=E2=80=94

Ultimately, you have = to like what you=E2=80=99re doing enough to get up and do it = again tomorrow...

= --Apple-Mail=_E34B4DEC-A566-4F78-A512-5B7F8D54348F-- From owner-chemistry@ccl.net Fri Jan 6 13:14:01 2017 From: "Rajarshi Guha rajarshi.guha[-]gmail.com" To: CCL Subject: CCL: DFT discovers it's recapitulating QSAR Message-Id: <-52593-170106131237-30383-F9O6ilOh9MuXbV3Zba17Hg]*[server.ccl.net> X-Original-From: Rajarshi Guha Content-Type: multipart/alternative; boundary=001a11473c642351b5054570f6cf Date: Fri, 6 Jan 2017 13:12:31 -0500 MIME-Version: 1.0 Sent to CCL by: Rajarshi Guha [rajarshi.guha() gmail.com] --001a11473c642351b5054570f6cf Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: quoted-printable Nicely put! On Fri, Jan 6, 2017 at 11:59 AM, Joe Leonard jleonard42%%gmail.com < owner-chemistry]^[ccl.net> wrote: > Odd blog to find this article, but it echoes concerns raised as ACS SD > when talks described functionals built on statistical models but did NOT > mention test/training/validation sets or other safety measures that the > QSAR literature has been dealing with for decades=E2=80=A6 > > FYI > > http://arstechnica.com/science/2017/01/in-chemistry- > computational-models-may-be-getting-worse/ > > =E2=80=94 > Ultimately, you have to like what you=E2=80=99re doing enough to get up a= nd do it > again tomorrow... > > --=20 Rajarshi Guha | http://blog.rguha.net NIH Center for Advancing Translational Science --001a11473c642351b5054570f6cf Content-Type: text/html; charset=UTF-8 Content-Transfer-Encoding: quoted-printable

Nicely put!

On Fri, Jan 6, 2017 at 11:59 AM, Joe Leonard jleonard42%%<= a href=3D"http://gmail.com">gmail.com <owner-chemistry]^[ccl.net> wrote:

Odd blog to find this article, but it echoes concerns raise= d as ACS SD when talks described functionals built on statistical models bu= t did NOT mention test/training/validation sets or other safety measures th= at the QSAR literature has been dealing with for decades=E2=80=A6

<= /div>
FYI

http://arstechnica.com/science/2017/01/in-chemistry-computational-models-may-be-getting-worse/

=E2=80=94
=
<= div style=3D"text-align:start;text-indent:0px;word-wrap:break-word">
Ultimately, you have to like wh= at you=E2=80=99re=C2=A0doing enough to get up and do it again=C2=A0tomorrow= ...

-- =

Rajar= shi Guha | http://blog.= rguha.net
NIH Center for Advancing Translational Science

--001a11473c642351b5054570f6cf-- From owner-chemistry@ccl.net Fri Jan 6 18:08:00 2017 From: "Susi Lehtola susi.lehtola_._alumni.helsinki.fi" To: CCL Subject: CCL:G: DFT discovers it's recapitulating QSAR Message-Id: <-52594-170106152419-23259-ht0dyurEnJR2VyPJHQunJw^_^server.ccl.net> X-Original-From: Susi Lehtola Content-Transfer-Encoding: 8bit Content-Type: text/plain; charset=utf-8; format=flowed Date: Fri, 6 Jan 2017 12:24:08 -0800 MIME-Version: 1.0 Sent to CCL by: Susi Lehtola [susi.lehtola .. alumni.helsinki.fi] On 01/06/2017 08:59 AM, Joe Leonard jleonard42%%gmail.com wrote: > Odd blog to find this article, but it echoes concerns raised as ACS SD > when talks described functionals built on statistical models but did NOT > mention test/training/validation sets or other safety measures that the > QSAR literature has been dealing with for decades… > > FYI > > http://arstechnica.com/science/2017/01/in-chemistry-computational-models-may-be-getting-worse/ This has also been featured on phys.org http://www.phys.org/news/2017-01-electron-density-discovery-disturbing-trends.html and C&EN http://cen.acs.org/articles/95/i2/Density-functional-theory-heads-wrong.html However, I wouldn't read too much into this. The list of functionals they studied is very small, due to limitations in the programs they used. Most of the more recent functionals in that study were the ones by Truhlar and collaborators, which also had the poorest performance (see Table 2 in the paper). This is not surprising, since the newer Minnesota functionals are known to be somewhat pathological with respect to integration grids and basis sets [JCTC 9, 4453 (2013); JCTC 12, 4303 (2016)]. The only other newer functionals in the study were ones by Perdew and coworkers. Many other people have also been active in functional development in recent years, and those functionals have not been included in that study. I doubt their conclusion would hold if one would actually run a decent sample of modern functionals. Moreover, the paper is somewhat funny, as they only looked at the isoelectronic series of He, Be, and Ne. Also the argument against the use of anions is somewhat funny: "Anions were not included, because semilocal functionals can only bind a fraction of an electron to any neutral atom. Thus, the errors in the densities of anions would reflect not only the functional but also the basis set, and would not fairly measure the relative performances of functionals." If an anion is bound, then it can be properly described in a Gaussian (or Slater type orbital) basis set, and coupled-cluster theory will give you a good density. The problem of pure DFT in binding anions is in the functional itself, and studying anionic densities *is* a sensible measure of the goodness of the functional. The only thing is that the basis set calculation will give you an *underestimate* of the error in the cases where DFT does not bind, because the L^2 basis set will still result in artificial binding. However, within a fixed basis set, you will still be able to rank functionals with respect to the goodness of the density, because the error is constant in each calculation. -- ----------------------------------------------------------------------- Mr. Susi Lehtola, PhD Chemist Postdoctoral Fellow susi.lehtola]![alumni.helsinki.fi Lawrence Berkeley National Laboratory http://www.helsinki.fi/~jzlehtol USA -----------------------------------------------------------------------