计算化学公社

标题: INTERVIEW: STEFAN GRIMME [打印本页]

作者
Author: sobereva 时间: 2015-9-22 22:59
标题: INTERVIEW: STEFAN GRIMME
2014年Computational Organic Chemistry 2ed书中的一段，感觉这段讨论比较有料，有意思，说了一些实在话，值得看看。Grimme比较致力于实实在在真真正正地解决较大体系的各种计算问题，不像很多搞理论的人老是搞一些玄虚没实际意义（而且往往也称不上有多大理论价值）的东西。

Interviewed August 30, 2012

"I am looking for solutions that work in practice,” says Stefan Grimme, University Professor at Universitat Bonn. “I want to have methods that are as black box as possible, as robust as possible, that can be used for large systems. Large chemical systems are always my focus. I am not always looking for high accuracy methods; a little bit of empiricism is ok. But this kind of empiricism must be physically well-founded.” These principles have guided Grimme’s pursuit of new quantum chemical methods.

Grimme’s three major contributions to quantum chemistry all derive from a desire to deal with large systems, where standard quantum mechanical approaches, be they wavefunction or density functional methods, are computationally intractable. His work on dispersion corrections came about through work on surface chemistry, such as for adsorption, where dispersion is inherent. Grimme was not the first to suggest a dispersion correction; early attempts date back to Ahlrich’s corrected HF in the 1970s. After a few years of working with various approaches, Grimme found success with the “-D” correction, first published in 2004. He has twice since developed improvements, the latest being the “-D3” correction. He indicates that “my name is connected with it because my corrections are the best around, but others had the original idea. I just developed them further and improved them.”

In contrast, the SCS-MP2 idea came to Grimme suddenly, and he had it all worked out in about 30 min. “I thought ‘MP2 is a nice method; it can be used for large systems (which is again my point), but is too inaccurate for some reactions.’ I saw that there are physically these different spin components of the correlation, and no one had ever looked at their relations to questions of accuracy.” The computer program he was using was able to print out the values of these components, so he could immediately test out his hypothesis. “I want to put my ideas into practice very quickly! I don’t want to spend my time on tiny improvements. I normally will invest a week or a month on a new idea, but not a year,” he explains. Within a couple of days, he knew the method was “a great idea.” He observed huge improvement in a few simple test cases. “Writing the paper took many months,” he notes “but from the idea to the first results only one day!”

Grimme’s third major methodological development is the notion of double hybrid density functionals. This development came when he recognized that the principle of perturbation theory, expressed through MP2, could be applied to DFT. He characterizes this realization as “not a big step.”

While he recognizes that SCS-MP2 is perhaps his most important individual contribution, he asserts that the dispersion correction work is much more significant. “There are many problems that cannot be solved without dispersion. We can live without double hybrids or SCS-MP2, but proper accounting for dispersion is critical within DFT.”

Grimme expressed some surprise that adoption of his dispersion corrections took so long. Today he is seeing exponential growth in the use of dispersion corrections. “This is really quite normal,” he realizes. “People need time to recognize new ideas. I was just too naive.”

Getting quantum chemical methods into some of the standard programs is very important to any methods developer. “It is necessary to have your methods in some code that is noncommercial, like Orca by Frank Neese. Everybody can use it. But having it in Gaussian is fantastic. When you have something in Gaussian connected to your name, you are totally on the safe side.”

As we discussed the general state of large-molecule computational methodologies, Grimme noted that he, along with many others, have repeatedly pointed out the problems with the B3LYP functional. Nonetheless, B3LYP remains a very widely used functional, a phenomenon that is often discussed in his group. He claims “I would not in general say that people should not use B3LYP. Rather they should understand it. When you properly correct B3LYP for its over-repulsive behavior, dispersion correct it, then it’s not bad. It’s not the best functional. It has the nice property that it is the best tested functional. It works in very many cases. This is a very important property, it’s robustness in a sense. But this only holds when you properly dispersion-correct it.”

Grimme concurs that CCSD(T) remains the gold standard method. “It is our go-to method. If you have more-or-less a single reference situation, as in most chemical reactions, then CCSD(T) is OK. In cases where this approximation breaks down, we need then to choose on a case-by-case basis. The situation is no longer black box. There is much room for development for this situation – to make these more difficult cases black box.”

The Grimme group has evolved a standard protocol for addressing new problems. “This is important when you have so many different projects that you don’t get confused, that the students don’t get confused,” he notes. “We use TPSS with dispersion correction for geometry optimization – its fast in Orca. For the energies, we employ a hybrid or double hybrid functional. To correct for solvation we use COSMO-RS. Most problematic is the computation of the costly frequencies, where we often run into numerical problems for large systems.”

At the beginning of his career, Grimme was in close contact with experimentalists. This began with his examination of electronic spectroscopy of cyclophanes as he sought ways to compute its CD spectrum. Grimme’s work on large systems was largely inspired by the supramolecular chemist Fritz Vögtle and the organic chemist Gerhard Erker. “Usually experimentalists come to me with their problems. I could have applied standard methods but they inspired me to think of new ways to answer their problems accurately and efficiently.”

Occasionally, he has approached an experimentalist with a computationally inspired problem. In working with Bader’s Atoms-In-Molecules theory, Grimme was interested in whether bond paths really correlate with bonds. Grimme proposed to Erker to synthesize 4,5-dideuterated phenanthrene “and we got spectroscopists to measure this!”

When asked about problems and bottlenecks facing computational chemist today, Grimme commented “In my experience with larger systems, the problem is conformational space: the degrees of freedom, the sheer number of isomers and conformers. What quantum chemists often fail to recognize is that it’s not only the electrons, it’s the nuclei as well! We can narrow down the error in the electronic energy to 1 kcal mol-1, but the error in the ZPE can easily reach 3 kcal mol-1, which is not uncommon for a big system. With all those degrees of freedom, the errors can mount up. There is room for improvement here in dealing with large conformational space. It’s not just computing power, there is factorial growth. And there is also the entropy term. The harmonic oscillator approximation is simply too crude. But moving beyond the harmonic approximation is just impossible now. Its unavoidable today, but we know that this approach is seriously deficient. We are starting to think about this more. If you want to do accurate thermochemistry for a big system you are faced with this problem.”

Finally, he noted a real dearth of solid computational methods to deal with the electronic structure of transition metal compounds. “For main groups, we have appropriate methods,” he says. “But for metallic clusters, this is really hard. Computations of metal surfaces at high accuracy are impossible today. This is a case where the physicists are leading the chemists.” These problems will no doubt keep Grimme inspired for years to come!

作者
Author: wxy 时间: 2015-9-23 00:12
Gaussian09带有B2PLYP和mPW2PLYP二个双杂化泛函，不知道此二个与B3LYP相比计算耗时会增加多少？谢了！

作者
Author: sobereva 时间: 2015-9-23 14:26

wxy 发表于 2015-9-23 00:12
Gaussian09带有B2PLYP和mPW2PLYP二个双杂化泛函，不知道此二个与B3LYP相比计算耗时会增加多少？谢了！

双杂化泛函耗时和MP2基本一样，B3LYP耗时和HF基本一样，可想而知，耗时会增加N倍或一个数量级以上，使体系而定，越大的体系增加的程度越大。

作者
Author: wxy 时间: 2015-9-23 15:22
刚才试了下，B2PLYP的耗时比mPW2PLYP多50%，约是B3LYP的30倍。

作者
Author: yflchx 时间: 2015-9-23 16:22

wxy 发表于 2015-9-23 15:22
刚才试了下，B2PLYP的耗时比mPW2PLYP多50%，约是B3LYP的30倍。

实用性要得益于RI。

作者
Author: yflchx 时间: 2015-9-23 16:29
虽然经常用DFT-D，但是不太欣赏它。物理上说，DFT-D3(BJ)是最好的。

一堆参数，表现好是应该的，什么时候表现不好才是有意思的问题。

我猜肯定在某种情况下，DFT-D会过高地计算Dispersion部分。

作者
Author: scf 时间: 2015-9-25 03:34

Finally, he noted a real dearth of solid computational methods to deal with the electronic structure of transition metal compounds. “For main groups, we have appropriate methods,” he says. “But for metallic clusters, this is really hard. Computations of metal surfaces at high accuracy are impossible today. This is a case where the physicists are leading the chemists.” These problems will no doubt keep Grimme inspired for years to come!

在dmrg的现状下, 强关联系统的坑还能撑多久？新一代的研究者是填坑还是开坑呢？

欢迎光临计算化学公社 (http://bbs.keinsci.com/)