Exploring Material Design Space with a Deep-Learning Guided Genetic Algorithm
Exploring Material Design Space with a Deep-Learning Guided Genetic Algorithm
复制标题
DOI：
10.4230/lipics.dna.28.4
复制
发表时间：
2022
期刊：
影响因子：
--
通讯作者：
Kuan-Lin Chen;Rebecca Schulman
中科院分区：
文献类型：
--
作者：
Kuan-Lin Chen;Rebecca Schulman
设计复杂的、动态的、多功能的材料和设备是一项挑战，因为这些材料的设计空间有许多相互依赖且经常相互冲突的限制。从人工智能的进展及其在材料发现中的应用中获得灵感，我们提出了一种设计变性DNA共聚水凝胶结构的计算方法。该方法由粗粒度模拟和深度学习引导的优化系统组成，用于探索这些结构的巨大设计空间。在这里，我们开发了一个简单的DNA共聚水凝胶形状变化的数值模拟，并试图找到结构化水凝胶的设计，可以折叠成不同的阿拉伯数字的形状在不同的驱动状态。我们训练了一个卷积神经网络来对粗粒度模拟的几何输出进行分类和评分，为设计优化提供自主反馈。然后，我们构建了一个遗传算法，生成和选择大量的材料设计，相互竞争，以发展和收敛到最佳的目标匹配设计。我们表明，我们能够探索大的设计空间，学习重要的参数和特点。我们确定的材料规模大小和形状变化的范围，可以实现由个别域之间的重要关系，我们阐明不同的设计参数之间的权衡。最后，我们发现材料设计能够在不同的驱动状态下转换成多个不同的数字。
Designing complex, dynamic yet multi-functional materials and devices is challenging because the design spaces for these materials have numerous interdependent and often conflicting constraints. Taking inspiration from advances in artificial intelligence and their applications in material discovery, we propose a computational method for designing metamorphic DNA-co-polymerized hydrogel structures. The method consists of a coarse-grained simulation and a deep learning-guided optimization system for exploring the immense design space of these structures. Here, we develop a simple numeric simulation of DNA-co-polymerized hydrogel shape change and seek to find designs for structured hydrogels that can fold into the shapes of different Arabic numerals in different actuation states. We train a convolutional neural network to classify and score the geometric outputs of the coarse-grained simulation to provide autonomous feedback for design optimization. We then construct a genetic algorithm that generates and selects large batches of material designs that compete with one another to evolve and converge on optimal objective-matching designs. We show that we are able to explore the large design space and learn important parameters and traits. We identify vital relationships between the material scale size and the range of shape change that can be achieved by individual domains and we elucidate trade-offs between different design parameters. Finally, we discover material designs capable of transforming into multiple different digits in different actuation states.