SI2-SSE: General Tensor Software Elements for Quantum Chemistry, Tensor Network Theories, and Beyond

PI将开发可重用的软件组件，以加速依赖高维张量计算的科学创新。虽然这些组件是由来自量子化学的用例提供的，但张量计算的挑战是普遍的，并且跨越了科学和工程的各个领域。这些问题的范围从核自旋谱的模拟，到分子的量子化学计算，到心理测量分析，再到数值广义相对论。资助工作的总体目标是基于现代可持续软件实践开发可重用的张量软件元素，以使科学界的张量算法开发受益。此外，除了软件的广泛科学影响之外，我们的教育和推广议程包括多层努力，以提高科学和工程界对高维张量和矩阵计算的推理能力。软件元素的重要方面将包括(i)用于基于张量理论的快速原型的表达性编程接口（ii）用于密集、块稀疏和核外张量的分层张量库，提供上述接口的峰值性能实现，（iii）用于一般高维计算的基于矩阵乘积状态方法的多线性代数包。（iv）张量虚拟机技术，它将算法开发从硬件中抽象出来，并为优化编译器转换提供了一个框架，以使算法适应现代计算机体系结构的内存访问、通信网络和处理器特性。

The PI will develop reusable software components to accelerate innovation in science that relies on high-dimensional tensor computations. While the components are informed by use-cases taken from quantum chemistry, the challenges of tensor computation are universal, and span diverse areas of science and engineering. These problems range from simulations of nuclear spin spectra, to quantum chemical calculations on molecules, to psychometric analysis, to numerical general relativity. The overarching aim of the funded work is to develop reusable tensor software elements, based on modern sustainable software practices, to benefit tensor algorithmic development in the scientific community at large. In addition, beyond the broad scientific impacts of the software, our education and outreach agenda comprises a multi-tiered effort to uplift the ability of the science and engineering community to reason about high-dimensional tensor and matrix computations.Important aspects of the software elements will include (i) expressive programming interfaces for rapid prototyping of tensor based theories (ii) layered tensor libraries for dense, block-sparse, and out-of-core tensors that provide peak-performance implementations of the above interfaces, (iii) a multi-linear algebra package for general high dimensional computations, based on the matrix product state approach, and (iv) tensor virtual machine technology that abstracts algorithm development from hardware, and which provides a framework for optimizing compiler transformations to adapt algorithms to the memory access, communication networks, and processor characteristics of modern computer architectures.