Computationalchemistry:CriticalappraisalMario BarbattiA*Midex Chair Professormario.barbatti@univ‐amu.fr

Aix Marseille Université, Institut de Chimie Radicalaire 

L10

L I G H T A N DM O L E C U L E S

Results depend on the method

L I G H T A N DM O L E C U L E S

3

Ab initio and semiempirical MRCI produce  divergent results

26

(°)

(°)

L I G H T A N DM O L E C U L E S

4• Plasser, Crespo‐Otero, Pederzoli, Pittner, Lischka, Barbatti, JCTC 10, 1395 (2014) 

Experi

mental

ADC(2)MRCIS

OM2/MRCI

TD-PBE

TD-B97

XDTD-B

3LYP

TD-PBE0

TD-CAM-B

3LYP

TD-BHLY

PTD-M

06-H

F

0

20

40

60

80

S 0 po

pula

tion

at 1

ps

(%)

C2 puckering C6 puckering H elimination Experimental682

5712

858

598

58128

2015

0

20212516

1016

TDDFT: unable to predict ps behavior

Right for different reasons

L I G H T A N DM O L E C U L E S

5

2.0

2.5

3.0

3.5

4.0

4.5

B97X-D(6-311G(df,2p))

M06-HFB3LYP PBE0 B97X-D(6-31G(d))

M06-2X CAM-B3LYP

LC-BLYP( = 0.47)

LC-BLYP( = 0.29)

Ener

gy (e

V)

LC-BLYP( = 0.2)

f > 0.2

LOC(A)LOC(D)CT DACT ADDELOC

• Sen, Crespo‐Otero, Weingart, Thiel, Barbatti, JCTC 9, 533 (2013)

D

A

L I G H T A N DM O L E C U L E S

6

• Excited‐state spectral region has high density of states• Small variation in geometry leads to a change of electronic character• No affordable method can describe all characters at the same level

L I G H T A N DM O L E C U L E S

Till where Can We Go with Single-Reference Methods?

L I G H T A N DM O L E C U L E S

8

Energy

Time

SR

ia SR

ja SR

SR

i ja a SR

i ja a SR

MR

MR

Dynamic Electron Correlation

Non‐Dynamic EC

If DEC is needed everywhere and NDEC only at the intersection, why choosing CASSCF (only DEC) for dynamics?

L I G H T A N DM O L E C U L E S

9

Up to here, dynamics may be OK.But we need better research protocolsto guarantee quality

L I G H T A N DM O L E C U L E S

10

Pyramidalization (°)

L I G H T A N DM O L E C U L E S

11• Barbatti, Crespo‐Otero, Top Curr Chem 368, 415 (2015)

Diagonal XC kernel for CT stateshas wrong behavior 

Single‐determinant KSapproximation fails; self‐

interaction errors

Linear response and ALDA approximations 

may fail

No problem, in principleFunctional‐dependent

predictions

DFT is, in principle, valid

Functional‐dependentpredictions

Energy

xy Linear response fails for

multiple excitations

L I G H T A N DM O L E C U L E S

1T 2T 3T 4T0.0

0.5

1.0 9T (expt)

Excited Population at 300 fs

12• Fazzi, Barbatti, Thiel, PCCP, accepted (2015)

TD‐B97XD/6‐31G*

L I G H T A N DM O L E C U L E S

The best strategy

L I G H T A N DM O L E C U L E S

14

Computational chemistry of excited states:

• Powerful tool to understand mechanisms and interpret experiments• Due to very complex electronic densities, they demand high expertise• Used like black‐box  they may lead to qualitatively wrong results

L I G H T A N DM O L E C U L E S

15

Reaction pathways:

• They can be computed with very high computational levels• However, they are limited to few biased coordinates based on “educated guess”

L I G H T A N DM O L E C U L E S

16

Semiclassical Dynamics:

• All degrees of freedom can be explored• It reveals features that we would otherwise miss • However, the computational level must be strongly downgraded

L I G H T A N DM O L E C U L E S

17

The best we can do is to combine them, reaction paths and dynamics

• Start with reaction paths comparing low/high levels and different methods• Do low level‐dynamics to let the important degrees revel themselves• Move to high‐level reaction paths to check the results• If things are not consistent, go back, improve dynamics level• Repeat the steps till self‐consistence is achieved  

L I G H T A N DM O L E C U L E S

18

The Five Commandments

I. Thou shalt use dynamics only for what is really necessary

II. Test thy inputs and outputs before submitting thy real production jobs

III. Thou shalt NOT let computers running alone without checking what is going on

IV. Thou shalt NOT start dynamics without a very careful investigation of the theoretical level

V. Thou art a limited and biased being, therefore, thou shalt NOT rely on thy eyes to perform analysis

L I G H T A N DM O L E C U L E S

19

• Use automatic approach for data input and analysis.

• Test the results with methods from different families (DFT x CC, for example).

• Improve statistics and compute statistical errors.

• Limit analysis to significant results with the error bar.

• Don’t blindly rely on previous works. Reanalyze their data.

• Be rigorous and honest. Report negative results and failures.