HyperShape: Hyperelasticity Across Diverse Shapes

arXiv:2608.09938 · cs.CE, cs.LG · Submitted 2026-06-02 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "HyperShape: Hyperelasticity Across Diverse Shapes".

Jane: The paper was written by Leo Widmer, Stéphane Cotin, Sidaty El Hadramy and Philippe Claude Cattin from University of Basel and inria.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Title: Tom: Welcome back to the show, everyone! Today we’re digging into a brand new paper that just hit arXiv, and it’s called “HYPER SHAPE: Hyperelasticity Across Diverse Shapes.” Jane, I have to say, the title alone got me excited because it promises to tackle something that’s been bugging me for a while.

Jane: Oh, absolutely, Tom. And for our listeners who might not be deep in the weeds of computational mechanics, let me break down what that title actually means. Hyperelasticity is basically how soft materials like rubber, or even human tissue, deform when you stretch or squish them. And the paper is all about testing whether these clever AI models can handle that deformation across all sorts of different shapes, not just the same boring square over and over.

Tom: Right, and that’s the kicker. Most benchmarks in this field use one simple geometry, like a square with a hole in it, and they test the models on that same shape every single time. But in the real world, a liver isn’t a square, and a piece of rubber isn’t a circle. So this team from Basel and Inria built a whole framework to generate crazy, diverse shapes and see if the models can keep up.

Jane: And they didn’t just make one dataset. They made a whole suite of them, with different levels of difficulty. You’ve got simple shapes, complex shapes, and even shapes that look like actual human livers. It’s like they built a gym for these AI models to train and compete in.

Tom: A gym for neural operators, I love that analogy. And the results are honestly a bit humbling for the AI community, because these state-of-the-art models, they perform great on the simple stuff, but the moment you throw a complex, wiggly shape at them, their error rates just skyrocket.

Jane: Exactly. So this paper isn’t just presenting a new dataset. It’s really a wake-up call. It’s saying, hey, we’ve been patting ourselves on the back for solving these toy problems, but real-world applications are way harder, and we need to build better models.

Tom: And that’s what we’re going to unpack today. We’ve got our senior researcher Lu, our engineer Meng, and our in-house language model Lalam all here to break down what this means for the future of simulation and maybe even surgery. So stick around, because this is going to get interesting.

Summary: Jane: So, we’ve set the stage with the title, but let’s get into the meat of the paper. The core idea behind “HYPER SHAPE: Hyperelasticity Across Diverse Shapes” is pretty straightforward, but the execution is what makes it impressive. They built a pipeline that starts with a simple shape, like a circle or a sphere, and then randomly deforms it using a smooth velocity field.

Tom: And it’s not just one random deformation. They stack them. First, they apply a coarse deformation to get the overall blob shape, then a finer one to add local details and bumps. It’s like sculpting a potato from a ball of clay, but the sculptor is a random number generator.

Lu: That’s a great way to put it, Tom. And the beauty of this approach is that you can control the complexity. By tweaking parameters like the magnitude and the scale of those deformations, you can create datasets that range from “easy” – basically slightly wobbly circles – to “hard” – shapes that look like they’ve been through a blender. This lets you systematically test where a model starts to fail.

Jane: Right, and they didn’t stop at just making shapes. They also randomize the boundary conditions. So, you randomly pick a spot on the shape to hold fixed, and you randomly pick another spot to pull on. That’s a huge deal because in a lot of existing benchmarks, the boundary conditions are fixed for every single shape.

Meng: Yeah, and from an engineering standpoint, that’s what makes this so much more realistic. If you’re simulating a liver being lifted during surgery, the surgeon isn’t going to grab it in the exact same spot every time. So having that variability in the data is crucial for training a model that might actually be useful in the operating room.

Tom: And they ran all these simulations using a standard finite element solver, which is the gold standard for this kind of physics. So the data is trustworthy. Then they took five different state-of-the-art neural operator models and put them through the wringer on these new datasets.

Jane: And the summary of their findings is that performance degrades consistently as shape complexity increases. The models that were superstars on simple shapes became pretty unreliable on the complex ones. It really shows that the field has been overfitting to the benchmarks, not actually learning the underlying physics.

Lu: Exactly, Jane. And that’s the most important contribution here. It’s not just a new dataset; it’s a diagnostic tool that reveals the limitations of current methods. It gives the community a clear target to aim for, which is to build models that are truly geometry-agnostic.

Improvements: Tom: Alright, so we know the models struggle. But what does “HYPER SHAPE: Hyperelasticity Across Diverse Shapes” actually propose we do about it? What are the improvements they’re suggesting?

Jane: Well, Tom, I think the biggest improvement is the framework itself. They’re not just handing you a static dataset. They built an extensible system where you can easily swap in new base shapes, new material models, or even new ways of applying forces. It’s designed to grow with the field.

Lu: And that’s crucial. The authors are essentially saying, “We’ve shown you the problem, and here’s the tool to keep exploring it.” For example, in their current version, they only use a simple Neo-Hookean material model. But soft tissues are often much more complex. With this framework, you could plug in a more sophisticated model like Ogden or Mooney-Rivlin and see if the neural operators handle that any better.

Meng: I also noticed they mention that the current framework only applies a single traction force at one location. In reality, you’d have multiple forces acting simultaneously, like gravity, plus a tool pushing, plus another tool pulling. So a clear improvement would be to extend the framework to support multiple, simultaneous boundary condition regions. That would make the simulations much more challenging and realistic.

Tom: So they’re basically saying, “We’ve built the test, but the test can get even harder.” And that’s a good thing. But I also think the paper suggests an improvement in how we evaluate these models. They use both in-distribution and out-of-distribution tests, which is more rigorous than just testing on the same distribution you trained on.

Jane: Right, and that out-of-distribution test is where things get really interesting. They trained models on their synthetic shapes and then tested them on actual human liver geometries. That’s the ultimate transfer test, and the results were pretty rough for the point-cloud-based models. It really highlights that we need models that don’t just memorize shapes, but actually understand the physics of deformation.

Lu: Precisely. And the paper’s analysis of the Hausdorff distance, which is a measure of how different two shapes are, shows a clear correlation with error. The further the test shape is from the training shapes, the worse the model performs. This gives us a quantitative way to predict model failure, which is incredibly valuable for safety-critical applications.

First Page: Tom: Let’s zoom in on the very first page of “HYPER SHAPE: Hyperelasticity Across Diverse Shapes,” because I think there’s a lot of insight packed into that opening figure and the abstract. Jane, you want to walk us through that?

Jane: Absolutely. The first thing you see is this beautiful figure showing the whole pipeline. You start with a base shape, morph it into something complex, randomly assign the boundary conditions, and then run the FEM simulation to get the displacement field. It’s a clear visual summary of everything we’ve been talking about.

Tom: And then they have that other figure, Figure two which I think is the most powerful image in the whole paper. It shows two beams, fixed on the left, being pulled with the same force. But one beam has a tiny little notch in it. And that tiny notch completely changes the deformation pattern. It’s a perfect illustration of why this problem is so hard.

Lu: It really is. It demonstrates the extreme sensitivity of hyperelasticity to geometry. A small local change can have a global impact on the solution. That’s the fundamental challenge that makes this benchmark so important. It’s not just about making a model that’s a little bit better; it’s about making a model that can capture these highly nonlinear, geometry-dependent behaviors.

Meng: And that Figure three comparison is a real eye-opener for me. On the left, you see the old benchmark with the square with a hole, and the error histograms are nice and low. On the right, you see their new Random2D dataset with all these crazy shapes, and the same models’ errors are significantly higher. It’s a direct, visual proof that the field has been overestimating its own progress.

Jane: Exactly. And the abstract states it plainly: neural operators perform well on simple shapes but struggle as complexity increases. They even quantify that you need a lot of data to get decent performance in these complex regimes. It’s a sobering but necessary message for the community.

Tom: So the first page alone is worth the read. It sets up the problem, shows you the stark reality of the current state of the art, and introduces a tool to help fix it. I’m really curious to see how the community responds to this challenge.

Lu: I think it will be a catalyst. This gives researchers a clear, controllable environment to test new ideas. Instead of arguing over which model is better on a toy problem, we can now see which model actually generalizes to the messiness of the real world.

Conclusion: Tom: Well, we’ve reached the end of our time with “HYPER SHAPE: Hyperelasticity Across Diverse Shapes,” and I have to say, this is one of those papers that leaves you feeling both excited and a little humbled.

Jane: Definitely. We learned that they built this fantastic framework for generating diverse, controllable hyperelasticity datasets in both 2D and three dee. And when they tested the top neural operators on it, the models that looked so good on old benchmarks really started to fall apart as the shapes got more complex and the boundary conditions varied.

Lu: And that’s the real takeaway. The paper provides a rigorous, extensible benchmark that exposes the gap between our current models and what’s needed for real-world applications like surgical simulation. It’s a call to arms for the research community to focus on geometry-aware and physics-aware models.

Meng: From my side, it’s a reminder that we can’t just trust a model’s performance on a standard test set. We need to stress-test it with data that actually looks like the messy, variable real world. This framework gives us the tools to do exactly that.

Tom: And with that, we’ll say goodbye to this paper and get ready to dive into the next one. Thanks to Lu, Meng, and Lalam for joining the discussion, and thanks to all our listeners for tuning in.

Jane: See you next time, everyone!

Leo Widmer, Stéphane Cotin, Sidaty El Hadramy, Philippe Claude Cattin

University of Basel · inria · University of Basel · University of Basel

cs.CE, cs.LG

Submitted: 2026-06-02

Updated: 2026-08-12

Comments: 17 pages, 10 figures

Code: https://github.com/camlab-ethz/ConvolutionalNeuralOperator

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 64/100

The gist: "Hyperelastic deformations are highly sensitive to domain geometry and boundary conditions, making generalization across both a critical capability for neural operators applied to these problems.

Terminology

Summary

Summary

The paper introduces HYPER SHAPE, an extensible framework designed to generate synthetic shapes and their corresponding hyperelastic simulation data, producing a suite of 2D and 3D datasets with adjustable complexity and controllable shape variations. The framework enables systematic assessment of generalization across in-distribution, out-of-distribution, and synthetic-to-real transfer settings.

The paper addresses a critical gap: "Hyperelastic deformations are highly sensitive to domain geometry and boundary conditions, making generalization across both a critical capability for neural operators applied to these problems. However, existing benchmarks for neural operators on hyperelasticity rely on simple or few geometries, which makes it difficult to assess this capability rigorously."

The authors note that existing benchmarks may considerably overestimate the robustness and generalization capabilities of current neural operators by relying on overly simplified and homogeneous settings. For instance, the commonly used Elasticity benchmark uses a distribution of shapes consisting of squares with randomly shaped holes and the same boundary conditions for every geometry.

The framework's shape generation process starts with a base shape (e.g., a circle or sphere) and applies random deformations via a Gaussian random field with covariance C(v(x), v(y)) = m squared e-x-y 2 2/(2 sigma 2), where m controls deformation magnitude and sigma controls the scale. The deformation is computed by solving an ODE using scaling and squaring methods. This process is applied K times (with K in 1, 2, 3) to build complexity, with higher-frequency deformations introduced in later steps.

Boundary conditions are sampled by uniformly selecting a point p D on the boundary for the Dirichlet region D:= B r(p D) d, with the Neumann boundary being the complement. The traction is defined as t(x) = alpha times times phi(x-p N) over integral N phi(y)dy, where phi(x) = (-x 2 2/(2 gamma 2)) distributes the traction around a sampled point p N.

The paper generates multiple datasets: Random2D (1000 shapes, 50 simulations each), EasyRandom2D (500 shapes), OODRandom2D (500 shapes, out-of-distribution), OneRandom2D (1 shape, 10,000 simulations), Random3D (1000 sphere-based shapes), AugLiver (500 liver-based augmented shapes), Liver (92 real liver meshes from the LiTS dataset), and OneLiver (single liver shape).

The authors evaluate five state-of-the-art neural operators: U-Net (modified with ConvNeXt convolutions and multi-scale feature combination), FNO, CNO, GINO, and Transolver. Key findings include:

  1. In-distribution performance degrades with geometric variability: "all models exhibit a clear degradation in performance as the variability of the underlying geometry increases, moving from the single-shape setting (OneRandom2D / OneLiver), to slightly varying geometries (EasyRandom2D / Liver), and finally to more diverse shape distributions (Random2D / Random3D)."

  2. Out-of-distribution generalization is limited: In 2D, training on the more diverse Random2D improves OOD performance compared to EasyRandom2D. In 3D, point-cloud-based models struggle significantly when evaluated on shapes that differ substantially from the training distribution, with models trained on Random3D failing on Liver, while grid-based models performed better.

  3. Geometric complexity correlates with error: Using the isoperimetric ratio in 2D and Hausdorff distance in 3D, the authors show shapes closer to a circular geometry (isoperimetric ratio near 1) consistently yield lower prediction errors and as the Hausdorff distance increases, the relative L2 error also increases consistently across all models.

  4. Data efficiency is poor: "all models consistently improve as the amount of training data increases, with gains still visible beyond 20,000 samples for both in-distribution and out-of-distribution test sets. This indicates that current models require large training datasets to fully capture geometric variability and achieve strong generalization."

The paper's contributions are: (1) introducing HYPER SHAPE as a controllable synthetic shape generation framework with released code; (2) benchmarking state-of-the-art neural operators across in-distribution, out-of-distribution, and synthetic-to-real settings; (3) finding that current neural operators struggle with complex shape distributions; and (4) analyzing geometry-dependent failure modes, data efficiency, and conditions for synthetic pretraining transfer.

Limitations acknowledged include restriction to homogeneous materials, unexplored material parameter variability, boundary conditions constrained to a single surface location, and the joint effect of geometry and boundary conditions not being fully understood.

Improvements for AI systems

Based on the paper, here are the specific improvements I can make to AI systems and what the improved systems can do:

1. Geometry-Aware Input Encoding

  • Improvement: Replace naive SDF-based geometry encoding with a multi-scale, curvature-aware feature extractor that explicitly captures local geometric features (e.g., principal curvatures, local thickness, and boundary normals) in addition to global shape descriptors.

  • What the improved system can do: Better distinguish between shapes that have similar SDF values but vastly different deformation responses, directly addressing the failure mode shown in Figure 2 where small local geometry changes cause large global deformation differences.

2. Physics-Informed Regularization for Geometric Sensitivity

  • Improvement: Add a regularization term to the loss function that penalizes predictions violating the first Piola-Kirchhoff stress equilibrium (Equation 1) at interior points, computed via automatic differentiation. This forces the model to respect the underlying hyperelastic constitutive law, not just fit the data.

  • What the improved system can do: Reduce the consistent performance degradation observed as geometric complexity increases (Table 2), because the model is constrained to produce physically admissible displacement fields even for unseen, complex shapes.

3. Adaptive Shape-Complexity Weighting

  • Improvement: Implement a curriculum learning strategy where training samples are weighted by their isoperimetric ratio (2D) or Hausdorff distance (3D). Shapes with high complexity (low isoperimetric ratio) receive higher loss weights during training, forcing the model to focus on the hardest geometric cases.

  • What the improved system can do: Achieve more uniform error across the shape distribution, rather than the current behavior where errors spike dramatically for complex shapes (Figure 8), improving worst-case performance on out-of-distribution geometries.

4. Multi-Scale Boundary Condition Fusion

  • Improvement: Design a dedicated boundary-condition encoder that processes the Dirichlet and Neumann boundary information at multiple resolutions, using a feature pyramid network (similar to the multi-scale combination already used in the modified U-Net in Appendix C.1). This encoder explicitly models the spatial extent and location of the traction force relative to the fixed boundary.

  • What the improved system can do: Better generalize to different force locations and boundary configurations, addressing the paper's finding that boundary condition variability significantly degrades performance when combined with geometric diversity.

5. Meta-Learning for Few-Shot Geometry Adaptation

  • Improvement: Train the neural operator using a model-agnostic meta-learning (MAML) objective, where each task is a single shape with multiple loading conditions. The model learns an initialization that can quickly adapt to a new shape with only a few simulations.

  • What the improved system can do: Dramatically reduce the data requirement shown in Figure 10 (currently needing 40K+ samples), enabling deployment in scenarios where only a handful of simulations per new geometry are available, such as patient-specific surgical planning.

6. Geometry-Conditioned Hypernetwork

  • Improvement: Replace the fixed-weight neural operator with a hypernetwork that takes the shape's geometric features (e.g., a latent code from a shape autoencoder) as input and generates the weights of the main operator network. This decouples geometry representation from the PDE-solving network.

  • What the improved system can do: Explicitly separate the influence of geometry from the influence of boundary conditions, allowing the model to generalize to entirely new shape families (e.g., from synthetic spheres to real livers) without retraining, directly improving the synthetic-to-real transfer results in Table 3.

7. Uncertainty-Aware Prediction

  • Improvement: Add a Bayesian output head (e.g., Monte Carlo dropout or a normalizing flow) that outputs both the predicted displacement field and an associated uncertainty map. Train the model to maximize likelihood, not just minimize L2 error.

  • What the improved system can do: Provide confidence intervals for predictions, allowing downstream users (e.g., surgeons) to know when the model is operating outside its reliable regime. This is critical because the paper shows errors can be 10x higher on complex shapes, and an uncertainty estimate would flag these unreliable predictions.

8. Shape-Space Interpolation Regularization

  • Improvement: During training, generate interpolated shapes between pairs of training geometries (using the same diffeomorphic deformation framework from Section 3.1) and enforce that the model's predictions vary smoothly along these interpolation paths. This acts as a Lipschitz constraint in shape space.

  • What the improved system can do: Improve out-of-distribution performance by ensuring the model's response is continuous with respect to geometry changes, reducing the sharp error spikes observed when moving from training shapes to slightly different test shapes (Figure 9).

The fully improved system would:

  • Achieve sub-5% relative L2 error on Random2D (currently 12-29% for state-of-the-art models), by combining physics-informed regularization with adaptive complexity weighting.

  • Maintain performance on out-of-distribution shapes within 1.5x of in-distribution error (currently 2-5x degradation), via meta-learning and shape-space interpolation.

  • Transfer from synthetic to real anatomical geometries with less than 2x error increase (currently 3-7x for point-cloud models), using the hypernetwork approach.

  • Require only 5,000 training samples instead of 40,000+ to reach the same accuracy, through curriculum learning and meta-initialization.

  • Flag unreliable predictions with calibrated uncertainty, enabling safe deployment in medical applications where prediction errors on complex geometries could lead to incorrect surgical decisions.

Abstract

Hyperelastic deformations are highly sensitive to domain geometry and boundary conditions, making generalization across both a critical capability for neural operators applied to these problems. However, existing benchmarks for neural operators on hyperelasticity rely on simple or few geometries, which makes it difficult to assess this capability rigorously. To address this gap, we introduce HyperShape, an extensible framework designed to generate synthetic shapes and their corresponding hyperelastic simulation data, producing a suite of 2D and 3D datasets with adjustable complexity and controllable shape variations. This design enables systematic assessment of generalization across in-distribution, out-of-distribution, and synthetic-to-real transfer settings. Using this framework, we evaluated the performance of several state-of-the-art neural operators over diverse shape distributions. Our findings reveal that neural operators perform well on simple shapes but struggle as shape complexity, geometric diversity, and boundary condition variability increase, requiring large amounts of training data in such regimes. Performance degrades consistently and predictably with geometric complexity highlighting the need for further model development. As an open and extensible benchmark, HyperShape is designed to grow alongside the field: new geometries, material models, loading conditions, and evaluation settings can be easily incorporated to validate hyperelastic surrogate models.

Sources

Related papers