Heterogeneous Metamaterials Design via Multiscale Neural Implicit Representation

TL;DR

The paper tackles the challenge of designing heterogeneous metamaterials by introducing a four-dimensional, coordinate-based neural network that jointly represents macro and micro structures as a continuous density field. A compatibility loss is employed to ensure smooth transitions between neighboring unit cells, enabling seamless, high-resolution designs without relying on a predefined microstructure library. The approach leverages homogenization to connect micro and macro scales and demonstrates effectiveness across displacement-targeting, NPR, mechanical cloaking, and bulk-modulus optimization, with a mini-epoch training strategy that preserves connectivity while reducing computation. The results show strong connectivity and fidelity, with substantial runtime advantages over traditional concurrent multiscale topology optimization and data-driven pipelines, making high-resolution metamaterial design more scalable and practical for engineering applications.

Abstract

Metamaterials are engineered materials composed of specially designed unit cells that exhibit extraordinary properties beyond those of natural materials. Complex engineering tasks often require heterogeneous unit cells to accommodate spatially varying property requirements. However, designing heterogeneous metamaterials poses significant challenges due to the enormous design space and strict compatibility requirements between neighboring cells. Traditional concurrent multiscale design methods require solving an expensive optimization problem for each unit cell and often suffer from discontinuities at cell boundaries. On the other hand, data-driven approaches that assemble structures from a fixed library of microstructures are limited by the dataset and require additional post-processing to ensure seamless connections. In this work, we propose a neural network-based metamaterial design framework that learns a continuous two-scale representation of the structure, thereby jointly addressing these challenges. Central to our framework is a multiscale neural representation in which the neural network takes both global (macroscale) and local (microscale) coordinates as inputs, outputting an implicit field that represents multiscale structures with compatible unit cell geometries across the domain, without the need for a predefined dataset. We use a compatibility loss term during training to enforce connectivity between adjacent unit cells. Once trained, the network can produce metamaterial designs at arbitrarily high resolution, hence enabling infinite upsampling for fabrication or simulation. We demonstrate the effectiveness of the proposed approach on mechanical metamaterial design, negative Poisson's ratio, and mechanical cloaking problems with potential applications in robotics, bioengineering, and aerospace.

Figures (12)

• Figure 1: The neural network (c) takes the global $(x,y)$ and local $(u,v)$ coordinates sampled across the design domain (a) as input and outputs density $\rho$ at each coordinate point. The density $\rho$ is then assembled and fed into the finite element analysis and other performance metric analysis to form a combined loss function. We use backpropagation to update the trainable kernel and weights of the neural network. We can obtain the metamaterial (d) by querying the design domain at its original resolution. We can also upsample the global coordinates and solve for a subset of cells during each epoch with mini-epoch solving. This enables us to solve a bigger problem with smoother continuity without a large increase in runtime. (e) Once good continuity is observed, we can then infinitely upsample the global coordinates to obtain an extremely large-scale metamaterial design without compromising the connectivity between cells.
• Figure 2: (a) The boundary condition setup for the metamaterial design problem, where the two sides are fixed, and a load is applied from the bottom center with the target displacement forming a bump at the top. (b) The deformed mesh corresponding to the optimized structure. (c) The converged geometry after 300 epochs of optimization using Adam with a learning rate of 0.001. (d) The deformed structure under the applied displacement field, demonstrating the connectivity and structural integrity of the optimized design. The Root Mean Square Error (RMSE) between the computed and target displacements is 0.556
• Figure 3: (a, c, e) The results of mini-epoch training with different global coordinate upsampling scales: no mini-epoch, 2× upsampling, and 4× upsampling, respectively. (b) The no-mini-epoch case upsampled post-training to 4× global coordinates, showing poor connectivity in the transition between microstructure cells. (d) The 2× upsampling case, further upsampled to 4× global coordinates after training, demonstrating improved connectivity. (f) The native 4× upsampled mini-epoch case, which exhibits the best connectivity among the tested configurations. A slight asymmetry is observed in the mini-epoch cases, likely due to the linear grid-like selection of representative cells during training, which does not preserve perfect symmetry across epochs. (g) Further upsampling of the 4$\times$ mini-epoch result to 16$\times$, we observe that good connectivity is still maintained.
• Figure 4: (a) The displacement boundary condition where we apply a unit displacement on the right-hand side. (b) Plotting out the full scale evaluated displacement with the target displacement. We observe that all full-scale analyses showed slightly smaller displacements than the target. (c) 12$\times$4 resolution without mini-epoch (u_11), with a full-scale resolution of 360$\times$120. (d) 24$\times$8 resolution with mini-epoch (u_12), with each batch consisting of 12$\times$4 macro cells and a full-scale resolution of 720$\times$240. (e) 24$\times$8 resolution without mini-epoch (u_22), with a full-scale resolution of 720$\times$240.
• Figure 5: Island removal with thresholding. We first remove the dangling islands at a 0.3 cutoff. Then the cutoff is increased to 0.5, which causes some dangling edges to become a disconnected island. We then use these new disconnected islands as a mask to remove the dangling edges in the previous 0.3 cutoff. Here we use the result shown in Figure \ref{['fig:mm_nn']} (c) to demonstrate the post-processing. (a) Dangling edges and islands are detected and highlighted in red. (b) Result with dangling edges and islands removed. We note that some dangling edges remain, which would necessitate more comprehensive methods.