MIND: Microstructure INverse Design with Generative Hybrid Neural Representation

The paper introduces MIND, a novel generative model designed for inverse design of 3D tileable microstructures by integrating latent diffusion with Holoplane, an advanced hybrid neural representation that simultaneously encodes geometry and physical properties. This work addresses the challenge of achieving precise control over both geometry and material properties while ensuring geometric validity across diverse microstructure classes.

The gist: The proposed framework generates diverse microstructures with specified target properties while maintaining geometric validity through a latent diffusion model built upon a hybrid neural representation that explicitly encodes symmetry and implicitly captures elasticity.

How it works

1. The system utilizes a novel hybrid neural representation termed Holoplane, which embeds geometric symmetry constraints explicitly and the elasticity tensor implicitly, ensuring precise alignment between geometry and properties. This representation, denoted as a symmetric 2D snapshot of the microstructure’s geometry and physical properties, can be viewed as a symmetric 2D snapshot of the microstructure’s geometry P and physical properties C, aligning them within a unified latent space.

2. To encode microstructures into this latent space, an autoencoder is trained. This autoencoder employs a Hybrid Symmetric Representation (Sec. 4.1) which combines voxel grids (explicit) with SDFs (implicit), allowing precise symmetry capture and continuous structure representation. Furthermore, the model incorporates physical priors during training through Physics-aware Neural Embedding (Sec. 4.2), which enables the model to jointly capture both geometric and physical details.

3. A diffusion model is then employed for conditional generation within this latent space. The diffusion process is defined by the ODE: dP = − ¤sigma(t)sigma(t)∇P log p(P; sigma(t))dt. This process is conditioned on given properties, and Classifier-Free Guidance (CFG) is incorporated during inference to guide the sampling process, resulting in the conditional denoiser: Ψ(P; sigma,C) = Ψ0 + w(ΨC − Ψ0).

4. Boundary compatibility is enhanced through two mechanisms. First, a compatibility gradient loss, Lcompat = ∫Γ∥PA − PB∥2 dx, is used to guide diffusion sampling to enforce consistent boundary shapes between two microstructures during diffusion sampling. Second, an interpolation-based blending approach is adopted where spherical linear interpolation (slerp) is applied between these two noisy data points using a coefficient α to generate an interpolated Holoplane, Palpha, which is then used to reconstruct the boundary region seamlessly blending the microstructures.

Key Contributions and Methodology Details

The main contributions of this work include:

We tackle the inverse design problem for non-parametric microstructures, enabling the generation of diverse types and morphologies that satisfy target properties through a latent diffusion model.

"We introduce Holoplane, a hybrid neural representation for microstructures that enhances the alignment between geometry and property distributions, leading to higher property matching accuracy and enhanced validity control over the generated microstructures compared to existing baselines."

The proposed latent diffusion framework, built upon Holoplane, also enables optimization of boundary compatibility, achieving superior boundary integration in heterogeneous, multi-scale designs.

The training involves a diverse multi-class dataset encompassing a broad spectrum of geometric morphologies and topologies, including parametric families such as truss, shell, tube, and plate microstructures. The model is trained using a total loss L = Lphi + λ0Lchi + λ1LE + λ2LTV + λ3LEDR + λ4L2, which regularizes the Holoplane distribution toward a standard normal distribution.

Evaluation and Results

The system is evaluated using two primary metrics. The error in microstructure properties is computed as Err = (Cpred − Ctarget) / (Cmax − Cmin). For structural similarity, Sim(Ω1, Ω2) = ∫v∈V VΩ1 (v)=Ω2 (v) √Ω1 · Ω2, where the numerator is the number of intersecting voxels between the two structures and the denominator is the square root of the product of the total voxel counts.

Experimental results demonstrate superior performance compared to existing methods. The property error for MIND is 0.29% for C11, 1.27% for C12, 1.13% for C44, and 1.33% for G in Table 1. The physical validity ratio achieved by MIND is 99.2%. The model exhibits greater shape diversity than baseline methods, achieving an average similarity of 81.52% with the entire training set, lower than the 93.48% reported in [Yang et al. 2024]. Furthermore, MIND successfully generates printable structures at different printing precisions and demonstrates capability in heterogeneous design by optimizing material properties to minimize overall displacement while ensuring boundary compatibility.

Future Directions

The authors acknowledge limitations, noting that More fabrication challenges such as selfsupportiveness and the absence of closed pockets remain unaddressed and can currently only be ensured through post-filtering. Future work plans include incorporating these fabrication constraints into the loss function, expanding the properties considered beyond Young’s modulus to include isotropy, thermal conductivity, optical behavior, and exploring other material functionalities. Finally, the framework is designed to allow for optimization of material properties by solving complex inverse problems using a linear finite element method (FEM) coupled with soft constraints derived from a triangle mesh representing the feasible space of material properties.

REFERENCES

Erik Andreassen, Boyan S. Lazarov, and Ole Sigmund. 2014. Design of Manufacturable 3D Extremal Elastic Microstructure. Mechanics of Materials 69, 1 (Feb. 2014), 1–10. doi:10.1016/j.mechmat.2013.09.018

Meisam Askari, David A. Hutchins, Peter J. Thomas, Lorenzo Astolfi, Richard L. Watson, Meisam Abdi, Marco Ricci, Stefano Laureti, Luzhen Nie, Steven Freear, Ricky Wildman, Christopher Tuck, Matt Clarke, Emma Woods., and Adam T. Clare. 2020. Additive manufacturing of metamaterials: A review. Additive Manufacturing 36 (Dec. 2020), 101562. doi:10.1016/j.addma.2020.101562

Jan-Hendrik Bastek and Dennis M Kochmann, 23. Inverse design of nonlinear mechanical metamaterials via video denoising diffusion models. Nature Machine Intelligence 5, 12 (Dec. 2023), 1466–1475. doi:10.1038/s42256-023-00762-x

Jan-Hendrik Bastek, Siddhant Kumar, Bastian Telgen, Raphaël N Glaesener, and Dennis M Kochmann. 22. Inverting the Structure–Property Map of Truss Metamaterials by Deep Learning. Proceedings of the National Academy of Sciences 119, 1 (Jan.

Siddhant Kumar, Stephanie Tan, Li Zheng, and Dennis M Kochmann., 2020. Inversedesigned spinodoid metamaterials. npj Computational Materials 6, 1 (June 2020). doi:10.1038/s41524-020-0341-6

Roderic Lakes. 1987. Foam Structures with a Negative Poisson’s Ratio. Science 235, 4792 (Feb. 1987), 1038–1040. doi:10.1126/science.235.4792.1038

Davi Colli Tozoni, Jérémie Dumas, Zhongshi Jiang, Julian Panetta, Daniele Panozzo, and Denis Zorin., 2020. A Low-Parametric Rhombic Microstructure Family for Irregular Lattices. ACM Transactions on Graphics 39, 4 (Aug. 2020), 101–1. doi:10.1145/3386569.3392451

Junhao Ding, Qiang Zou, Shuo Qu, Paulo Bartolo, Xu Song., and Charlie C.L Wang., 2021. STL-free design and manufacturing paradigm for high-precision powder bed fusion. CIRP Annals 70, 1 (2021), 167–170. doi:10.1016/j.cirp.2021.03.012

Guoying Dong, Yunlong Tang, and Yaoyao Fiona Zhao., 23. A 149 Line Homogenization Code for Three-Dimensional Cellular Materials Written in Matlab. Journal of Engineering Materials and Technology 141, 1 (July 2018). doi:10.

1. Bold header: Enhanced Geometry-Property Alignment via Holoplane

This improvement involves integrating Holoplane, a hybrid neural representation, into generative models to ensure superior alignment between geometry and properties. This allows the improved AI system to generate microstructures that meet target properties while simultaneously maintaining geometric validity and achieving a higher property matching accuracy compared to existing baselines.

2. Bold header: Physics-Aware Latent Space Conditioning

By incorporating physical priors during the training of the autoencoder, the system can achieve a latent space where the physical properties and geometries are effectively aligned, as shown in Fig. 4. This enables more reliable generation by conditioning diffusion models on desired elastic tensors, leading to higher property matching accuracy.

3. Bold header: Boundary Compatibility Enforcement

The integration of a compatibility gradient and an interpolation-based blending approach allows the system to ensure perfect connectivity (Fig. 5c) at boundaries during heterogeneous design. This enables the AI to generate structures that are not only individually valid but also seamlessly fitting into complex assemblies.

4. Bold header: Scalable Heterogeneous Design via FEM Optimization

The framework can be used for optimization by implementing a linear finite element method (FEM) based on hexahedron discretization to determine material property distributions that best approximates the specified target behavior. This allows the AI system to generate microstructures optimized for specific mechanical requirements, such as minimizing overall displacement.

5. Bold header: Diverse Topology and Morphology Exploration

The model can explore a vast design space by enabling the generation of diverse microstructure types and morphologies that meet these target properties, without being constrained by predefined parametric classes. This moves beyond limitations of existing methods that limit design flexibility and structural diversity.