Efficiency is one of the key problems in the design of high-throughput materials computing. In this paper, we provide a Self-Evaluation High-throughput Computing framework (SEHC). The framework introduces an automatic self-evaluation filtering mechanism, which is based on machine learning, for high-throughput computing architectures to stop unexpected materials calculation tasks in advance during high-throughput calculation. The time-consuming high-throughput computing process is disassembled into several finer-grained high-throughput Stages. Multiple high-throughput Stages with the same standard design specifications can be assembled into a Pipeline model. Combined with the public service like data storage and system monitoring, the SEHC with a “Stage-Pipeline-Framework” three-tier structure is formed. To search for diamond-like structures with higher group velocity in a space of 254 compounds, a SEHC-based prototype was implemented. The experiment result shows that this prototype achieved a significant improvement in efficiency by reducing the amount of invalid computation remarkably.

效率是高通量材料计算设计中的关键问题之一。在本文中，我们提供了一个自我评估高通量计算框架（SEHC）。该框架引入了一种基于机器学习的自动自我评估过滤机制，用于高通量计算体系结构，可以在高通量计算过程中提前停止意外的物料计算任务。耗时的高通量计算过程分解为几个更细粒度的高通量阶段。可以将具有相同标准设计规范的多个高通量级组装为流水线模型。结合数据存储和系统监控等公共服务，形成具有“阶段-管道-框架”三层结构的SEHC。为了在254种化合物的空间中搜索具有更高群速度的类金刚石结构，实施了基于SEHC的原型。实验结果表明，该样机通过显着减少了无效计算量，实现了效率的显着提高。