基于机器学习的分子动力学模拟获得戈登·贝尔奖

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"2020年"},{"type":"link","attrs":{"href":"https:\/\/www.acm.org\/","title":"","type":null},"content":[{"type":"text","text":"美国计算机协会"}]},{"type":"text","text":"（ACM）的"},{"type":"link","attrs":{"href":"https:\/\/www.acm.org\/media-center\/2020\/november\/gordon-bell-prize-2020","title":"","type":null},"content":[{"type":"text","text":"戈登·贝尔奖"}]},{"type":"text","text":"授予了来自美国和中国机构的研究团队，表彰他们题为“用机器学习将分子动力学的从头计算方法的精度推至1亿个原子”的项目。据团队介绍，深势分子动力学（Deep Potential Molecular dynamics，DPMD）是一种基于机器学习的新协议，它能够每天模拟超过1亿个原子的1纳米的轨迹。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"http:\/\/www.sciencedirect.com\/topics\/biochemistry-genetics-and-molecular-biology\/molecular-dynamics","title":"","type":null},"content":[{"type":"text","text":"分子动力学（Molecular Dynamics）"}]},{"type":"text","text":"是一种计算机模拟方法，用来分析在特定的时间段内原子的运动和相互作用。从小到单细胞的系统，到大到气体云的复杂系统，科学家都能利用分子动力学模拟的方式来了解这些分子化合物在一段时间内的行动。三十五年来，研究人员一直在使用一种被称为"},{"type":"link","attrs":{"href":"https:\/\/www.pnas.org\/content\/102\/19\/6654","title":"","type":null},"content":[{"type":"text","text":"从头计算（ab initio）的模拟方法进行分子动力学研究"}]},{"type":"text","text":"，因为事实证明这种方式是最有效的。尽管从头计算方法（在拉丁语中，它的意思是“从第一原理”）在模拟中取得了很高的精确度，但是这种方式需要大量的计算资源，这限制了它的应用，最多只能用于包含数千个原子的小型系统。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"DPMD背后的团队在"},{"type":"link","attrs":{"href":"https:\/\/dl.acm.org\/doi\/pdf\/10.5555\/3433701.3433707","title":"","type":null},"content":[{"type":"text","text":"本论文"}]},{"type":"text","text":"中详细介绍了“从头计算”方法的局限性，发现它随电子自由度（electronic degrees of freedom）的数量呈立方级关系。采用从头计算方法可以实现的典型时空比例的设置是~100个原子和~10皮秒。从头计算方法几乎完美地遵守立方扩展定律。即便是世界上最大的超级计算机，也无法进行复杂的化学反应、电化学电池、纳米晶体材料和辐射损伤等方面的模拟。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"DP（深度势能，Deep Potential）模型的精确性来源于"},{"type":"link","attrs":{"href":"https:\/\/www.nextplatform.com\/2020\/04\/07\/changing-conditions-for-neural-network-processing\/","title":"","type":null},"content":[{"type":"text","text":"深度神经网络（DNN）"}]},{"type":"text","text":"对高维函数独特的近似能力、对对称性约束等物理要求的恰当处理以及并发的学习模式，能够生成紧凑的训练数据集，并保证在相关配置空间中统一的精确度。"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"DPMD团队选择利用世界第二快的超级计算机"},{"type":"link","attrs":{"href":"https:\/\/www.ibm.com\/thought-leadership\/summit-supercomputer\/","title":"","type":null},"content":[{"type":"text","text":"IBM的Summit系统"}]},{"type":"text","text":"上的GPU来运行几乎所有的计算和通信任务。由于“深度势能”模型中的计算粒度的限制，该团队发现，仅仅严重依赖GPU的效率会很低。通过算法创新，包括为相邻列表提供新的数据布局以避免嵌入式矩阵计算中的分支，将新数据结构中的元素压缩为64位整数以提高GPU对自定义"},{"type":"link","attrs":{"href":"https:\/\/www.tensorflow.org\/","title":"","type":null},"content":[{"type":"text","text":"TensorFlow"}]},{"type":"text","text":"操作的优化，以及为深度势能模型创建混合精度计算，团队针对GPU相关的低效率进行了优化。通过这些改进，研究人员能够以从头计算计算相同的精度模拟前所未有的规模和时间范围。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"戈登·贝尔奖旨在表彰高性能计算领域的成就，入围者必须要证明他们的算法能够在世界最强大的超级计算机上进行扩展。GPU Deep MD-Kit能够有效地扩展到整个Summit超级计算机上，在单\/半混合精度下达到91 PFLOPS（一个PFLOPS指的是每秒1千万亿次的浮点运算——译注）和162\/275 PFLOPS。这一成绩为下一代超级计算机更好地实现机器学习和物理建模的结合提出了新的挑战。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"原文链接："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/www.infoq.com\/news\/2020\/12\/ml-based-molecular-dynamics\/","title":"","type":null},"content":[{"type":"text","text":"Molecular Dynamics Simulation Based on Machine Learning Wins Gordon Bell Prize"}]}]}]}