No abstract

Metamaterials can be designed to contain functional gradients with negative Poisson's ratio (NPR) that have counterintuitive behavior compared with monolithic materials. These NPR materials, referred to as auxetics, are relevant to engineering sciences because of their unique mechanical expansion. Previous studies have explored compliant actuators using analytical and numerically derived mechanics of materials principles. However, the control of compliant gradient mechanisms frequently uses complex analytical equations combined with traditional control algorithms, making them difficult to design. To confront the design processes and computational load, herein, machine learning is used to predict errors in compliant auxetic designs based on a mathematically optimal deformation. Finite element analysis and experimental specimens validate the theoretical mechanical behavior of a specific auxetic configuration as well as demonstrate the capabilities of additive manufacturing of graded auxetic materials. Pseudorandomized images and their respective computational deformation results are used to train a regressive model and predict the deviation from optimal behavior. The model predicts the deviation from the desired behavior with a mean average percent error below 5% for the validation set. Subsequently, a scalable workflow design process connecting the unique performance of auxetics to machine learning design predictions is proposed.

Neural network-based generative models have been actively investigated as an inverse design method for finding novel materials in a vast design space. However, the applicability of conventional generative models is limited because they cannot access data outside the range of training sets. Advanced generative models that were devised to overcome the limitation also suffer from the weak predictive power on the unseen domain. In this study, we propose a deep neural network-based forward design approach that enables an efficient search for superior materials far beyond the domain of the initial training set. This approach compensates for the weak predictive power of neural networks on an unseen domain through gradual updates of the neural network with active transfer learning and data augmentation methods. We demonstrate the potential of our framework with a grid composite optimization problem that has an astronomical number of possible design configurations. Results show that our proposed framework can provide excellent designs close to the global optima, even with the addition of a very small dataset corresponding to less than 0.5% of the initial training dataset size.

High myopia, which is extremely prevalent in the Chinese population, is one of the leading causes of blindness in the world. Genetic factors play a critical role in the development of the condition. To identify the genetic variants associated with high myopia in the Han Chinese, we conducted a genome-wide association study (GWAS) of 493,947 SNPs in 1088 individuals (419 cases and 669 controls) from a Han Chinese cohort and followed up on signals that were associated with p < 1.0 × 10(-4) in three independent cohorts (combined, 2803 cases and 5642 controls). We identified a significant association between high myopia and a variant at 13q12.12 (rs9318086, combined p = 1.91 × 10(-16), heterozygous odds ratio = 1.32, and homozygous odds ratio = 1.64). Furthermore, five additional SNPs (rs9510902, rs3794338, rs1886970, rs7325450, and rs7331047) in the same linkage disequilibrium (LD) block with rs9318086 also proved to be significantly associated with high myopia in the Han Chinese population; p values ranged from 5.46 × 10(-11) to 6.16 × 10(-16). This associated locus contains three genes-MIPEP, C1QTNF9B-AS1, and C1QTNF9B. MIPEP and C1QTNF9B were found to be expressed in the retina and retinal pigment epithelium (RPE) and are more likely than C1QTNF9B-AS1 to be associated with high myopia given the evidence of retinal signaling that controls eye growth. Our results suggest that the variants at 13q12.12 are associated with high myopia.

Over the past decades, significant effort has been made to improve the adhesive properties of adhesive pillars, by searching for pillar shapes with optimized interfacial stress distribution. However, the shape optimizations in the previous studies are conducted by considering specific pillar forms with a few parameters, hence with limited design space. In this study, we present a framework to find a free-form optimized adhesive pillar shape out of extensive design space. We generate 200 000 different shapes of adhesive pillars based on the Bézier curve with a few control points by considering two distinct edge shapes, sharp and truncated edges, to account for the limitation in the realistic manufacturing resolution. The resulting interfacial stress distributions from numerical simulations are used to train deep neural networks for each edge type. Our deep learning model shows greater than 99% classification accuracy on a limited data set with orders of magnitude speedup in computation time compared to finite element analyses. On the basis of the trained neural network, we conduct genetic optimization by maximizing a fitness function that prefers the uniform interfacial stress distribution with neither stress peak nor singularity. The optimized adhesive pillar shape is composed of smoothly mixed convex and concave parts and shows improved uniformity in the interfacial stress distribution. Our study also demonstrates that the deep learning can be used for nonparametric curve optimization task with diverse fitness function.