主 题: 气固两相流的粗粒化颗粒轨道模拟
报告人: 葛蔚研究员 （中国科学院过程研究所）
时 间: 2014-05-29 16:00-17:00
地 点: 理科一号楼1303室（主持人：李铁军）
Discrete elements (particles, grains, powders, droplet or bubbles, etc.) are encountered in process engineering as often as continuua, not to mention that continuua themselves are discrete at microscale. It is estimated that bulk materials in the form of powders and grains hold the second largest quantity in process engineering, only next to water. However, their properties and interactions are much more complex and less understood theoretically than continua, and experiments can hardly reveal their dynamical details. Direct simulation of the interactions and movement of the individual elements, therefore, becomes important for process development and engineering, such as for material design and reactor scale-up. However, industrial systems typically involve billions to trillions of elements, so the computational costs will be virtually unaffordable.
To address such challenges, we conducted a systematic co-design of the physical model, numerical software and computer hardware, which can greatly accelerate discrete simulation while reduces its cost. We notice that discrete elements share some very favorable features for scalable parallel computing, such as locality and additivity, which ensures elements at different locations and different interactions on the same element can be processed simultaneously, and more importantly, a coarse-grained element may represent a collection of real elements to reduce the computational cost greatly. Taking advantages of these features, we developed a general algorithmic platform for different discrete systems without trading off its accuracy and efficiency. Furthermore, we introduced many-core processors to the traditional supercomputer architecture which can accelerate the execution of the software by several orders.
The software and hardware described above can be applied in multi-scale simulation of discrete systems on different levels: Molecular dynamics simulation of condensed matter can now achieve petaflops sustainable performance, linking angstrom and millimeter scales directly; direct numerical simulation of multiphase fluid flow with micron-scale particle can be carried at meter-scale; and at the reactor scale, with coarse-grained models, discrete simulation can already serve as a practical alternative of traditional continuum-based simulations as employed in mainstream commercial softwares.
We believe, with further optimization of the model and algorithm, the speed and scale of discrete simulation can be elevated again by several orders in near future, which may finally lead to realtime simulation of industrial systems without scarifying its resolution and accuracy. That means virtual reality of industrial processes is turning from fictions to practice, which will be an exciting revolution to the technological development and engineering of process industries.