🔬 materials science

Specimen design for material parameter identification using topology optimization

This paper proposes a framework that integrates Bayesian optimal experimental design with topology optimization to automatically generate specimen geometries that maximize information gain for identifying material constitutive parameters using full-field measurement data.

Original authors: Adeline Wihardja, Kaushik Bhattacharya

Published 2026-07-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Adeline Wihardja, Kaushik Bhattacharya

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but the culprit is invisible. In the world of materials science, the "culprit" is the hidden set of rules that dictate how a material behaves when you squeeze, stretch, or twist it. These rules are called constitutive relations. Think of them as the material's secret personality: does it snap like a dry twig, stretch like a rubber band, or flow like honey? To figure out this personality, scientists usually put a piece of the material in a machine and pull on it. But here's the catch: the machine can only measure how hard it's pulling and how far the ends move. It can't see the invisible stress and strain happening inside the material.

Traditionally, scientists have used simple shapes, like a dumbbell-shaped piece of rubber (called a "dogbone"), to get a clear, uniform stretch. It's like trying to learn about a person's personality by only asking them to say "hello" in a quiet room. It works for simple cases, but if the material is complex—like human skin or a biological tissue that has fibers running in different directions—a simple "hello" isn't enough. You need a more chaotic, interesting conversation to really understand them. This is where the challenge lies: how do you design a test that reveals the most secrets with the least amount of guessing?

This is exactly what Adeline Wihardja and Kaushik Bhattacharya tackle in their paper. They propose a clever new way to design the shape of the test specimen itself, using a mix of advanced math and computer simulations. Instead of guessing what shape might be interesting, they use a method called Bayesian Optimal Experimental Design. Think of this as a super-smart planner that asks, "If I try this specific weird shape, how much new information will I learn compared to what I already think I know?" They combine this with topology optimization, which is like a digital sculptor that can carve away any part of a block of material to find the perfect, non-intuitive shape.

The authors' main finding is that by letting the computer design the specimen, they can create shapes that are far more informative than the standard dumbbell shapes. In their simulations, they tested this on both simple, uniform materials and complex, fiber-reinforced materials (like those found in biological tissues). They found that the computer-designed shapes generated a much richer variety of internal stresses and strains, even when pulled in just one direction. This "richness" allowed them to identify the material's hidden parameters with much greater accuracy. For instance, when testing a material with fibers running in specific directions, the optimized shape had strange, linear patterns that forced the material to reveal its directional secrets, whereas a standard shape would have missed them.

However, the paper is careful to note that these results come from computer simulations, not physical experiments in a lab. The authors suggest that while the method works beautifully in the digital world, there are challenges in the real world. For example, they found that trying to use raw camera images directly to guide the design was too "noisy" and unstable for the computer to handle effectively. Instead, they found that using the calculated movement (displacement) of the material was a more stable route for the design process. They also point out that while their method suggests these shapes are superior, the ultimate test will be seeing if they work when actually printed and pulled in a real lab, where manufacturing defects and real-world noise might change the outcome.

In short, the paper suggests that we don't have to stick to boring, standard test shapes. By using a smart, math-driven approach to design the specimen's geometry, we can squeeze more information out of a single test. It's like realizing that to understand a complex puzzle, you shouldn't just look at the edge pieces; you need to arrange the pieces in a way that forces the hidden picture to reveal itself. While this is currently a powerful simulation, it offers a promising path toward faster, cheaper, and more accurate ways to understand the materials that make up our world, from the tires on our cars to the tissues in our bodies.

Technical Summary: Specimen Design for Material Parameter Identification using Topology Optimization

Problem Statement
Constitutive relations are essential for closing the equations of continuum mechanics and predicting material behavior in engineering applications. However, identifying these relations is challenging because internal quantities like stress and strain are not directly measurable; only boundary loads and displacements are observable. Traditional mechanical testing relies on simple specimen geometries (e.g., dogbones) to achieve homogeneous deformation states, which limits the data to sparse, single strain states. While recent advances in full-field imaging (e.g., Digital Image Correlation) allow for the observation of heterogeneous deformation states, the inverse problem of inferring material parameters from such data remains ill-posed. Furthermore, standard experimental designs often rely on intuition or conventional protocols that do not maximize the information content of the experiment, leading to sparse data and potentially inaccurate model calibration, particularly for anisotropic materials.

Methodology
The authors propose a framework that couples Bayesian Optimal Experimental Design (OED) with Topology Optimization to design specimen geometries that are maximally informative for constitutive model discovery.

Bayesian Inverse Problem Formulation:
The parameter identification problem is formulated as a Bayesian inverse problem. The goal is to estimate unknown material parameters $P$ from experimental data $\tilde{y}$ . The framework assumes Gaussian priors and Gaussian noise. The posterior distribution is approximated using a Laplace approximation (Gaussian approximation centered at the Maximum A Posteriori estimate), which allows for the analytical derivation of the posterior covariance matrix $C_{post}$ .
Optimal Experimental Design Criterion:
The objective is to maximize the Expected Information Gain (EIG), defined as the expected Kullback-Leibler divergence between the prior and posterior distributions. Under the Gaussian approximation, maximizing EIG is equivalent to minimizing the determinant of the posterior covariance matrix, known as the D-optimal criterion. The authors adopt a local Bayesian D-optimal approach, linearizing the forward map around the current best parameter estimate to compute the Hessian of the negative log-likelihood.
Topology Optimization Integration:
Instead of optimizing a few geometric parameters (e.g., notch radius), the design variable is a high-dimensional spatially varying material density field $\rho$ . The authors use the SIMP (Solid Isotropic Material with Penalization) method to relax the discrete design problem (solid/void) into a continuous one. The objective function (negative EIG) is minimized subject to the constraints of the forward physical problem (quasistatic equilibrium of a hyperelastic material).
Gradient-Based Optimization:
The optimization is solved using a gradient-based approach. To handle the computational cost of PDE-constrained optimization, the authors employ the adjoint method. They derive the sensitivity of the objective function with respect to the design density $\rho$ by solving two forward problems (displacement field and parameter sensitivity) and two adjoint problems. This allows for efficient gradient computation without requiring thousands of forward solves per iteration.
Data Formulations:
The framework supports two data formulations:
- Image-based: Uses raw image data and boundary forces directly.
- Displacement-based: Uses displacement fields inferred from images (via DIC) and boundary forces. The authors note that the image-based approach can lead to noisy sensitivities due to the differentiation of image intensities, making the displacement-based formulation more stable for the optimization process in their examples.

Key Results
The framework was demonstrated on hyperelastic materials in a uniaxial tension setting:

Isotropic Materials: The optimized specimen topologies differed significantly from standard dogbone shapes, often featuring internal strips or complex voids. When tested with synthetic data, the optimized designs yielded more accurate parameter inference (lower error in bulk modulus) compared to random designs with the same volume fraction.
Anisotropic (Fiber-Reinforced) Materials: The method successfully designed specimens with linear elements aligned in multiple directions to probe anisotropic behavior. The results showed that the optimized designs consistently achieved higher Expected Information Gain (EIG) and lower posterior variance (A-optimal criterion) compared to arbitrary or conventional designs.
Robustness to Priors and Initial Guesses: The optimized designs were found to be relatively insensitive to variations in the prior parameter estimates and initial design guesses, suggesting that meaningful designs can be obtained even without iterative refinement.
Performance: The optimized designs led to reduced errors in inferred parameters (e.g., shear modulus, fiber orientation angles) when compared to non-optimized designs.

Significance and Claims
The paper claims that coupling Bayesian OED with topology optimization provides a principled foundation for designing experimental configurations that maximize information gain. Key contributions include:

Moving beyond conventional, intuition-based specimen geometries to discover non-intuitive topologies that generate rich, heterogeneous deformation fields.
Providing a quantitative framework (D-optimal criterion) to guide the design of experiments for complex, anisotropic materials where traditional testing is insufficient.
Demonstrating that while direct image-based inference is valuable for high-fidelity modeling, a displacement-based formulation offers a more stable and practical route for the optimal design of experiments, avoiding the noise issues associated with differentiating raw image data.
Establishing that optimized specimens can significantly improve the accuracy of material parameter identification, thereby enhancing the fidelity of constitutive models used in engineering predictions.

The authors conclude that this approach reduces the number of experiments required for high-fidelity model discovery and note that future work could extend this to history-dependent materials and incorporate robustness against manufacturing variability.

More like this