计算化学公社

标题: 算来算去都是算个寂寞 [打印本页]

作者
Author: gauss98 时间: 2022-6-17 13:06
标题: 算来算去都是算个寂寞
算来算去，都是算个寂寞
一个反应，一两个千卡就可以把选择性从 10%搞到 90%，把活性增减十倍一百倍
结果，算起来，基组不同都可以误差七八个千卡，方法选择七八个千卡，溶剂效应，频率计算，自由能模型的选择
哪个环节都可以整出几个千卡的误差出来，哪怕号称最精确的 dlpno-ccsd（t），最高的基组选择 def2qzvpp到 CBS都能整出个一两个千卡的误差出来。

如果有实验结果，总有一套参数去凑，如果没有，嘿嘿，天知道
整疲倦了。

作者
Author: wzkchem5 时间: 2022-6-17 14:18
其实还好，如果是计算两个特别类似的反应的选择性，那么是可以达到远高于1~2kcal/mol的精度的，比如手性催化，生成两个手性对映体的反应极其类似，误差抵消得很充分，因此只要做好构象搜索，把ee值算准一般不成问题。

作者
Author: chands 时间: 2022-6-17 14:23
本帖最后由 chands 于 2022-6-17 14:24 编辑

这么说来，做实验做来做去也是个寂寞。本人领域内的动力学数据compilation，会对各实验室用各种方法各种表征手段测出的数据做对比，然后做评估，挑出比较可信的，但是那些动力学数据，有差几倍的，也有差一两个量级的。做过实验估计都有体会，明明每一步都控制很好，最后就不知道误差哪里来。文献中的某些热力学数据差几个kcal/mol也不是什么大事。实验有误差，计算有误差，能控制的尽量控制，正视误差，小心下结论。

作者
Author: 风起~ 时间: 2022-6-17 15:43
本帖最后由风起~ 于 2022-6-17 17:06 编辑

我一直觉得，某一套计算级别对于某一类反应的‘准确度’是比较好的
而摸索这一套计算级别就太花时间了。。。

作者
Author: sobereva 时间: 2022-6-17 16:12
这正体现出理论知识和经验积累的重要性，从而准确判断出什么计算模型对当前研究的问题能得到靠谱的结果，在什么方面什么计算方法/方式会引入大约何等程度的误差，以及认清对于一个问题以当下的计算方法和计算条件是否真有办法进行合理的计算。

作者
Author: gauss98 时间: 2022-6-17 16:37

wzkchem5 发表于 2022-6-17 14:18
其实还好，如果是计算两个特别类似的反应的选择性，那么是可以达到远高于1~2kcal/mol的精度的，比如手性催 ...

我原来也是这么认为的，
这次就算个模型反应，打算用高精度方法来验证下，才发现精度远远的没那么高。

就算个烯烃在催化剂上的12加成与 2,1加成的选择性，
按照低精度热力学，高精度电子能的做法，
发现光是 PBE0 与 B3LYP 在频率自由能校正上，都有1kcal的不同，用低频校正算自由能，低频选取 50,100cm-1又有不同，溶剂中算频率和气象算频率也不同，电子能就算用了 dlpnoccsdt(t), 基组取到 qzvpp了还跟CBS不同，差距都在1kcal左右，可是远没有站长说的那么理想哈

原来用DFT算了几个，每次兴冲冲费劲巴力搞来药品拿去做实验，嘿嘿，一个都没成

作者
Author: sobereva 时间: 2022-6-17 18:33

gauss98 发表于 2022-6-17 16:37
我原来也是这么认为的，
这次就算个模型反应，打算用高精度方法来验证下，才发现精度远远的没那么高。
...

后HF相关能随基组尺寸收敛较慢，def2-QZVPP和CBS有不可忽略的差异很正常，要不然的话基组外推就没意义了

不要用Truhlar的QRRHO那种一刀切的做法，始终用Grimme的QRRHO，原理好得多也没有阈值的任意性

作者
Author: wzkchem5 时间: 2022-6-17 18:59

gauss98 发表于 2022-6-17 09:37
我原来也是这么认为的，
这次就算个模型反应，打算用高精度方法来验证下，才发现精度远远的没那么高。
...

另外也可以注意一下构象搜索的问题，如果你现在选的是单一构象的话，可能因为两个过渡态的构象差别比较大，误差抵消不充分。如果做了构象搜索，往往会发现一个过渡态的所有主要构象在另一个过渡态里面都有对应，所以做构象搜索以后不仅结果更准，结果受理论级别的影响可能也会小些

作者
Author: chemicalchange 时间: 2022-6-17 19:20

gauss98 发表于 2022-6-17 16:37
我原来也是这么认为的，
这次就算个模型反应，打算用高精度方法来验证下，才发现精度远远的没那么高。
...

虽然不知道楼主具体的体系，但烯烃上加成的区域选择性有时候也不一定是过渡态理论能定量解释的（比如和硼烷加成的过程），这个能量算准了可能也不一定就解决问题。

作者
Author: linhai093 时间: 2022-6-19 15:24

sobereva 发表于 2022-6-17 16:12
这正体现出理论知识和经验积累的重要性，从而准确判断出什么计算模型对当前研究的问题能得到靠谱的结果，在 ...

请问基础班在方面会有多少涉及和讲解？

作者
Author: zjxitcc 时间: 2022-6-19 23:07
本帖最后由 zjxitcc 于 2022-6-19 23:08 编辑

gauss98 发表于 2022-6-17 16:37
我原来也是这么认为的，
这次就算个模型反应，打算用高精度方法来验证下，才发现精度远远的没那么高。
...

一个都没成，正常，因为常规计算（高斯找过渡态、做频率分析、IRC这类静态计算）没考虑的因素太多了，这话不是指计算级别不够，即使你是Full CI/CBS搭配隐式溶剂模型，把主、副反应机理算得很全，也未必能解释正确的实验结果。
Singleton这篇烯烃与硼烷加成的经典文章Dynamics and the Failure of Transition State Theory in Alkene Hydroboration展示了（至少在这类反应里）即使你把反应物和过渡态的电子能量和自由能算得很准也不够，要考虑显式溶剂和动态学效应。因此理论计算设计新反应，一定要找那些条件简单的（反应物、催化剂结构尽可能简单、参与反应分子数少，反应条件也简单的），这样成功率才能高一些。
如果你做出来实验与计算不匹配，反过来在计算里把一个个没考虑的因素考虑进去（不是指换泛函这种考虑），不断提升计算复杂度，最后能解释实验，也是可以的。

作者
Author: sobereva 时间: 2022-6-20 10:21

linhai093 发表于 2022-6-19 15:24
请问基础班在方面会有多少涉及和讲解？

几乎所有问题在培训里都会讲到

作者
Author: qinjiu 时间: 2022-7-28 19:05
本帖最后由 qinjiu 于 2022-7-28 19:28 编辑

Computational chemistry is rapidly emerging as a subfield of theoretical chemistry, where the primary focus is on solving chemically related problems by calculations. For the newcomer to the field, there are three main problems:
(1) Deciphering the code. The language of computational chemistry is littered with acronyms, what do these abbreviations stand for in terms of underlying assumptions and approximations?
(2) Technical problems. How does one actually run the program and what to look for in the output?
(3) Quality assessment. How good is the number that has been calculated?
Point (1) is part of every new field: there is not much to do about it. If you want to live in another country, you have to learn the language. If you want to use computational chemistry methods, you need to learn the acronyms. I have tried in the present book to include a good fraction of the most commonly used abbreviations and standard procedures.
Point (2) is both hardware and software specific. It is not well suited for a textbook, as the information rapidly becomes out of date. The average lifetime of computer hardware is a few years, the time between new versions of software is even less. Problems of type (2) need to be solved “on location”. I have made one exception, however, and have included a short discussion of how to make Z-matrices. A Z-matrix is a convenient way of specifying a molecular geometry in terms of internal coordinates, and it is used by many electronic structure programs. Furthermore, geometry optimizations are often performed in Z-matrix variables, and since optimizations in a good set of internal coordinates are significantly faster than in Cartesian coordinates, it is important to have a reasonable understanding of Z-matrix construction.
As computer programs evolve they become easier to use. Modern programs often communicate with the user in terms of a graphical interface, and many methods have become essential “black box” procedures: if you can draw the molecule, you can also do the calculation. This effectively means that you no longer have to be a highly trained theoretician to run even quite sophisticated calculations.
The ease with which calculations can be performed means that point (3) has become the central theme in computational chemistry. It is quite easy to run a series of calculations that produce results that are absolutely meaningless. The program will not tell you whether the chosen method is valid for the problem you are studying. Quality assessment is thus an absolute requirement. This, however, requires much more experience and insight than just running the program. A basic understanding of the theory behind the method is needed, and a knowledge of the performance of the method for other systems. If you are breaking new ground, where there is no previous experience, you need a way of calibrating the results.The lack of quality assessment is probably one of the reasons why computational chemistry has (had) a somewhat bleak reputation. “If five different computational methods give five widely different results, what has computational chemistry contributed? You just pick the number closest to experiments and claim that you can reproduce experimental data accurately.” One commonly sees statements of the type “The theoretical results for property X are in disagreement. Calculation at the CCSD(T)/6-31G(d,p) level predicts that…, while the MINDO/3 method gives opposing results. There is thus no clear consent from theory. ” This is clearly a lack of understanding of the quality of the calculations. If the results disagree, there is a very high probability that the CCSD(T) results are basically correct, and the MINDO/3 results are wrong. If you want to make predictions, and not merely reproduce known results, you need to be able to judge the quality of your results. This is by far the most difficult task in computational chemistry. I hope the present book will give some idea of the limitations of different methods.
Computers don’t solve problems, people do. Computers just generate numbers. Although computational chemistry has evolved to the stage where it often can be competitive with experimental methods for generating a value for a given property of a given molecule, the number of possible molecules (there are an estimated 10200 molecules with a molecular weight less than 850) and their associated properties is so huge that only a very tiny fraction will ever be amenable to calculations (or experiments). Furthermore, with the constant increase in computational power, a calculation that barely can be done today will be possible on medium-sized machines in 5–10 years. Prediction of properties with methods that do not provide converged results (with respect to theoretical level) will typically only have a lifetime of a few years before being surpassed by more accurate calculations.

from 《Introduction to Computational Chemistry》

作者
Author: Q-Chembio-llg 时间: 2022-7-28 20:02
别的不了解，但是对映选择性这种俩chirality-determining TS的能量差不到2 kcal/mol的东西，实验上如果没做出来，单独理论上的预测扔出去应该很难让别人信服，个人意见。

作者
Author: EdwardLimit 时间: 2022-7-29 15:43
哎，怎么说了
计算结果解释实验现象 ---》调参，总能给个说法
计算结果指导实验---》难，相当难
前者的问题在于，总是在现有的理论范式下构建模型，其所能得到得结果也因此受到局限，或者结果基本上都是能够根据理论得到定性规律的，无非多了一个所谓精确的值。后者的问题在于达到目的的过程总是在会不可避免的少考虑一些影响因素，毕竟靠人的先验来挖掘影响因素局限性还是太大，这种局限性会导致模型存在缺陷，而接近不了实际。当然，现在人工智能这几年似乎快渗透到计算化学的方方面面了，我也看了两百来篇，虽然大多数都是水文章，但也有很多确实推进了行业发展，总之，还是在看看吧，我觉得后者未来肯定能行。

作者
Author: 牧生 时间: 2022-7-29 16:09
本帖最后由牧生于 2022-7-29 16:13 编辑

EdwardLimit 发表于 2022-7-29 15:43
哎，怎么说了
计算结果解释实验现象 ---》调参，总能给个说法
计算结果指导实验---》难，相当难

虽然大多数都是水文章，但也有很多确实推进了行业发展

复制代码

赞同。

虽然我觉得水文章不好，但是实实在在来说，我个人还是很乐意看到这种现象的。在当前社会下，作者们也是要为五斗米折腰的。毕竟，很多单位的发展，人们需要生活下去，都需要这些硬指标来评价的，先解决有无，其他的后面再说。金字塔的底部永远是大多数。

我们行业的某个大牛给很多项目签过字，我以前对此有点不屑。那些项目，连我都能看出漏洞，还都评价是国际先进，国内领先。后来某一天我就突然醒悟了，行业要发展，单位要产出经济效益，各大单位还有很多很多人就等着这个签字活下去。

欢迎光临计算化学公社 (http://bbs.keinsci.com/)