Database Reference
In-Depth Information
structure parameters:
Bulk modulus
=−
1
.
00096
EN
−
0
.
35682
x
−
0
.
77228
BL
A
−
N
−
0
.
83367
BL
B
−
N
03296
Q
tet
+
18484
Q
oct
−
13503
Q
N
+
0
.
0
.
0
.
where
EN
is the weighted electronegativity difference,
x
is the internal anion
parameter,
BL
A
−
N
is the A-N bond length,
BL
B
−
N
is the B-N bond length,
Q
tet
is the Mulliken effective charge for tetrahedral site ion,
Q
oct
is the Mul-
liken effective charge for octahedral site ion, and
Q
N
is the Mulliken effective
charge for N ion.
By systematically exploring the number and type of variables needed, Suh
and Rajan found very strong agreement in being to able to predict properties
consistent with ab-initio calculations based strictly on a data-driven analysis.
Based on our QSAR formulation, the role of the effective charge (
Q
∗
) in en-
hancing modulus is particularly notable. This is consistent with theoretical
studies, which show that it is the effective charge parameter that helps to
define the degree of charge transfer and the level of covalency associated with
the specific site occupancy of a given species. Ab-initio calculations of this
effective charge can then be used as a major screening parameter in identi-
fying promising crystal chemistries for promoting the modulus. Hence, using
PLS to develop a QSAR formulation, combined with an interpretation of the
physics governing these materials, can indeed be valuable. Our predictions fit
well with systems of similar electronic structure and allow us to clearly iden-
tify outliers based on these quantum mechanical calculations. Based on these
predictions, we can now seriously and effectively accelerate materials design
by focusing on promising candidate chemistries. Those selected can then be
subjected to further analysis via experimentation and computational methods
to validate crystal-structure-level properties. The data generated by these se-
lective experiments and computations also serve to refine the next generation
of “training” data for another iterative round of data mining, which permits
a further refinement of high-throughput predictions.
8.5.3 Data Mining for Descriptor Development:
Enhancing Databases for Predictions
Standard materials databases containing experimental and theoretical data
about properties and structures can be enhanced by the addition of chemistry-
structure-property descriptors for specific applications such as materials
design. While these descriptors are usually statistically derived, they must in-
corporate the physics of the problem. In this case study, the starting database
is the zeolite framework database of the Structure Commission of the Inter-
national Zeolite Association (http://www.iza-sc.ethz.ch/IZA-SC/). However
this database is not useful to predict the mesopore sizes of the framework,
suggesting the need for additional descriptors or types of data. Using prin-
cipal component analysis and PLS, we have developed secondary descriptors
