EAGER: Collaborative Research: Inverse Procedural Material Modeling for Battery Design

如今普遍使用的几乎所有便携式电子设备--相机、手机、音乐播放器等--都依赖可充电的锂离子电池。这些设备的能力可以通过改进这些电池的设计来实现。这项工作将为设计具有高功率输出和长寿命等理想性能的电池提供新的计算方法。新的计算方法将使用在电影和游戏的计算机图形学中成功描述复杂体积结构(如多孔岩石和海绵)的技术。这些计算机图形技术将被应用于描述电池中的材料。新的计算方法将专注于产生正确的性能特征(如功率密度)的结构，而不是专注于寻找能够提供正确视觉外观的体积结构。新的体积描述将用于生成大量潜在的体积材料，这些模型将从电池属性和性能方面进行表征。利用最近开发的机器学习技术，这些大量的潜在模型将被转换成一种便于在电池设计中使用的形式。除了提供制造改进的便携式电池的工具外，新的计算方法还有可能进一步扩展和应用于其他涉及复杂体积结构的材料的问题，例如了解地质测量和设计文化遗产古迹和文物的保护策略。电池设计的一个直接方法是对材料微结构进行理论推导，运行正向模拟来评估其性能，并评估结果。然而，模拟需要几个小时(在当前的多核系统上，功率密度计算需要长达50个小时)，使得正向模拟对于迭代设计来说成本高得令人望而却步。如果有反函数可用来产生给定所需性能特性的微结构描述，则设计过程可以显著改进。创建这种反函数的障碍是微观结构描述的复杂性以及结构和性能之间的关系。要创建反函数，我们需要比高分辨率网格的完整枚举在维度上更低的微结构描述。程序模型可以提供这样一个较低维度的描述。在这个项目中探索的寻找合适的程序模型的方法是基于组合和转换在其他问题领域已经成功的模型来拟合来自真实电池材料测量的数据。在给定适当的程序模型的情况下，设计问题简化为确定生成输入的程序模型参数；这一问题称为“反向程序建模”。即使有一个紧凑的微观结构描述，这个问题也太复杂了，不能用数学上的反转。不是尝试寻找数学函数，而是使用机器学习(深度神经网络)。微结构及其性能特征的数据库将与根据程序模型参数的大样本计算的示例微结构进行综合填充。将在这些样本上运行正向模拟，以计算属性(曲折度和面密度)和性能特征(功率和能量密度)。然后，将使用机器学习优化来找出模型参数和性能特性之间的关系，并将这种关系用于设计过程。找到适合数据的程序模型，然后从模型生成的合成数据中学习关系的整体方法将新的数据驱动方法的力量带到了电池设计领域。该项目产生的软件、数据和出版物将在项目网站(http://hpcg.purdue.edu/Eager2018/).)上提供

Nearly all portable electronic devices commonly used today -- cameras, phones, music players and the like -- rely on rechargeable Lithium-ion batteries. Improvements in the capabilities of these devices can be achieved by improving the design of these batteries. This work will produce new computational methods for designing batteries with desirable properties such as high power output and long lifespan. The new computational methods will use techniques that have successfully described complex volumetric structures (such as porous rocks and sponges) in computer graphics for film and games. These computer graphics techniques will be applied to describing the materials in batteries. Instead of focusing on finding volumetric structures that give the correct visual appearance, the new computational methods will focus on structures that produce the correct performance characteristics such as power density. The new volumetric descriptions will be used to generate a large number of potential volumetric materials, and these models will be characterized in terms of battery properties and performance. Using recently developed machine learning techniques, this large number of potential models will be converted into a form that is convenient to use in battery design. In addition to providing tools to create improved portable batteries, the new computational methods have the potential to be further extended and applied to other problems involving materials with complex volumetric structure such as understanding geologic measurements and designing conservation strategies for cultural heritage monuments and artifacts.A straightforward approach to battery design is to theorize material microstructures, run forward simulations to assess their performance, and evaluate the results. However, simulations require hours (up to 50 hours on current multi-core systems for power density calculations), making forward simulation prohibitively expensive for iterative design. The design process can be dramatically improved if an inverse function is available that can produce a microstructure description given desired performance characteristics. Barriers to creating such an inverse function are the complexity of microstructure descriptions and the relationship between structure and performance. To create an inverse function, we need a microstructure description that is lower in dimension than a full enumeration of a high-resolution grid. A procedural model can provide such a lower dimensional description. The approach explored in this project for finding appropriate procedural models is based on combining and transforming models that have been successful in other problem domains to fit data from real battery material measurements. Given an appropriate procedural model, the design problem is reduced to determining the procedural model parameters that generate the input; a problem called "inverse procedural modeling". Even with a compact microstructure description, the problem is too complex to be mathematically inverted. Rather than attempt to find a mathematical function, machine learning (deep neural networks) are used. A database of microstructures and their performance characteristics will be populated synthetically with example microstructures computed from a large sampling of procedural model parameters. Forward simulations will be run on these samples to compute properties (tortuosity and area density) and performance characteristics (power and energy density.) Machine learning optimizations will then be used to find the relationship between model parameters and performance characteristics and this relationship will be used in the design process. The overall method of finding procedural models to fit data and then learning the relationships from synthetic data generated from the models brings the power of new data-driven approaches to the domain of battery design. The software, data and publications resulting from this project will be available at the project website (http://hpcg.purdue.edu/Eager2018/).