HyPRNN

Published:

HyPRNN

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

This work was performed in collaboration with Iuri Rocha, Sarah Schyck, Kunal Masania, and Frans van der Meer.

Multiscale optimization of microstructured sustainable materials needs surrogate models, but training one across a range of microstructures normally demands prohibitive amounts of data. The work solves this by conditioning a hybrid physics-data surrogate on microstructural variables through a hypernetwork, a HyPRNN, which stays accurate on small datasets and is validated against a full FE² simulation of a mycelium-woodchip composite. Applied to a functionally graded disk it cuts peak stress by 42% versus a random microstructure, and conditioning directly on manufacturing variables extends the approach into a practical route for engineering the microscale to hit target macroscale behavior.

Link to preprint: https://arxiv.org/abs/2607.13688

GitHub Repository link: https://github.com/JoepStorm/HyPRNN

Overview
Overview for optimizing a graded multiscale material using a surrogate model conditioned on either microscale or manufacturing variables.

Multiscale optimization

The HyPRNN makes accurate predictions based on only 32 training samples. We validate this in an FE² simulation

val_deformation
Multiscale simulations of the ground truth (top left), and various surrogate models. The HyPRNN performs well, even in the low-data regime.

With a well-trained surrogate model, we can run multiscale simulations in seconds. This allows for optimizing the microscale variables:

disk_grading_comparison
Optimization, made possible by the HyPRNN, leads to a significant reduction in peak stress.

Conditioning on microscale variables

Parametrizing the microstructure is challenging, and some other works resolve this by, for example, training a variational-autoencoder to capture the microstructure, and conditioning on the latent space. We avoid this complexity by instead conditioning directly on manufacturing variables. This is possible by simulating how the manufacturing properties affect the resulting microscale. We do this using a discrete-element simulation.

grav_dep_init_0 grav_dep_3d_box
Visualization of the gravity deposition simulation that mimics the manufacturing process. This links the manufacturing variables to the generated microstructure.

This allows realistic optimization of multiscale simulation, such as to control the macroscopic deformation.

grav_dep_baseline grav_dep_optimal_crop
Macroscale deformation can be controlled. The left figure shows the reference deformation and the right figure shows the optimized deformation and woodchip placement. The goal was to minimize the horizontal deformation of the center hole.