maze in the dark). EUCLID's method is like solving a maze with a bright flashlight; it's faster, more stable, and less likely to get stuck.

The Bottom Line

This paper proves that we don't need to be experts guessing the right mathematical formulas anymore. We can just feed the data into a smart system like EUCLID, and it will automatically "discover" the laws of physics governing the material. It turns the difficult art of material modeling into a reliable, automated science, making it easier to design better tires, medical implants, and soft robots.

Technical Summary

1. Problem Statement

Determining accurate constitutive laws for nonlinear materials (specifically hyperelastic materials like natural rubber) from experimental data is a persistent challenge in mechanics. Traditional approaches suffer from two main limitations:

• Subjectivity and Trial-and-Error: Engineers must pre-select a functional form for the material model (e.g., Mooney-Rivlin, Ogden, Gent-Thomas) before identifying parameters. If the chosen form is incorrect, the model fails, leading to tedious iterative cycles.
• Data Limitations: Many methods rely on global data (force-displacement) from simple tests (uniaxial tension, pure shear), which may not sufficiently cover the material's state space (multiaxial stress states) or handle noisy data effectively.

The paper aims to evaluate EUCLID (Efficient Unsupervised Constitutive Law Identification and Discovery), a framework that automates both model selection and parameter identification, and to compare its performance against traditional methods using real-world experimental data.

2. Methodology

A. The EUCLID Framework

EUCLID unifies model selection and parameter identification by treating the constitutive law discovery as a sparse regression problem.

• Model Library: The method utilizes a comprehensive library of strain energy density functions ($W$) containing:
  • Generalized Mooney-Rivlin (GMR) terms (polynomials of invariants $I_1, I_2$).
  • Gent-Thomas (GT) terms (logarithmic terms).
  • Ogden terms (functions of principal stretches $\lambda_i$).
  • The library includes 521 candidate terms.
• Optimization: The problem is formulated to minimize the discrepancy between experimental data and the model prediction while enforcing sparsity (selecting only the necessary terms) using LASSO (Least Absolute Shrinkage and Selection Operator) regularization.
  • Objective Function: Minimizes the equilibrium gap and boundary constraint errors (based on the Virtual Fields Method/Equilibrium Gap Method) plus an $L_1$-norm penalty on the parameters.
  • Sparsity Control: A hyperparameter ($\lambda_p$) is tuned via Pareto analysis to balance model accuracy (Mean Square Error) and complexity (number of active terms).

B. Experimental Campaign

The study used Natural Rubber (NR-40) specimens tested under quasi-static conditions.

• Specimens:
  1. Dogbone: For Uniaxial Tension (UT).
  2. Wide Rectangular: For Pure Shear (PS).
  3. Complex Geometries: Specimens with circular and elliptical holes (TTa–TTf) to induce heterogeneous, multiaxial stress states.
• Data Acquisition:
  • Global Data: Reaction forces and machine displacements.
  • Local Data: Full-field displacement maps obtained via Digital Image Correlation (DIC) (stereo DIC system, 4096x3000 pixels).
• Noise Assessment: The authors quantified DIC noise (temporal and spatial) to ensure data reliability, finding it within acceptable limits for 8-bit imaging.

C. Comparative Approaches

The study compared three scenarios:

1. Traditional Identification: Pre-selecting a model form (e.g., 2-term Ogden) and fitting parameters using nonlinear least squares.
2. EUCLID with Global Data: Using only force-displacement curves from UT and PS tests.
3. EUCLID with Local Data: Using full-field displacement and reaction forces.
4. EUCLID with Mixed Local Data: Using UT (large stretch, homogeneous) and TTf (complex geometry, diverse stress states) local data.

3. Key Contributions

1. First Experimental Validation of EUCLID: While EUCLID was previously tested on synthetic data, this paper validates it on real, noisy experimental data from natural rubber.
2. Comparison of Data Types: It rigorously compares the efficacy of global data (simple tests) vs. local data (full-field DIC) for model discovery.
3. State Space Coverage Analysis: The authors analyze how different specimen geometries cover the strain invariant plane ($I_1, I_2$). They demonstrate that complex geometries (holes) provide a richer diversity of stress states compared to simple specimens, even if the maximum stretch is lower.
4. Generalization Assessment: The discovered models were tested against unseen geometries (complex specimens not used in calibration) to evaluate predictive capability.

4. Results

• Model Discovery Success: EUCLID successfully identified a constitutive law that combined terms from GMR and GT formulations (or Ogden terms depending on the dataset) without prior knowledge of the correct form.
• Accuracy Comparison:
  • The 2-term Ogden model was the best-performing pre-selected model across all datasets.
  • EUCLID matched or slightly outperformed the 2-term Ogden model in predictive accuracy (measured by $L_2$ error in stress and displacement fields) but did so automatically, removing the need for subjective model selection.
• Data Type Performance:
  • Local Data (UT + TTf): This combination yielded the lowest prediction errors for both global and local quantities. The diversity of stress states provided by the complex geometry (TTf) combined with the large stretch range of UT proved superior to using only simple tests (UT+PS).
  • Global vs. Local: Models identified using local data generally showed better generalization to complex geometries than those identified using only global data.
• Robustness: EUCLID demonstrated robustness to experimental noise and was able to discover accurate models even when the library was restricted (e.g., excluding Ogden terms, though accuracy dropped slightly).
• Computational Efficiency: EUCLID relies on convex minimization, making it computationally more efficient and stable than traditional methods for multi-term Ogden models, which require non-convex optimization with multiple random initial guesses.

5. Significance and Conclusions

• Paradigm Shift: The paper validates a shift from "Model Selection + Parameter Identification" to "Automated Model Discovery." EUCLID eliminates the trial-and-error process of choosing the right functional form.
• Data Strategy: The study concludes that for material model discovery, synergy is key. Combining a simple test (providing large strain ranges) with a complex geometry test (providing diverse multiaxial stress states) is the most effective strategy for covering the material state space, even with standard uniaxial testing machines.
• Practical Application: EUCLID offers a flexible, data-driven tool for reliable material modeling. It is particularly valuable when the underlying physics is complex or unknown, as it can automatically select the most appropriate terms from a vast library to fit experimental observations.
• Future Outlook: The authors suggest that while complex loading setups (biaxial, torsion) would further improve state space coverage, the proposed method using standard equipment and complex specimen geometries is sufficient for robust discovery of hyperelastic laws.

In summary, the paper demonstrates that EUCLID is a powerful, automated framework capable of discovering accurate hyperelastic constitutive laws from experimental data, outperforming or matching traditional methods while removing the subjectivity of model selection.