🔬 materials science

Microstructure-Conditioned Surrogate Models for Graded Multiscale Optimization of Mycelium Composites

This paper introduces a microstructure-conditioned hypernetwork surrogate model that enables efficient multiscale optimization of mycelium composites, achieving a 42% reduction in peak stress for functionally graded structures while demonstrating a practical pathway to engineer microstructures directly via manufacturing variables.

Original authors: J. Storm, I. B. C. M. Rocha, S. Schyck, K. Masania, F. P. van der Meer

Published 2026-07-16

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: J. Storm, I. B. C. M. Rocha, S. Schyck, K. Masania, F. P. van der Meer

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 building a house out of LEGO bricks. If you use the exact same brick for every single wall, floor, and roof tile, the house is easy to build, but it might be weak in the spots where the wind hits hardest. Nature, however, is a master architect. Think of a tree trunk or a human bone: they aren't made of one uniform stuff. Instead, they are "functionally graded," meaning their internal structure changes smoothly from the outside in. The outside might be tough and dense to protect against bumps, while the inside is lighter and more flexible to save weight. This smart design makes them incredibly strong and efficient.

For decades, engineers have wanted to copy nature's trick by building materials that change their internal "recipe" from one spot to another. But there's a catch: figuring out exactly how to arrange the tiny bits inside a material to get the perfect strength is like trying to solve a million-piece puzzle while blindfolded. To do this, computers usually have to simulate the behavior of every single tiny piece, which takes so much time and power that it's often impossible to test many different designs. This is where the story gets interesting: scientists are now teaching computers to "guess" the answer using smart shortcuts, but they need a way to make those guesses work for any tiny pattern, not just one specific design.

This paper tells the story of how researchers at Delft University of Technology created a clever new "guessing machine" to solve this puzzle, specifically for a sustainable material made of fungus (mycelium) and wood chips.

The Problem: The "One-Size-Fits-None" Dilemma

Imagine you have a super-smart robot that can predict how a specific LEGO wall will hold up under pressure. If you show it a wall made of red bricks, it's great. But if you ask it about a wall made of blue bricks, or a mix of red and blue, it gets confused and needs to be retrained from scratch. In the world of advanced materials, this is a huge problem. To design a perfect, nature-inspired structure, engineers need to test thousands of different internal patterns (microstructures). Training a new robot for every single pattern would require an impossible amount of data and time.

The researchers wanted to build a robot that could look at any pattern of wood chips and fungus and instantly tell them how strong it would be, without needing a new lesson for every single variation.

The Solution: The "Shape-Shifting" Brain

The team developed a new type of artificial intelligence called a HyPRNN (Hybrid Physics-Data Recurrent Neural Network). To understand how it works, think of it as a master chef with a special assistant.

The Chef (The Physics Model): The core of the system is a "chef" who knows the basic rules of cooking (physics). This chef knows how fungus and wood chips react to being squished or stretched. This part is rigid and reliable; it never forgets the laws of physics.
The Assistant (The Hypernetwork): The problem is that the chef needs to know exactly what ingredients are in the pot. Is it 10% wood chips? 50%? Are the chips long and thin, or short and round?
- Old methods tried to give the chef a new recipe book for every single ingredient mix.
- The new method uses a "smart assistant" (the hypernetwork). This assistant looks at the ingredients (the microstructure) and instantly whispers the right instructions to the chef. It tells the chef, "Hey, today we have 30% wood chips arranged in circles, so adjust your cooking style accordingly."

This allows the system to handle a vast range of different internal structures without needing to be retrained for each one. It's like having a chef who can instantly adapt to any recipe you throw at them, as long as you tell them what's in the pantry.

The Test: Fungus and Wood Chips

To prove this idea worked, the researchers chose a very trendy, eco-friendly material: mycelium-woodchip composites. Mycelium is the root-like part of mushrooms. When it grows on wood chips, it acts like a natural glue, creating a lightweight, biodegradable material that could replace plastic or foam.

They simulated a manufacturing process where wood chips and mycelium "pellets" are dropped into a mold, much like pouring cereal into a bowl. The way the chips fall and settle creates a unique, random pattern.

The Challenge: They wanted to see if their "Shape-Shifting Brain" could predict how strong the material would be based on how many pellets were added or how the chips were oriented before they were dropped.
The Result: They trained the AI on a relatively small dataset (only 512 examples). Usually, AI needs millions of examples to learn this well. However, because their AI was "conditioned" on the physics of the material, it learned incredibly fast. It could predict the strength of new, unseen patterns with high accuracy.

The Big Win: Designing the Perfect Disk

The real magic happened when they used this AI to design a new object. They set up a computer simulation of a disk (like a wheel) that needed to withstand internal pressure.

The Random Approach: If you just mix the wood chips and fungus randomly, the disk has weak spots and high stress.
The Optimized Approach: The AI ran an optimization loop, trying thousands of different "gradients" (changing the mix of wood chips and fungus from the center of the disk to the edge).
The Outcome: The AI found a design where the material properties changed smoothly across the disk. This new design reduced the peak stress (the point where the material is most likely to break) by 42% compared to the random version.

Even more impressively, they tested this on a 3-point bending beam (a classic engineering test). A traditional computer simulation that calculates every tiny detail took nearly 2 hours (5,900 seconds) to run. The new AI-powered simulation did the same job in just 2 seconds. That is a speed-up of nearly 3,000 times!

Why This Matters

The paper doesn't just show a faster computer; it shows a new way to engineer the future. By connecting the manufacturing process (how you drop the chips) directly to the final strength of the material, engineers can now design structures that are as efficient as nature's bones.

The researchers also showed that they could condition the AI directly on "manufacturing variables." Instead of trying to describe the complex shape of the wood chips with math, they just told the AI, "Here is the setting we used to drop the chips," and the AI learned how that specific setting changed the material's strength. This means that in the future, we might be able to tell a 3D printer or a factory robot, "Make this part stronger here and lighter there," and the machine will automatically figure out the perfect internal arrangement of fibers and particles to make it happen.

In short, this paper proves that by teaching AI to respect the laws of physics while learning from data, we can unlock the potential of sustainable, bio-based materials and design structures that are lighter, stronger, and smarter than anything we've built before.

Technical Summary: Microstructure-Conditioned Surrogate Models for Graded Multiscale Optimization of Mycelium Composites

Problem Statement
Emerging sustainable materials, such as mycelium-woodchip composites, rely on engineered microstructures to achieve specific mechanical properties. Optimizing these materials for functionally graded structures requires efficient multiscale simulations. Traditional concurrent multiscale methods (FE2), which embed representative volume elements (RVEs) within a macroscopic domain, are computationally prohibitive because they require solving a separate boundary value problem for every quadrature point. While data-driven surrogate models can accelerate these simulations, existing approaches face significant limitations:

Fixed Microstructure: Most surrogates are trained on a single, fixed microstructure and cannot generalize to varying underlying geometries, which is essential for graded materials.
Data Scarcity: Approaches that attempt to handle varying microstructures (e.g., using CNNs or GNNs) typically require vast amounts of training data, which is costly to generate.
Manufacturing Complexity: Directly parametrizing complex microstructural geometries is often difficult. A practical route requires linking manufacturing variables directly to macroscale behavior without intermediate, complex latent space models.

Methodology
The authors propose a HyPRNN (Hypernetwork-Physically Recurrent Neural Network), a hybrid physics-data surrogate model conditioned on microstructural and manufacturing variables.

Architecture: The core model is a PRNN, which mimics computational homogenization via an encoder-material-decoder framework. The encoder maps macroscopic strain to microscopic strains; a set of fictitious material points (using embedded physics-based constitutive models, specifically neo-Hookean) computes microscopic stresses; and a decoder homogenizes these into macroscopic stress.
Conditioning via Hypernetwork: To handle varying microstructures, the PRNN weights are not fixed. Instead, a hypernetwork ( $H$ ) takes microstructural variables (e.g., volume fraction, aspect ratio) as input and predicts the weights for the PRNN's encoder and decoder. This allows the model to learn a continuous mapping from microstructural parameters to the network's weight space.
Finite Strain Adaptations:
- Linear Encoder: Uses a Cholesky-based parametrization to ensure the predicted microscopic deformation tensors remain positive definite, satisfying physical constraints even with limited data.
- Nonlinear Encoder: Introduces a flexible neural network layer for the encoder to capture geometric nonlinearities in finite strains, constrained to ensure zero strain in the undeformed state and positive definiteness of the resulting tensors.
- Material Parameters: Material properties (e.g., shear modulus $\mu$ ) are handled by directly modifying the embedded constitutive models within the fictitious material points, avoiding the need for retraining.
Manufacturing Conditioning: In a second approach, the hypernetwork is conditioned directly on manufacturing variables (e.g., pellet fraction and initial woodchip orientation) derived from a discrete-element method (DEM) simulation, bypassing the need for an explicit geometric parametrization or a generative model.

Key Contributions

Microstructure-Conditioned PRNN: The authors extend PRNNs to account for geometric variations in microstructure using a hypernetwork, enabling accurate predictions for graded materials with small datasets.
Finite Strain PRNNs: They introduce a more flexible nonlinear encoder and modify the linear encoder to remain valid for large deformations, ensuring physical consistency (positive definiteness) in finite strain regimes.
Tractable Multiscale Optimization: The HyPRNN is integrated into an optimization loop to design graded multiscale structures, demonstrating a significant reduction in computational cost compared to full FE2 simulations.
Direct Manufacturing Conditioning: The paper demonstrates a practical route to engineer microscale behavior by conditioning the surrogate directly on manufacturing process variables (from DEM simulations), eliminating the need for intermediate latent space models.

Results

Performance vs. Data: In learning curve comparisons, the HyPRNN with a linear encoder significantly outperforms standard neural networks (NN) and nonlinear PRNNs in the low-data regime. The linear PRNN's physics-based bias allows it to transfer knowledge across different material parameters effectively. As data volume increases, the nonlinear PRNN and NN improve, but the linear HyPRNN remains competitive and robust.
FE2 Validation: In a 3-point bending beam simulation with functionally graded parameters, the surrogate-based simulations achieved near-perfect agreement with full RVE-based FE2 simulations (using 512 training samples) while reducing computation time from ~5900 seconds to ~2 seconds. Standard NNs showed poor performance in the multiscale context, particularly with limited data, often producing non-zero stresses in undeformed states.
Optimization Outcomes:
- Graded Disk: Optimizing a rotating disk with graded microstructural parameters (volume fraction, aspect ratio, shear modulus) reduced the peak Cauchy stress by 42% compared to a random microstructure and by 5% compared to a uniform optimization.
- Manufacturing Variables: In a compression test of a sample with a hole, optimizing manufacturing variables (pellet fraction and initial orientation) via the HyPRNN reduced horizontal bulging from 1.48 to 0.00001, demonstrating the ability to find specific manufacturing inputs that yield desired macroscale deformations.

Significance and Claims
The paper claims that microstructure-aware, conditioned surrogate models are essential for the efficient design of future sustainable materials. By combining physics-based inductive biases with data-driven flexibility, the HyPRNN enables the multiscale optimization of functionally graded structures that would otherwise be computationally intractable. The work highlights that conditioning directly on manufacturing variables provides a practical pathway to control microscale geometry for desired macroscale performance, potentially accelerating the development of bio-based composites like mycelium-woodchip materials. The authors emphasize that this approach reduces the data requirements for training surrogates and avoids the complexity of additional generative models for microstructure parametrization.