aCenter

for Computationally Assisted Science and Technology, North Dakota State University, Fargo, ND, United States
bDepartment of Computer Science, North Dakota State University, Fargo, ND, United States
cDepartment of Chemistry, North Dakota State University, Fargo, ND, United States

CCAST

*E-mail: bakhtiyor.rasulev@ndsu.edu

Introduction

Table 1. The list of investigated polymers, splitting,

experimental and predicted values of glass
transition temperature (Tg).

The reduction of the number of experimental trials become more and more popular in the development of new polymers. The new polymers with most wanted properties can
be predicted by computer algorithms. In this sense, the trend is encouraged by the economic advantages that experimental research can get from it in terms of time and money
saving. The glass transition temperature (Tg) is one of the properties of polymers which indicates the temperature below which the material becomes rigid and brittle due to the loss
of molecular mobility [1].
In this context, the relationships between polymers structural characteristics and Tg can be established by quantitative structure-property relationship (QSPR). Numerous
attempts have been made to predict Tg by various approaches.
In this work, we propose a seven-descriptor QSPR model for the prediction of glass transition temperature with a diverse set of 80 polymers. The aim of this work was to use
the proposed model which can be used for predicting glass transition temperature of new design polymers from their molecular structures.

ID

Figure 1. Schematic diagram of a solid supported

phospholipid bilayer. The membrane is separated
from the substrate by a10-20A thick .

Materials and Methods

Experimental dataset
A list of 80 polymers with their experimental Tg were collected from diverse compilations [13], [14] and [15]. 80% of the 80 polymers was
used as training set and the rest 20% is used as test set.

Computational details

Results

Descriptors Generation. The molecular structure of all polymers were drawn using ChemDraw 14 [16]. The final geometries of the

Statistical Data for Models 1-10

minimum energy conformation were obtained using Avogado [17]. Then the molecular descriptors of each polymer were calculated using
DRAGON software on the minimal energy conformation.

0.9
0.8
0.7
0.6

Model Development and Validation. The models were developed using QSARINS 2.2 software

0.5
0.4

developed by group of Prof. P. Gramatica. The dataset has been splitted to training and test sets
to get robust and externally validated models. The ratio of training and test set compounds was 80:20.

0.3
0.2
0.1

Virtual Library. A virtual library of the novel 54 polymers has been designed and Tg values
will be predicted by using the developed QSPR model.

Results

Figure 2. Schematic diagram of transition of polymer from

glassy state to a rubbery state and melting point. Polymer chains
move from frozen state to a flexible state. Copyright Macromolecular Materials and Engineering
2014, 299 (5), 524539.

4
R2 train

Q2 train

7
R2 ext

Figure 1. The plot of R2, Q2 and R2ext

values for the models with 1 variable to 10
variables in the equation.

7-variable Model
The best predictive ability showed a model with 7 descriptors involved in MLR equation:

Prospective work

Log(Tg) = 0.25(0.08) AVS_B(e) 0.46(0.29) SM6_X + 0.01(0.002) SpAbs_Dz(e)

0.04(0.06) MATS7e 0.03(0.01) L1p 0.05(0.02) F-083 + 0.22(0.07) SsssSiH +
2.24(0.35)

10

Prediction Tg for new polymer systems.

NAME
1hexafluoropropylene_oxide.mol
2tetrafluoroethylene.mol
3dimenthyl_siloxane.mol
4trifluoroethyl_acrylate.mol
5vinylidene_fluoride.mol
6vinyl_isobutyl_ether.mol
7ethoxyethylene.mol
8vinyl_ethyl_ether.mol
9vinyl_n_butyl_ether.mol
10oxyethylene.mol
11propylene_oxide.mol
123_butoxylpropylene_oxide.mol
13vinyl_n_pentyl_ether.mol
143_hexoxylpropylene_oxide.mol
15vinyl_n_octyl_ether.mol
16vinyl_2_ethylhexyl_ether.mol
17vinyl_n_decyl_ether.mol
18tert_butyl_methacrylate.mol
194_methyl_1_pentene.mol
20n_butyl_acrylate.mol
21vinyl_propionate.mol
22vinyl_methyl_ether.mol
23vinyl_acetate.mol
24ethyl_acrylate.mol
25isopropyl_methacrylate.mol
26propylene.mol
27mol27.mol
28dodecyl_methacrylate.mol
29ethylene_succinate.mol
30vinyl_formate.mol
31ethylene.mol
32isobutyl_methacrylate.mol
33methyl_acrylate.mol
34oxymethylene.mol
35n_hexyl_methacrylate.mol
36n_butyl_methacrylate.mol
37n_propyl_methacrylate.mol
38ethyl_methacrylate.mol
39vinyl_butyral.mol
40methyl_methacrylate.mol
41sec_butyl_alpha_chloroacrylate.mol
421_2_butadiene.mol
43mol43.mol
44ethyl_alpha_chloroacrylate.mol
452_methylcyclohexyl_methacrylate.mol
46cyclohexyl_methacrylate.mol
472_hydroxyethyl_methacrylate.mol
48vinyl_chloroacetate.mol
492_chloroethyl_methacrylate.mol
50acrylonitrile.mol
51isoprene.mol
52acrylic_acid.mol
53N_vinylpyrrolidone.mol
54oxy_methylphenylsilylene.mol
55cyclohexyl_alpha_chloroacrylate.mol
56vinyl_chloride.mol
57chloroprene.mol
58benzyl_methacrylate.mol
59phenyl_methacrylate.mol
60ethylene_terephthalate.mol
61oxy_2_6_dimethyl_1_4_phenylene.mol
62vinyl_benzoate.mol
63o_methoxy_ styrene.mol
64styrene.mol
65propylene_sulfide.mol
66N_benzyl_methacrylamide.mol
67p_methoxy_styrene.mol
68vinylidene_chloride.mol
69o_chloro_styrene.mol
702_6_dichlorostyrene.mol
71chloro_p_xylylene.mol
72oxy_2_6_diphenyl_1_4_phenylene.mol
73p_xylylene.mol
74alpha_vinyle_naphthalene.mol
75N_vinyl_carbazole.mol
76pentafluoropropyl_acrylate.mol
77chlorotrifluoroethylene.mol
782_methoxyethyl_acrylate.mol
792_ethoxyethyl_acrylate.mol
802_chlorostyrene.mol

Exp.
Pred. by Pred.Mod
Status
endpoint model eq. .Eq.Res.
Training
2.3483
2.3409 -0.0074
Training
2.5888
2.5678
-0.021
Training
2.1703
2.1964
0.0261
Training
2.42
2.4538
0.0338
Prediction
2.3674
2.4873
0.12
Training
2.3997
2.3834 -0.0163
Training
2.3711
2.387
0.016
Training
2.4048
2.3768
-0.028
Training
2.3444
2.3204
-0.024
Prediction
2.3139
2.3854
0.0715
Training
2.2967
2.4451
0.1484
Training
2.2878
2.3424
0.0546
Training
2.316
2.3017 -0.0143
Training
2.2742
2.3206
0.0464
Prediction
2.2878
2.2673 -0.0205
Training
2.316
2.4032
0.0872
Training
2.2945
2.2522 -0.0423
Training
2.5922
2.5467 -0.0455
Training
2.48
2.4398 -0.0402
Prediction
2.3404
2.4379
0.0975
Training
2.4518
2.4011 -0.0507
Training
2.415
2.3902 -0.0247
Training
2.4786
2.4417 -0.0369
Training
2.3997
2.4366
0.0369
Prediction
2.5145
2.5098 -0.0048
Training
2.3674
2.3571 -0.0102
Training
2.4031
2.4466
0.0435
Training
2.3385
2.3382 -0.0003
Training
2.4346
2.4617
0.0271
Prediction
2.4829
2.455 -0.0279
Training
2.1703
2.2579
0.0876
Training
2.5065
2.4891 -0.0174
Training
2.4487
2.481
0.0323
Training
2.3385
2.3033 -0.0352
Prediction
2.4281
2.4187 -0.0094
Training
2.4669
2.452 -0.0148
Training
2.4886
2.4347 -0.0538
Training
2.5105
2.4396 -0.0709
Training
2.5105
2.5221
0.0115
Prediction
2.5775
2.4941 -0.0834
Training
2.5403
2.5364
-0.004
Training
2.4298
2.4654
0.0357
Training
2.5539
2.3936 -0.1603
Training
2.29
2.4737
0.1837
Prediction
2.4669
2.6068
0.1399
Training
2.5763
2.5419 -0.0344
Training
2.5551
2.4624 -0.0927
Training
2.4829
2.4654 -0.0175
Training
2.5623
2.4795 -0.0828
Prediction
2.5775
2.5023 -0.0752
Training
2.3075
2.473
0.1655
Training
2.5786
2.5185 -0.0601
Training
2.6212
2.6018 -0.0193
Training
2.2718
2.2524 -0.0195
Prediction
2.5877
2.6635
0.0758
Training
2.5416
2.3779 -0.1636
Training
2.3522
2.4749
0.1228
Training
2.5145
2.5589
0.0443
Training
2.5944
2.5983
0.0039
Prediction
2.5378
2.6176
0.0798
Training
2.683
2.7476
0.0646
Training
2.5366
2.5311 -0.0054
Training
2.6117
2.5895 -0.0222
Training
2.5717
2.5424 -0.0293
Prediction
2.3541
2.4334
0.0793
Training
2.6946
2.5703 -0.1243
Training
2.5587
2.5713
0.0126
Training
2.4082
2.4564
0.0482
Training
2.5933
2.5856 -0.0077
Prediction
2.6435
2.6114
-0.032
Training
2.5478
2.5848
0.0371
Training
2.6928
2.6287 -0.0641
Training
2.5224
2.5513
0.0289
Training
2.6355
2.637
0.0015
Prediction
2.6263
2.716
0.0897
Training
2.5353
2.5138 -0.0215
Training
2.5717
2.5477
-0.024
Training
2.3483
2.441
0.0927
Training
2.3483
2.3625
0.0142
Prediction
2.5933
2.5898 -0.0035

(n=80, r2train=0.73, q2loo=0.68, r2test=0.67, RMSEtr=0.064, F=21.998)

Conclusions

Descriptors
AVS_B(e) = 2D-based descriptor - average vertex sum from Burden matrix weighted by Sanderson electronegativity
SM6_X = 2D-based descriptor - spectral moment of order 6 from chi matrix
SpAbs_Dz(e) = 2D-based descriptor - graph energy from Barysz matrix weighted by Sanderson electronegativity
MATS7e = 2D-based descriptor - Moran autocorrelation of lag 7 weighted by Sanderson electronegativity
L1p = WHIM directional descriptors - 1st component size directional WHIM index / weighted by polarizability
F-083 = F attached to C3(sp3)
SsssSiH = Atom type E-states - Sum of sssSiH E-states

It worth to note that most of the descriptors involved in the model are
electronegativity based descriptors (AVS_B(e), SpAbs_Dz(e), MATS7e or polarizability
(L1p) which supports the idea that glass transition temperature depends on interactions
between polymer chains that govern by electronegative groups.
Figure 4. The plots of correlation between observed and predicted
values of Tg (a), the Williams plot (applicability domain) of the model (b),
the Y-scrambling plot (c).

Figure 3. An example of importance of polymers with high glass

transition temperature feature in electronics. (a), OFET structure,
conducting channel is formed at the interface between polymer
and organic semiconductor. (b), electron transport in a prefect
conducting channel below Tg, charge transport is located in a
well-defined conducting channel. (c), electron hindered in a
disrupted conducting channel, above Tg, the chain segments shift
from a frozen in or cessation state (the glassy state, T < Tg) to a
cooperative motion state (the rubbery state, T Tg). Copyright Scientific Reports, 2014, 4, 6387.

The glass transition temperature of 80 polymers was

investigated by computational methods, applying QSAR
methodology. The 7-descriptor model is developed, showing a
good performance, internally and externally. It worth to note that
most of the descriptors involved in the model are electronegativity
based descriptors (AVS_B(e), SpAbs_Dz(e), MATS7e or
polarizability (L1p) which supports the idea that glass transition
temperature depends on interactions between polymer chains that
govern by electronegative groups.
The next step is to use generated virtual library to predict Tg
temperatures for new polymers.

Acknowledgements
The authors gratefully acknowledge support from the North Dakota State University Center for Computationally
Assisted Science and Technology and the U.S. Department of Energy through Grant No. DE-SC0001717.