Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
DOI: 10.1002/chem.2008(……))
Abstract: Few solutions that aim at improvements, and the resulting toward a total automation of such a
identifying crystalline materials from remarkable properties of our routine procedure, eliminating
the analysis of powder X-ray methodology. As the use of high- laborious and time-consuming controls,
diffraction (XRD) data have been throughput techniques for the discovery preliminary treatments and settings.
reported so far. A careful inspection of or the synthesis space enlargement of Consequently, the proposed solution is
the latter and the corresponding new microporous crystalline structures of great interest and appears to be very
highlight of specific failures when used makes the trustworthiness of search- promising considering the numerous
for the discrimination of match methods a critical point to be potential applications of X-ray
crystallographic phases of zeolites assessed, a meticulous evaluation of the diffraction in material science, but also
among mixtures, has allowed the reliability and the robustness is the possible extent to several other
creation of the recently proposed provided and supported by both characterization techniques.
strategy called Adaptable Time empirical comparisons and
Warping (ATW). Here, the design mathematical proofs. The results Keywords: x-ray diffraction · full
process is thoroughly detailed in a step offered by our methodology, as it profile · warping distance · mixture
by step manner allowing a deep clearly outperforms the well- · high-throughput
understanding of the motivations, established solutions, open the way
1
which some corrective adaptations of the methodology are described Identification of Zeolite Crystallographic Phases
in the third section. It is shown how such modifications can be
merged together making the so-called Adaptable Time Warping a Zeolites are crystalline alumino-silicates, whose structures are
fundamentally different, powerful and innovative approach. Its 3D networks composed of channels and cavities of molecular
evaluation on numerous benchmarks allows supporting empirically dimensions. Their microporous framework structures, the wide
the mathematical proofs of ATW robustness and reliability property. range of chemical composition and surface acidity, and the
Finally, results using ATW for the same cases will show the clear possibility of tuning their electric fields are key factors that render
superiority of this methodology with respect to those previously zeolites versatile materials for an increasing number of
reported. applications.[8]
Figure 1. The automatic analysis of the 192 X-Ray diffraction data of the discovery Among them, is the use of zeolites as catalysts (for the
program of the zeolite ITQ-33 is expected to identify the different crystallographic
petrochemical industry, for pollution control and for the synthesis of
phases present in each sample, allowing the creation of the corresponding phase
diagram (8 different phases, numerous mixtures, and among them the presence of a new
special chemicals), for gas adsorption and separation, electronics
material has to be established). and medical uses.[9] The discovery of new structures or enlarging the
synthesis space, and optimization of existing ones require a
considerable experimental effort. This can be diminished by using
high throughput (HT) technology (usually parallelization and
miniaturization)[10] as also done for heterogeneous catalysts.[11] As a
recent example, this technique has allowed the synthesis of a very
unique zeolite structure that includes extra-large 18MR connected
with medium 10MR pores (ITQ-33).[12] The unusual conditions
2
needed to synthesize this new material were found by using HT When PolySnap2© is employed for the studies “ITQ-33”,
techniques to cover a wide range of possible synthesis conditions. “Zeolite β”, and “ITQ-21/30” the overall recognition rates are
That study required the generation of 192 diffractograms to follow 68.2%, 84.3%, 86.11% and respectively. Despite the relative
the formation of the different crystallographic phases and mixtures easiness of β study and ITQ-21/30, i.e. only three different phases
of phases depending on the synthesis conditions, and from those (including amorphous) are present with their corresponding
mixtures of phases the presence of a new zeolite structure had to be mixtures, 15.7% and 13.89% of errors are respectively obtained.
established, see Figure 1. To achieve this is not obvious when Among them, the majority corresponds to the fact that partially
complex mixtures of several crystalline phases exist in the same crystalline materials are not correctly identified. The same
samples and consequently, the identification of the phases is time observation can be done for “ITQ-33” study. This makes the number
consuming. At the same time when increasing the amount of of false negative, i.e. the presence of a given crystallographic phase
experiments to cover a broader synthesis range, it becomes which is not detected, relatively high, see the first line of the Tables
mandatory to develop an automatic data treatment not to slow down 1 and 2. Confusion matrix for “ITQ-21/30” study is given in
the whole process.[13] Therefore an important step in this direction is Supplementary information as Table S2. Among them, growing
the automatic classification of crystalline structures through XRD phases represent the principal source of misidentifications, and it
data, as it has been previously pointed by Takeushi et al.[6] For doing can be observed that mixtures of pure phases (without the presence
this PolySnap2©[3] has been selected as a reference because the of amorphous) are also poorly recognized.
software is representative and highly relevant considering the
current state of the art in HT identification of phases through full Table 1. Confusion matrix with the real phases in the experimental design “Beta study”
profile examination. Detailed explanations are available, while versus predicted classes obtained with PolySnap2 indicated at the top of the cell, and
versus ATW indicated at the bottom of the cell (in italic and between brackets).
merging various techniques: principle component analysis (PCA),
multi-dimensional scaling (MDS), cluster analysis,[14] and statistical
tests. As PolySnap2© offers an automatic analysis where user Real phases
interface is minimized to few options (“allows x-shift” and “check
Growing Mixture
Growing UDM
for amorphous” have been selected), this mode is chosen for
Growing Beta
Amorphous
Mixture
comparison. Three datasets of experiments have been selected:
UDM
Beta
“Zeolite β”,[10b] “ITQ-21/30”,[10a,10d] and “ITQ-33”[12].
TEA/(Si+Al) (4) from 0.1 to 0.6; OH-/(Si+Al) (4) from 0.15 to 0.52; UDM (9) (9)
Beta
synthesised considering this factorial design is 144. The materials (28) (28)
obtained in the studied area are amorphous materials, zeolite beta, Growing 0 0 0
and another dense material (named as UDM). Beta (1) (2) (3)
1 2 0 3
“ITQ-21/30” - The gel composition is explored by varying the Mixture
(1) (1)
following molar ratios: Al/(Si+Ge), MSPT/(Si+Ge), F-/(Si+Ge) and
Growing 0 0 0
Si/Ge. A factorial experimental design (4 x 32 x 22 =144) is selected
Mixture (1) (2) (3)
for studying simultaneously the crystallization time and the
composition of the gel varying the molar ratios of the components. 76 12 11 29 2 2 2 84.3%
Two zeolites are found in the area of the phase diagram studied: (97%)
ITQ-21 and ITQ-30.
Considering the selected applications, an adapted methodology
“ITQ-33” - Hexamethonium is used as structure directing agent
is still lacking taking into account that one specific structure can
(SDA). An initial experimental factorial design (3×43) is selected.
present large differences in peak intensity and diffraction angle
Si/Ge, TIII/(Si+Ge), OH-/(Si+Ge), and H2O/(Si+Ge) are the
within the XRD diffractogram, depending on the level of
synthesis variables. This experimental design considers the
crystallinity, crystallite size, and chemical composition.
following four molar ratios (level): Si/Ge (4) ranging from 2 to 30;
Subsequently, the synthesized powder which presents a mixture of
B/(Si+Ge) (4) from 0 to 0.05; OH-/(Si+Ge) (3) from 0.1 to 0.5; and
phases and impurities makes the automatic determination of
H2O/(Si+Ge) (4) from 5 to 30. The number of synthesized samples
crystalline phases through X-ray diffraction data a real challenge. A
is 192. Hexamethonium is a promising structure directing agent
closer look on mistakes should allow understanding why the
(SDA), since this molecule has a high number of degrees of freedom.
methodology is failing. Therefore, some representative failures are
Such flexibility allows different conformations that stabilize several
extracted from the above treatments, see Figures 1 to 3.
competing structures, such as EU-1, ITQ-17, ITQ-22, ITQ-24, SSZ-
31, a lamellar phase, and the new structure, ITQ-33.
PolySnap2© operations[4] can be summarized as follows:
i) each input sample is checked against a database of known phases,
ii) the criterion based on the Pearson and Spearman correlation
3
coefficients is calculated taking into account all measured data Figures 2-a ,3-b, and 2-c highlight a first problem inherent to
points (full profile), iii) if the pattern correlations do not indicate a methodologies for which a statistical criterion based on a correlation
good match, then the percentage composition of the sample is coefficient is employed to measure the strength (and direction) of
determined, i.e. quantitative analysis, using the singular value the relationship between diffractograms. As highest peaks get the
decomposition (SVD) for matrix inversion, and all the known greatest influence on the variation of the criteria, the local
phases as potential elements of the composition. overlapping of diffractograms considering the reflections with high
intensities, such as for ITQ-22 and ITQ-24 at 11° and between 20°
and 23°, see Figure 2-a, can mislead the methodology when used in
Predicted phases the case of ITQ-33. The same confusion appears for the sample
containing both zeolite Beta and UDM in Figure 3-b. In this case,
Other mixtures
Lamellar + 24
22 + 24 + 33
Amorphous
smaller peaks belonging to UDM are clearly visible, but their
Lamellar
22 + 24
17 + 24
24 + 33
SSZ-31
ITQ-17
ITQ-24
ITQ-22
EU-1 influence remains relatively too low compared to the main phase.
This problem is tackled in PolySnap2© as it integrates various
techniques and criteria in order to better handle such kind of
complications. Gilmore et al. define a general criterion as the linear
55 (55)
Am.
55
8 (17)
19
SSZ-31
13 (15)
(1)[b]
16
3
18 (46)
Lam.
46
28
ITQ-17
0 (0)
0
Real phases
Lam. + 24
0 (8)
8
24 + 33
0 (2)
2
17 + 24
0 (2)
2
22 + 24
0 (2)
2
22 + 24 + 33
0 (1)
1
1
(187/192=97.4%)
131/192=68.2%
19 (46)
13 (14)
10 (17)
40 (36)
95 (54)
13 (3)
0 (1)
0 (4)
0 (2)
0 (2)
0 (8)
1 (0)
1 (2)
Figure 2. Predicted errors by using PolySnap2. a) The real phase is ITQ-22 (#19),
Table 2. Confusion matrix with the real phases in the experimental design “ITQ-33” being the software prediction mixture ITQ-22 with ITQ-24, b) it is growing the
versus predicted classes obtained with PolySnap2 indicated on the left hand-side of Lamellar phase, being the software prediction “amorphous”, and c) it predicts pure ITQ-
the cell, and versus ATW indicated on the right hand-side of the cell (in italic and 17 (#217), being a mixture between ITQ-17 and ITQ-24.
between brackets).
[a]
ITQ-24 + amorphous. [b] SSZ-31 + amorphous.
4
Similarly, Figures 2-b and 3-a point the problem of detecting a To summarize, when facing difficult problems where samples
growing phase in presence of amorphous. As explained in Ref[3], if present large differences in the intensity of peaks and diffraction
the ratio of non-background to background intensity falls below a angles, the traditional methodologies obviously fail. As basic
default limit of 3% and if, additionally, there are less than a default statistical correlation coefficients allow the identification of pure
value of 3 peaks, the sample is flagged as non-crystalline. Here, the phases in most of the cases, except for the latter, the real strength
conception of the methodology consisting in an a priori separation and need of new strategies is expected where the current
of the different situations, i.e. one pure phase-mixture of phases- methodologies does not succeed or simply warns for further user
amorphous sample, through thresholds explains such errors. attention. Such tricky situations can be divided into 5 groups: i) the
treatment of the highest peaks which either mislead the
Another intricacy encountered during the analysis of “ITQ-33” crystallographic phase assignment due to local overlapping, Figure
data is represented in Figure 4. It can be observed how the 2-a, or hide the formation of competitive phases due to their
increasing integration of Germanium modifies the dimension of the relatively big weight into the criterion, Figure 2-c; ii) the
unit cell due to [Ge-O] longer bond length than [Si-O], and simultaneous occurrence of small peaks representative of a given
consequently shifts the diffractograms along x-axis. The difficulty growing phase does not take enough importance among the evident
of such situation arises from the fact that the shifting is not constant presence of another phase, Figure 3-b; iii) the treatment of partially
along the entire angle range as emphasized in Figure 4-b. This has a but predominantly amorphous samples, Figures 2-b and 3-a, is not
direct influence on the criteria and thus on the identification as most accurate enough leading to numerous false negatives; iv) similarly,
techniques use the Euclidian distance at the basis of their calculation, the shifting of diffractograms along angle axis, Figure 4,
i.e. the differences of intensity are calculated at the same angle undoubtedly appears as one of the limitation of Euclidian-based
positions. The resulting is that most of these diffractograms are criteria leading to uncertainty; iv) finally the effective detection of
interpreted as mixtures of ITQ-24 and another phase, i.e. false the presence of a new phase, or zeolite structure, which remains one
positive. of the most required feature of the strategy still needs improvements.
The three samples containing the growing ITQ-33 zeolite have been
classified as mixtures of other pure phases, Table 2.
5
In the following section, details about these modifications are that defines the distance between the two sub-sequences. DTW is
given and we show how both have been combined showing our defined following by the recursive Equation 1. The computation of
methodology its unique strength and reliability/robustness DTW needs a dynamic programming approach consisting in
characteristics. creating a t×t matrix, see Figure 6. Inside the cell (i, j) of the matrix
the value |ai – bj| is stored. After the cells are filled, the search for
the best warping path begins. It is the path beginning in position
Corrective actions
(1,1) and finishing in (t, t) that minimizes the sum of the cells it goes
through, see Figure 6.
The treatment of naturally ordered (i.e. time, temperature, etc…)
sequences represents an actual challenge in machine learning (ML) Table 3. Example of two series A and B
and data mining research due to the specific structure of the data.
Most classical algorithms initially created for multi-dimensional
data are not efficient when series are considered. The reason why t 1 2 3 4 5 6 7 8
their performances suffers is mainly due to the lack of knowledge of
the intrinsic ordering data particularity. The lack of adapted
A 1 2 4 3 1 1 2 1
methodologies has induced a new field of investigation called
temporal data mining[16] which includes association rules,[17]
indexing for large databases,[18] feature mining,[19] the discovery of B 2 1 1 2 4 3 1 1
recurrent or surprising motifs,[20] classification[21] and clustering.[22]
All the works named above use a distance measure between
instances but the effort is not focused on the elaboration of a suitable
criterion considering the problem to be handled but rather on the
algorithms making use of the distances. Due to the relative
importance/impact of the similarity expression, even very
sophisticated approaches may failed, and the lack of stability when
facing various problems remains their principal drawback. Among
the whole set of distances applied for the treatment of series, the
Dynamic Time Warping[15] (DTW) appears as the most famous one.
6
crystallographic phases, previous experiments, and theoretical classification function which relates a new case to the learned cases.
calculations. The aim of analyzing additional data is to modify the IBL approaches are chosen since the similarity function that is
similarity between diffractograms in order to improve recognition of employed plays a central role as stressed by the definition given
mixtures, with or without the presence of amorphous. A way to earlier. The following are the most common IBL methods: k-nearest
achieve this is to modify the influence of the intensities for each neighbors (k-nn), locally weighted regression, and radial basis
angle position. Since the matrix which is employed for DTW function (RBF). k-nn algorithm is selected for simplicity since it is
calculation is a squared matrix of dimension t, i.e. the total number the most basic of all IBL methods, but also because the aim is to
of angles steps, it can be used to integrate weights. A possible way stress on the proposed similarity measure and not on the algorithm
to correctly assign the values of each weight is to “learn” how which uses it. Moreover, superior performances may be easily
intensities must be interpreted or modified in order to get a correct obtained as k-nn algorithm can be refined by weighting the
answer from the labelling procedure. Due to the large number of contribution of each of the k neighbors according to their distance to
angle positions, a machine learning (ML) approach must be the query point giving greater weight to closer neighbors such as for
employed. By presenting various cases (better if complicated) either RBF. One very important trait of the methodology is its
from past experiments or theoretical calculations, we expect the independence from the classification algorithm, i.e. others could be
setting of weights to improve the classification results. employed as well.
Our methodology Adaptable Time Warping (ATW) is designed, At this level of the design, it remains defining the procedure for
as all warping (including the famous DTW), and p-norm distances weight assignment. Since the aim is to improve labelling results, the
are particular cases of ATW. Let Π a t×t matrix, be the set of best matrix Π is defined as the matrix that minimizes the
parameters required to compute ATW, Π = π i j ∈ ℝ
+
classification error rate with ATW and the chosen learning strategy,
∀i, j ∈ [1, t ] . If πij=∞, ai is not allowed to be matched with bj. see Figure 7. The “brute force” solution would perform the
γA(A’i, B’j, πij) is the recursive function that computes the distance classification with all the possible values of Π, and finally keeps the
between A’i and B’j. ATW is defined by Equation 2. one that corresponds to the best error rate. Each possible matrix Π
must be tested for obtaining the error rate. However, the resulting
complexity in time of the exhaustive search is intractable. This
(
ATW ( A, B, Π ) = γ A A't , Bt' , π t t ) (Equation 2)
implies the use of a heuristic which aims at quickly finding a
(
with γ A A'i , B 'j , π i j ) solution, i.e. obtaining a locally optimal matrix Π. Genetic
algorithms (GAs) are selected because there is a body of theory
∞ if π i j = ∞ providing theoretical evidence to complement the empirical proof of
their robustness. Such techniques are now consistently used. The
0, if i = j = 1
reader is referred to Ref[26] for an introduction and advanced
' '
( )
γ A Ai , B j −1 , π i ( j −1) , if i = 1 and j > 1 discussions.
=
'
(
'
)
γ A Ai −1 , B j , π (i −1) j , if i > 1 and j = 1
ai − b j × π i j + γ A A'i −1 , B 'j , π (i −1) j
( )
'
( '
min γ A Ai , B j −1 , π i ( j −1) ) , else
( '
)
γ A Ai −1 , B j −1 , π (i −1) ( j −1)
'
π 11 … π 1t
with Π = ⋮ ⋱ ⋮ π i j ∈ ℝ + , ∀i, j ∈ [1, t ]
π ⋯ π Figure 7. ATW conception. Based on classification errors of training samples, the ∏
t1 tt matrix is modified until reaching a given termination criterion, i.e. the convergence to a
minimum. Finally the best matrix is used for unseen test samples.
7
to give results at least equal or superior to other distances, and same phase. Therefore, one advantage of ATW is to consider peak
consequently ATW is always at least equivalent or superior to all position shifts, which are common for zeolites with different Si/Ge
other distance use. In the following section, results for 20 ratio. For this reason, the dataset “ITQ-33” is an excellent proof to
benchmarks[27] and the three real datasets of zeolite investigations check the ability of the ATW to classify such materials. There are
are given, confirming empirically the mathematical proof. eight different materials (amorphous, ITQ-17, ITQ-22, ITQ-24, EU-
1, Lamellar phase, SSZ-31, and ITQ-33) with numerous mixtures
between them. The ratio Si/Ge is also varied in the study getting
Results and Discussion
variations in the peaks positions. The last challenge that shows the
experimental design is the discovering of a new crystalline phase,
The k-nn algorithm is employed with k=3. The calculation of the
ITQ-33, in a very narrow condition range. In Figure 1 it can be
error rate on the learning dataset is done using k’-CV with k’=5.
observed the occurrence of each competing phase as a function of
The lowest error rate on the validation set (i.e. the winner matrix) is
the composition of the starting gel. Despite of the non-linearity of
kept for comparison with other techniques which have been
the system, it is clearly defined the range of composition in which a
evaluated with the same validation set. The number of iteration (r)
specific phase is formed. However, there are some narrow ranges of
for ATW matrices optimization is set to 50, and the population size
composition in which two or more phases are competing. Zeolite
(p) is set to 4. pmut and q are respectively set to {0.1; 5}.
ITQ-33 was growing with ITQ-24 in an specific area of the phase
diagram. The conditions for ITQ-33 synthesis are extremely unusual,
Figure 8 and Table S1 in supplementary material shows that
due to the low OH-/(Si+Ge) ratio in a concentrated gel H2O/(Si+Ge)
ATW always constitutes the lower bound of error rates when is
= 5, in the presence of a trivalent element ((Si+Ge)/B = 20 or 50)
compared to both DTW and Euclidian distance use on the 20
and germanium (Si/Ge = 2). When some variable was changed, the
different benchmarks.[27] ATW is superior to DTW (either with the
ITQ-33 phase disappears and ITQ-22 or ITQ-24 are formed. When
best delta value or without warping window) and the Euclidian
ATW is applied, the classification error is below 4% (see Table 2),
distance for 7 benchmarks (see bold values).
considering all the mixtures obtained in the experimental design.
0.6
Euclide DTWδ(best d value)
0.5 DTW (no warping window) ATW
0.4
0.3
0.2
0.1
0
ECG
OliveOil
Gun-Point
CBF
Synthetic Control
Trace
Fish
OSU Leaf
Face (all)
Lightning-2
Lightning-7
Face (four)
50Words
Two Patterns
Yoga
Coffee
Beef
Wafer
Adiac
Swedish Leaf
Figure 8. Evaluation of ATW of 20 different benchmarks.[27] For all the cases, ATW Moreover, if the confusion matrix is studied in detail, one can see
reaches the lowest classification error indicated on the y-axis.
that 2 of the 5 errors are due to the fact that the algorithm predicts
also the amorphous materials that are contained in these samples,
being more accurate than our manual classification. Amorphous
Considering “zeolite β” dataset, DTW performs nearly two
material is considered as a regular phase, even when the amorphous
times better than the Euclidian distance but here again ATW allows
content is “impurity”. Two other errors come from two predicted
to reach a better performance. For “ITQ-21/30”, DTW and
mixtures ITQ-22/ITQ-24 while the real phases are pure ITQ-22. In
Euclidian distances perform equally while ATW decreases the error
fact, ITQ-24 was in the limit of significance, being a most minority
rate by half. In those datasets there are only few different classes
phase, and the SVD procedure assigns less than 3%.
with low number of mixtures between them. However, in the zeolite
synthesis, it is normal to crystallise a high quantity of different
materials, with mixtures of two or more different zeolites. The ATW Conclusion
methodology can not only predict the pure phase but also the
mixture ordered by decreasing degree of crystallization of each one. ATW is a new time series distance that can be adapted
For doing that, the algorithm compares to next neighbours getting according to the classification issue, using a genetic algorithm. GA
the distance to them, and the output phase is ordered depending on components (chromosomes and operators) have been adapted for the
that distance. Moreover, depending on the crystal composition, it is task in-hand. Such hybridization is shown of great interest for the
possible to obtain some shifts in the peak positions, although it is the classification goal while the adaptation of the GA to the problem
8
form allows applying successfully the new strategy to complex [5] R. Storey, R. Docherty, P. D. Higginson, C. Dallman, C. J. Gilmore, G. Barr, W.
datasets. Besides its obvious identification superiority, a large Dong, Cryst. Rev. 2004, 10, 45-56.
number of advantages are inherent to ATW such as the following: [6] C. J. Long, J. Hattrick-Simpers, M. Murakami, R. C. Srivastava, I. Takeuchi, V.
there is no background removal, no peak extraction, no smoothing, L. Karen, X. Li. Rev. Sci. Instrum. 2007, 78, 072217.
no interpolation, no region exclusion from the user, no pre-defined [7] L. A. Baumes, M. Moliner, A. Corma. ChemEngComm. 2008, 10, 1321-1324
shifts, no false negative, the best δ is automatically found, it handles [8] A. Corma, J. Catal. 2003, 216, 298-312.
properly amorphous samples, captures phases characteristics while
[9] H. Lee, S. I. Zones, M. E. Davis, Nature. 2003, 425, 385-387.
reducing the user interaction to the minimum, and finally encounters
new crystallographic phases. [10] a) A. Corma, M. J. Díaz-Cabañas, J. Martínez-Triguero, F. Rey, J. Rius, Nature.
2002 418, 514-517. b) O. B. Vistad, D. E. Akporiaye, K. Mejland, R. Wendelbo,
A. Karlsson, M. Plassen, K. P. Lillerud, Stud. Surf. Sci. Catal. 2004, 154, 731-
738. c) A. Cantin, A. Corma, M. J. Diaz-Cabanas, J. L. Jordá, M. Moliner, J. Am.
Chem. Soc. 2006, 128, 4216-4217. d) A. Corma, M.J. Diaz- Cabañas, M.
Moliner, C. Martinez. J. Catal. 241, 2, 312-318, 2006. e) A.Corma, M. Moliner,
J. M. Serra, P. Serna, M. J. Díaz-Cabañas, L. A. Baumes, Chem. Mater. 2006,
18, 3287-3296.
which can be easily coupled with existing high-throughput synthesis [14] G. Barr, W. Dong, C. J. Gilmore, J. Appl. Cryst. 2004, 37, 874-882.
apparatus, see Figure 9, is a real breakthrough. It has been shown [15] a) D. J. Berndt, J. Clifford, in Advances in Knowledge Discovery and Data
how the phase diagram corresponding to a broad research space can Mining (Ed.: U. M. Fayyad, et al.) , AAAI Press/MIT Press, Menlo Park, CA
be quickly revealed making clearer the way towards new structures. 1996, pp. 229-248; b) E. Keogh, Data Mining and Machine Learning in Time
Series Databases, at the 18th ACM SIGKDD Int. Conf. on Knowledge
Discovery and Data Mining (KDD’04). Seattle, WA, USA, 2004.
Acknowledgements [16] a) W. Lin, M. A. Orgun, G. J. Williams, Australasian Data Mining Workshop,
Macquarie Univ. and CSIRO Data Mining, 2002; b) C. M. Antunes, A. L.
EU Commission FP6 (TOPCOMBI Project) is gratefully acknowledged. We also thank Oliveira, Workshop on Temporal Data Mining, at the 7th Int. Conf. on
Santiago Jimenez for his important collaboration to set the hITeQ platform which Knowledge Discovery and Data Mining, San Francisco, CA, 2001.
integrates all programming codes of the presented methodology. Laurent A. Baumes
thanks Stephan Schunk from hte Company for his comments on the technique. [17] G. Das, K. Lin, H. Mannila, G. Renganathan, P. Smyth, Global partial orders
from sequential data, at the 4th Int. Conf. on Knowledge Discovery and Data
Mining, New York, NY, 1998.
[1] a) M. Fransen, Am. Lab. 2002, 34, 42-49; b) R. Meier, M. Dirken, Int. Cement &
Lime Journal. 2002, 1, 44-47. [18] a) E. Keogh, C. A. Ratanamahatana, J. Knowl. Inf. Syst., 2005, 7, 358–386; b) B.
K. Yi, H. V. Jagadish, C. Faloutsos, Proceedings 14th IEEE Int. Conf. on Data
[2] a) I. Takeuchi, C. J. Long, O. O. Famodu, M. Murakami, J. Hattrick-Simpers, G. Engineering. 1998, pp: 201 – 208.
W. Rubloff, M. Stukowski, K. Rajan. Rev. Sci. Instrum. 2005, 76, 062223; b)
X'Pert HighScore Plus software from PANalytical, see website visited Nov. 20th [19] N. Lesh, M. J. Zaki, M. Ogihara, IEEE Intell. Syst. App. 2000, 15(2), 48-56.
2008, http://www.panalytical.com/index.cfm?pid=547; c) D. L. Bish, S. A.
[20] a) J. Buhler, M. Tompa, J. Comput. Biol. 2002, 9, 225-242; b) B. Chiu, E.
Howard, J Appl Crystallogr. 1988, 21, 86-91; d) D. L. Bish, S. J. Chipera, in
Keogh, S. Lonardi, Probabilistic Discovery of Time Series Motifs, at the 9th Int.
Advances in X-ray Analysis (Ed.: C. Barrett, et al.), Plenum Press, New York
Conf. on Knowledge Discovery and Data Mining, Washington, DC, 2003.
1988, 31, pp. 295-308; e) D. L. Bish, S. J. Chipera, in Advances in X-ray
Analysis (Ed.: P. Predecki, et al.), Plenum Press, New York 1995, 38, pp. 47- [21] a) J. C. Chappelier, M. Gori, A. Grumbach, in Sequence Learning: Paradigms,
57; f) S. J. Chipera, D. L. Bish, Powder Diffr. 1995, 10, 47-55; g) F. H. Chung, J. Algorithms and Applications (Ed.: R.Sun, G. L. Giles), Springer-Verlag, London,
Appl. Crystallogr. 1974, 7, 519-525; h) RockJock – Uses Microsoft Excel 2001, pp. 105-134; b) D. L. James, R. Miikkulainen, Advances in Neural
Macros and the Solver function to perform a whole-pattern modified Rietveld- Processing Systems. 1995, 7, 577-584; c) P. Somervuo, T. Kohonen, Neural
type refinement to perform quantitative analysis. Written by Dennis D. Eberl, Process. Lett. 1999, 10, 151-159.
the software was published in 2003 as U.S.G.S. Open-File Report 03-78,
“Determining Quantitative Mineralogy from Powder X-ray Diffration Data”. [22] a) F. R. Lin, L. S. Hsieh, S. M. Pan, Learning clinical pathway patterns by
Available via FTP from ftp://brrcrftp.cr.usgs.gov/pub/ddeberl/RockJock. i) See hidden markov model, at the 38th Annual Hawaii Int. Conf. on System Sciences
JADE from http://www.rigaku.com/index_world.html. (HICSS'05), Big Island, Hawaii, USA, 2005. b) J. Lin, M. Vlachos, E. Keogh, D.
Gunopulos, Iterative incremental clustering of time series, at the IX Conf. on
[3] a) C. J. Gilmore, G. Barr, J. Paisley, J. Appl. Crystallogr. 2004, 37, 231–242; b) Extending Database Technology (EDBT 2004), Crete, Greece, 2004.
G. Barr, W. Dong, C. J. Gilmore, J. Appl. Crystallogr. 2004, 37, 243–252.
9
[23] E. Keogh, M. Pazzani, Scaling up Dynamic Time Warping for Datamining Theory, MIT Press, Cambridge, MA. 1999. d) D. Whitley. A genetic algorithm
Applications, at the 6th Int. Conf. on Knowledge Discovery and Data Mining, tutorial. Statistics and Computing. 1994, 4, 65–85.
Boston, MA, 2000.
[27] E. Keogh, X. Xi, L. Wei, C.A. Ratanamahatana. 2006, The UCR Time Series
[24] E. Keogh, M. Pazzani, Derivative Dynamic Time Warping, at the 1st SIAM Benchmark: www.cs.ucr.edu/~eamonn/time_series_data/
International Conference on Data Mining, Chicago, USA, 2001.
Received: ((will be filled in by the editorial staff))
[25] H. Sakoe, S. Chiba, IEEE T. Acoust. Speech. 1978, 26, 43-49.
Revised: ((will be filled in by the editorial staff))
[26] a) D. E. Goldberg, The Design of Innovation: Lessons from and for Competent Published online: ((will be filled in by the editorial staff))
Genetic Algorithms, Kluwer Academic Publishers, Norwell, MA. 2002; b) L. M.
Schmitt, Theory of Genetic Algorithms, Theoretical Computer Science, 2001,
259, 1-61; c) M. D. Vose, The Simple Genetic Algorithm: Foundations and
10
Making clearer the way towards Design of ATW
new structures methodology for
automatically, quickly and
Laurent A. Baumes*, Manuel reliably deciphering X-ray
Moliner and Avelino Corma* diffraction patterns
Design of a Full-Profile
Matching Solution for High-
Throughput Analysis of Multi-
Phases Samples Through
Powder X-Ray Diffraction
11