A Programming Framework for Physics

The fundamental difficulty with scientific software is that the existing languages are either<br>too general-purpose and too low-level, or both. Thus, physics equations are entangled with the<br>way they are implemented, even to low level details such as grid structure, cache optimization,<br>etc. Because everything is coupled, it is difficult for the various scientific communities to adapt to<br>algorithm or hardware advances. More importantly, it means that researchers cannot specialize<br>as easily. Everyone must understand physics, software engineering, and computer science to a<br>to a very deep level, and the learning process, especially for students, is much more arduous<br>than it needs to be.<br>In principle, it should be possible to independently specify (1) the scientific equations to be<br>solved; (2) the type of grids; (3) the type of numerical methods, e.g. time integrators, elliptic<br>solvers; (4) the intended execution platform (desktop, accelerated cluster, etc.); (5) performance<br>goals, e.g. as fast as possible, minimum cost, etc.

科学软件面临的核心困境在于，现有编程语言要么通用性过强且层级过低，要么兼具二者弊端。因此，物理方程与其实现方式深度绑定，甚至涉及网格结构、缓存优化这类底层实现细节。由于所有模块高度耦合，不同科研领域的从业者难以适配算法或硬件的迭代升级。更关键的是，这导致研究者难以进行专业化深耕。所有人都需要在物理学、软件工程与计算机科学领域达到极深的造诣，而学习过程——尤其对于学生而言——远超必要的繁重程度。

原则上，我们应当可以独立指定以下五项内容：(1) 待求解的科学方程；(2) 网格类型；(3) 数值方法类型，例如时间积分器、椭圆方程求解器；(4) 目标执行平台（桌面设备、加速集群等）；(5) 性能目标，例如极致速度、最低成本等。