A Three-Dimensional Generative Framework for Designing High Performance Complex Material Structures

Generative design for materials has recently gained significant attention due to the rapid evolution of generative deep learning models. There have been successful demonstrations of molecular-level generative design, including conditional generation based on specific criteria¹. In the realm of macroscopic structures, voxel representations of two-dimensional^2,3 and three-dimensional⁴ structures have been extensively studied. However, the challenge lies in the insufficient resolution of these models, hindering their ability to accurately represent real structures. Various other approaches based on point cloud^5,6, mesh⁷, and neural radiance fields (NeRF)^8–10 have been explored but have yet to be introduced in the fields of material science. In this study, we aim to address these issues with a novel framework consisting of a mesh-based 3D structure generator and physics-informed functions. We utilize non-linear mapping networks as soft constraints to capture the complex nature of 3D material structures. The incorporated physics-based methods are capable of evaluating the performances of generated structures in real-time. The composition of the loss function considers these evaluation results to drive active structure optimization. Our framework endeavors to achieve high-resolution 3D structures generation guided by desired material properties, expected as a significant step forward toward the inverse design of novel functional materials. <br/><br/>References<br/>1. Luo, S., Guan, J., Ma, J. & Peng, J. A 3D generative model for structure-based drug design. <i>Adv. Neural Inf. Process. Syst.</i> <b>8</b>, 6229–6239 (2021).<br/>2. Mao, Y., He, Q. & Zhao, X. Designing complex architectured materials with generative adversarial networks. <i>Sci. Adv.</i> <b>6</b>, (2020).<br/>3. Qian, C., Tan, R. K. & Ye, W. Design of architectured composite materials with an efficient, adaptive artificial neural network-based generative design method. <i>Acta Mater.</i> <b>225</b>, 117548 (2022).<br/>4. Nguyen, P. C. H. <i>et al.</i> Synthesizing controlled microstructures of porous media using generative adversarial networks and reinforcement learning. <i>Sci. Rep.</i> <b>12</b>, 1–16 (2022).<br/>5. Luo, S. & Hu, W. Diffusion probabilistic models for 3D point cloud generation. in <i>2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)</i> (2021).<br/>6. Wu, W., Qi, Z. & Fuxin, L. PointConv: Deep Convolutional Networks on 3D Point Clouds. in <i>2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)</i> (2019).<br/>7. Gao, J. <i>et al.</i> GET3D: A generative model of high quality 3D textured shapes learned from images. in <i>Advances in Neural Information Processing Systems 35 (NeurIPS)</i> 1–14 (2022).<br/>8. Shen, J., Agudo, A., Moreno-Noguer, F. & Ruiz, A. Conditional-Flow NeRF: accurate 3D modelling with reliable uncertainty quantification. in <i>ECCV 2022: 17th European Conference</i> 540–557 (2022).<br/>9. Jun, H. & Nichol, A. Shap-E: Generating conditional 3D implicit functions. <i>arXiv</i> (2023).<br/>10. Mildenhall, B. <i>et al.</i> <i>NeRF: Representing scenes as neural radiance fields for view synthesis</i>. <i>European Conference on Computer Vision (ECCV)</i> (2020).