Accepted Manuscript

Adaptive refinement of hierarchical T-splines

L. Chen, R. de Borst

PII: S0045-7825(18)30156-7
DOI: https://doi.org/10.1016/j.cma.2018.03.032
Reference: CMA 11838

To appear in: Comput. Methods Appl. Mech. Engrg.

Received date : 31 May 2017

Revised date : 23 March 2018
Accepted date : 25 March 2018

Please cite this article as: L. Chen, R. de Borst, Adaptive refinement of hierarchical T-splines,
Comput. Methods Appl. Mech. Engrg. (2018), https://doi.org/10.1016/j.cma.2018.03.032

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form.
Please note that during the production process errors may be discovered which could affect the
content, and all legal disclaimers that apply to the journal pertain.
*Manuscript
Click here to download Manuscript: ref-tspline_annotated.pdf Click here to view linked References

1
2
3 Adaptive Refinement of Hierarchical T-splines
4
5
6 L. Chena , R. de Borsta,∗
7
a University of Sheffield, Department of Civil and Structural Engineering, Sir Frederick Mappin Building, Mappin Street, Sheffield S1 3JD, UK
8
9
10
11
12
13 Abstract
14 We present an adaptive local refinement technique for isogeometric analysis based on hierarchical T-splines. An
15 element-wise point of view is adopted, which exploits Bézier extraction, and allows adaptive refinement of standard
16 hierarchical T-splines and truncated hierarchical T-splines in a straightforward and unified manner. No explicit basis
17 function operations are required to build the hierarchical basis function space, as only matrix manipulations are in-
18
volved. This makes the efficiency superior to that of existing implementations. In particular, the implementation of
19
truncated hierarchical T-splines requires no explicit truncation of the basis functions. In the analysis, a multi-level
20
21 T-mesh is constructed by successive cell subdivisions of an initial, coarse T-mesh. An important feature is that Bézier
22 extraction is employed to compute the refinement operator between two successive hierarchical levels, and that, at
23 each level, Bézier extraction is applied to obtain the stiffness matrix without, initially, considering multi-level inter-
24 action. This interaction is recovered through a subdivision operator. Numerical examples are presented for validation
25 purposes, and to assess the convergence properties.
26
Keywords: Hierarchical T-splines; Bézier extraction; isogeometric analysis; adaptive refinement
27
28
29
30
31 1. Introduction
32
A main advantage of isogeometric analysis is that NURBS functions commonly employed in the Computer Aided
33
34 Geometric Design model, can directly be employed in the analysis model [1, 2], thus reducing the effort expended
35 in (re)meshing, and improving, or even eliminating the error committed in the geometry description. However, the
36 tensor product structure of NURBS prevents local mesh refinement. To obviate this drawback, various local refine-
37 ment strategies have been proposed, including T-splines [3–5] and hierarchical and truncated hierarchical T-splines as
38 further developments [6–8], LR-splines [9–12], hierarchical and truncated hierarchical B-splines [13–18], and PHT-
39 splines [19–21]. It is further noted that adaptive splines also behold promise as an effective tool for local refinement
40 in isogeometric analysis [6, 7, 13, 17, 22].
41 T-splines were introduced by Sederberg [3, 4]. They remove the rigidity of the tensor product structure of NURBS
42 by allowing extra vertices to be inserted. Their first use in (isogeometric) analysis is in Reference [23], and the pos-
43 sibility to use them in existing finite element datastructures through Bézier extraction has been described in [24].
44 Mathematical properties of T-splines, for instance linear independence and partition-of-unity property of basis func-
45 tions, are given in [25–27]. The local refinement of T-splines has been investigated in [5–7, 28, 29]. Of particular
46 relevance for the work reported here is Reference [6, 7], in which the concepts of hierarchical and truncated hierar-
47 chical T-splines were proposed. They enable to combine the ability to locally refine hierarchical B-splines with the
48 geometrical representation capability of T-splines.
49
In this contribution, we will develop the adaptive hierarchical refinement of T-splines. An element-wise point of
50
view, enabled through Bézier extraction, will be employed for implementation purposes. A multi-level, hierarchical
51
T-spline mesh is generated by successive cell subdivisions of an initial, coarse T-spline mesh. At each hierarchy level
52
53 the element stiffness matrices are obtained by applying Bézier extraction without, initially, considering multi-level
54 interaction. This interaction is enforced through the introduction of a subdivision operator. Two cases are considered:
55
56 ∗ Corresponding author
57 Email address: r.deborst@sheffield.ac.uk (R. de Borst)
58
59 Preprint submitted to Computer Methods in Applied Mechanics and Engineering March 23, 2018
60
61
62
63
64
65
 1
2
3 standard hierarchical T-spline bases and truncated hierarchical T-spline bases. It is noted that the hierarchical bases
4 are not implemented explicitly in the formulation, but implicitly through matrix multiplications.
5
The paper is structured as follows. Section 2 gives a succinct summary of T-splines and Bézier extraction, includ-
6
ing the construction of the Bézier extraction operator. Some basic notions of hierarchical T-spline basis functions are
7
8 discussed in Section 3. Section 4 illustrates the use of Bézier extraction for hierarchical T-splines, and Section 5 pro-
9 vides information on adaptive hierarchical refinement. The paper is completed with numerical examples in Section 6
10 to validate the approach, and with some concluding remarks.
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43 Figure 1: Example of a cubic T-spline mesh. The object is given in the index domain (i, j), in the physical domain
44
(x1 , x2 ), and in the parameter domain ξ1 , ξ2 . The element is also shown in the parent domain.
45
46
47
48
2. T-spline and Bézier extraction
49
50 In this Section, we will first give a concise overview of T-splines and Bézier extraction, please see [23, 24] for a
51 more in-depth discussion.
52
53
2.1. Fundamentals of T-splines
54
55 T-splines are constructed on a T-mesh. For two dimensional objects, the T-mesh is a mesh of quadrilateral el-
56 ements, which allows T-junctions. An example of a cubic T-spline mesh is given in Figure  1. In  the figure, the
57 index domain (i, j), the physical domain (x1 , x2 ), the parameter and sub-parameter domain ξ1 , ξ2 of T-splines are
58
59 2
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14 (a) Determination of the local knot nvector and theo corresponding
n o blending nfunction for aoquadratic
n T-spline
o mesh. The local knot
15 vectors for blue anchor A are Ξ1A = ξ31 , ξ51 , ξ61 , ξ71 = 13 , 32 , 1, 1 and Ξ2A = ξ12 , ξ22 , ξ32 , ξ42 = 0, 0, 13 , 32 .
16
17
18
19
20
21
22
23
24
25
26 (b) Determination of the local knotn vector and the corresponding
o n blending
o function
n for a cubic T-splineo n mesh. The o local knot
27 vectors for blue anchor A are Ξ1A = ξ21 , ξ31 , ξ51 , ξ61 , ξ71 = 0, 31 , 32 , 1, 1 and Ξ2A = ξ12 , ξ22 , ξ32 , ξ42 , ξ52 = 0, 0, 13 , 32 , 1 .
28
29 Figure 2: Determination of the local knot vector and the construction of blending functions for quadratic and cubic
30 T-spline meshes.
31
32
33 shown. To carry out Gaussian integration, the element e must be mapped from the physical domain onto the parameter
34 domain, and then onto the parent domain ξ̃l ∈ [−1, 1], see Figure 1.
35 In a T-spline mesh, elements are defined by the edges of T-spline mesh and continuity reduction lines [26], see
36 the areas that are shaded gray in Figures 1 and 2. Anchors are prescribed in the index domain and in the parameter
37 domain [23], and a multivariate blending function is attached to each of them. In Figure 1, the anchors are placed
38 at
 each
 vertex of a cubic T-mesh. To obtain the blending function associated with an anchor, a local knot vector
39 Ξi has to be defined, with n denoting the number of anchors. In Reference [26] the construction of local
i=0, ··· , n−1
40 knot vectors for T-meshes of even and of odd degrees has been described in detail. Figure 2 gives an example of the
41 determination of local knot vectors. On the basis of local knot vectors, we obtain the blending functions associated
42
with an anchor [24], see Figure 2.
43
44
2.2. Bézier extraction fundamentals
45
46 The blending function NA is defined over entire support of anchor A, see Figures 1 and 2. It is cumbersome to
47 directly incorporate a blending function in a standard finite element code. However, Bézier extraction overcomes this
48 by representing blending functions as element-wise Bernstein shape functions [24], and this approach will be adopted
49 herein. We suppose that the domain can be decomposed into E elements with n anchors. The local knot vectors of
50 anchor i are Ξi1 and Ξi2 . Then, the blending function Ni of anchor i over element e can be expressed as
51  
52 Nie (ξ) = Cei T Be (ξ) (1)
53 where Be (ξ) contains (element-local) Bernstein polynomials with dimension (p + 1)2 × 1 [26]. Here, we consider
54 T-splines with same polynomial degree, p, in the ξ1 and the ξ2 parametric direction.
55 Cei is the Bézier extraction operator of anchor i over element e, which is determined by the tensor product [26]:
56
57 Cei = Ce2 e1
i ⊗ Ci (2)
58
59 3
60
61
62
63
64
65
 1
2
3 where Ce1 e2 1 2
i and Ci are univariate Bézier extraction operators of anchor i over element e in the ξ and the ξ parametric
4 direction, respectively.
5
We take anchor A in Figure 2(a) as an example n to illustrate
o n the calculation o of Bézier n extraction
o operator
n Ceio.
6 A 1 1 1 1 1 2 A 2 2 2 2 1 2
7 The local knot vectors of anchor A are Ξ1 = ξ3 , ξ5 , ξ6 , ξ7 = 3 , 3 , 1, 1 and Ξ2 = ξ1 , ξ2 , ξ3 , ξ4 = 0, 0, 3 , 3 ,
8 respectively.
h We evaluate
i h thei Bézier extraction operator of anchorhA over i element
h i b, which is bounded in the parameter
2 1
9 domain by ξ5 , ξ6 × ξ2 , ξ3 and in the sub-parameter domain by 3 , 1 × 0, 3 . Figure 3 shows the blending function
1 1 2 2

10 NAb for each parametric direction. The part of NAb in the ξl parametric direction, i.e. NAbl , is plotted as a solid line over
11 element b. Now, we express the blending function NAbl as a linear combination of Bernstein polynomials Bbl over
12 element b:  b1   b1 
13  B1   B1 
   
14 h i   h i  
b1 
15 NA = [C A ] B = C A1 C A2 C A3  B2  = 2 1 0  Bb1
b1 b 1 b1 b1 b1  b1  1  (3)
16    2 
   
17 Bb1
3 Bb1
3
18  b2   b2 
 1  B B
 1 
19    
h i   h i  
20 NA = [C A ] B = C A1 C A2 C A3  B2  = 0 1 2  Bb2
b2 b 2 b2 b2 b2 b2  b2  1  (4)
21    2 
   
22 Bb2 Bb2
3 3
23
Employing Equation (2) and considering Equations (3) and (4), we obtain the Bézier extraction operator of anchor A
24
25 over element b, as follows: h i
26 CbA = 0 0 0 21 1 0 14 12 0 (5)
27 When we subsequently apply Equation (2) to all elements, the Bézier extraction operator of anchor i is obtained:
28  1
 Ci 
29  
30 Ci =  ...  (6)
 
31 CiE
32
33 The dimension of Ci is E (p + 1)2 × 1, where E is the total number of elements.
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50 Figure 3: The blending function NAbl of anchor A and the Bernstein polynomials over element b in the ξl parametric
51 direction.
52
53 When we consider Equation (6) for all anchors, we can express the n blending functions as a linear combination
54 of Bernstein basis functions:     
55 N1 (ξ) CT1   B1 (ξ) 
     
56 N (ξ) = CB (ξ) =  ...  =  ...   ...  (7)
     
57 Nn (ξ) T
Cn B (ξ) E
58
59 4
60
61
62
63
64
65
 1
2
3 where C has the dimension n × E (p + 1)2 and B (ξ) is of E (p + 1)2 × 1. The blending functions with support over
4 element e are then expressed as:
5
Ne (ξ) = Ce Be (ξ) (8)
6
7 with Ce element Bézier extraction operator.
8
9
10 3. Hierarchical T-splines
11
Hierarchical T-splines were introduced in Reference [6] and allow for local refinement of a given, normally coarse
12
13 T-spline mesh. The basic idea of hierarchical T-splines is to locally enrich the approximation space by replacing
14 selected coarse grid T-splines by fine grid T-splines. We now give a concise outline of the mechanism how to construct
15 hierarchical T-splines.
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32 (a) T-spline mesh T α (b) T α+1 created by subdividing T α
33
34 Figure 4: Construction of T-spline mesh T α+1 from T α . The anchors are indicated by circular dots. Black denotes the
35 mesh and the anchors of T-spline mesh T α , and red stands for the mesh and the anchors generated for T α+1 .
36
37
38
3.1. Nested spaces and domains
39  
40 A hierarchical T-spline space is constructed from a finite sequence of P nested T-spline spaces T l
  l=0, ··· , P−1
41 bounded by P open sets Ωl . The nested nature of T-spline space defines the nested domains for hierarchy:
42 l=0, ··· , P−1
43
44 T 0 ⊂ T 1 ⊂ · · · ⊂ T P−1 ΩP−1 ⊆ ΩP−2 ⊆ · · · ⊆ Ω0 (9)
45
46 The sequence of P T-spline meshes is built by subdividing each effective rectangular cell in T α into two or four
47 congruent cells such that T α ⊂ T α+1 , α = 0, · · · , P − 2, where the term ’effective rectangular cell’ refers to a cell
48 with non-zero parametric length in at least one parametric direction. Figure 4 illustrates the algorithm to generate the
49 T-spline mesh T α+1 from T α . Cell A has a zero parametric length in both directions, which results in no subdivision
50 in A. Cell B only has a non-zero length in the ξ2 parametric direction. It is divided into two congruent cells in T α+1 .
51 For cell C, the parametric length is non-zero in both directions, which leads to four congruent cells in T α+1 .
52
53 3.2. Hierarchical T-spline bases
54 The algorithm of [6] has been adopted for the construction of hierarchical T-spline bases, but we rephrase the
55 algorithm below. Defining τ as a T-spline basis function and by T we denote the T-spline basis function space [13].
56 The hierarchical T-spline bases H is recursively built as follows:
57
58
59 5
60
61
62
63
64
65
 1
2
n o
3 (1) Initialize: H 0 = τ ∈ T 0 : supp τ , ∅ .
4
5
6 (2) Construct: H α+1 from H α in a recursive manner: H α+1 = Hcoarse
α+1 α+1
∪ Hfine α+1
, α = 0, · · · , P − 2 where Hcoarse =
n o n o
7 τ ∈ H : supp τ * Ω
α α+1
; α+1
Hfine = τ ∈ T α+1
: supp τ ⊆ Ω α+1
.
8
9 (3) Set H = H P−1 .
10
11 Using linear combinations between basis functions on hierarchy level α and α + 1, we can obtain a truncated
12 hierarchical basis function space [7, 14]:
13 n o
14 (1) Initialize: H 0 = τ ∈ T 0 : supp τ , ∅ .
15
16 (2) Construct: H α+1 from H α in a recursive manner: H α+1 = Hcoarse
α+1 α+1
∪ Hfine α+1
, α = 0, · · · , P − 2 where Hcoarse =
17 n o n o
18 τ ∈ H ∧ supp τ * Ω
α α+1
; α+1
Hfine = τ ∈ T α+1
: supp τ ⊆ Ω α+1
.
19
20 (3) Set H = H P−1 .
21
22 Figure 5 illustrates the hierarchical and truncated hierarchical basis function spaces for the univariate case, while
23 Figures 6 and 7 do so for the bivariate case. The mathematical properties of hierarchical and truncated hierarchical
24 basis functions, e.g. nested nature, linear independence and partition of unity property have been discussed in [6, 7].
25
26
4. Bézier extraction of hierarchical T-splines
27
28 We will now extend the Bézier extraction framework to allow for the implementation of hierarchical T-splines.
29 This procedure is similar to that used for hierarchical B-splines [17]. As stated, an element-wise point of view is
30 adopted, which conforms ideally to Bézier extraction framework. The hierarchical T-spline basis functions are defined
31
over multiple hierarchy levels. Strong boundary condition is imposed over different hierarchy levels [13]. This results
32
in a nested hierarchical element structure.
33
34
35 4.1. Data structure
36 The element-wise implementation of hierarchical T-splines is a natural choice in adaptive finite element analysis.
37 The hierarchical T-spline bases consist of T-splines over multiple hierarchical levels with same polynomial degree.
38 Below, we will outline the data structure of multi-level hierarchical bases.
39
40 4.1.1. Multi-level mesh and Bézier extraction operator
41 We first construct a hierarchy of P levels. The T-spline basis functions at each hierarchy level are defined over
42 n oni −1
43 local knot vector set Ξi (i = 0, 1, ... P − 1). Ξi is defined as Ξi = Ξij , in which ni denotes the number of anchors
j=0
44 at level i. Ξi is obtained by successive cell subdivision within Ωd , starting from initial local knot vector Ξ0 , where Ωd
45 denotes the parameter domain. The algorithm is visualised in Figure 4. In this process, the nested parameter domains
46 Ωdi+1 ⊂ Ωid and the nested local knot vectors Ξi ⊂ Ξi+1 are obtained, see Figure 4. Each knot vector Ξi defines a set
n oni
47 of T-spline basis functions Ni = N ij , which in turn forms a nested T-spline approximation space T i , see Figure
48 j=1

49 6. With Ξi and Ωid , the Bézier extraction operator Ci can be obtained for each anchor at level i, see Section 2.2.
50 Furthermore, T-spline basis functions at level i can be defined in terms of elements at level i + 1 by the Bernstein
51 polynomials Bi+1 (ξ):
52 Ni = CiR Bi+1 (ξ) (10)
53 where CiR denotes Bézier extraction operator of each anchor at level i over elements at level i + 1.
54
55
56
57
58
59 6
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42 Figure 5: Definition of basis function space; (a) illustration of basis function sets Al , Al− and Al+ ; (b) final hierarchical
43 basis functions; (c) final truncated hierarchical basis functions.
44
45
46 4.1.2. Subdivision operator and control point
47 Due to the nested nature of T i , the T-spline basis functions at hierarchy level i can be described by the T-spline
48 basis functions at hierarchy level j:
49 Yj−1
i i, j j
50 N =S N = Sl,l+1 Nl+1 (11)
51 l=i
52
53 where Sl,l+1 is the subdivision or refinement operator [26]. For the NURBS basis functions, the subdivision operator
54 Sl,l+1 can be directly obtained from the linear relation between basis functions from two hierarchy levels [15, 18]. It is
55 noted that Sl,l+1 is a matrix with a high degree of sparsity. Using Equations (7) and (10), Sl,l+1 can now be computed:
56
57 Nl = ClR Bl+1 (ξ) = Sl,l+1 Cl+1 Bl+1 (ξ) (12)
58
59 7
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
(a) T-spline mesh at level 0 (b) T-spline bases at level 0
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38 (c) T-spline mesh at level 1 (d) T-spline bases at level 1
39
40 Figure 6: T-spline meshes and bases at hierarchy level 0 and 1
41
42
43 where ClR denotes Bézier extraction operator of each anchor at level l over elements at level l + 1, and Cl+1 represents
44 Bézier extraction operator of each anchor at level l + 1.
45 Expanding Equation (12) in a vector form, it obtains
46
 l T   l,l+1 T   
47  [CR0 ]  [S1 ]   [C1l+1 ]T 
48  .   .   .. 

 ..  =  ..   (13)
49  l T   l,l+1 T   l+1 T 
.
50 [CRnl ] [Sn −1 ] [Cnl+1 −1 ]
l
51
52 with nl and nl+1 the number of anchors at level l and l + 1, respectively. ClRi denotes Bézier extraction operator of
53 anchor i at level l over elements at level l + 1, whose dimension is El+1 (p + 1)2 × 1. Cil+1 represents Bézier extraction
54 operator of anchor i at level l + 1 with the dimension of El+1 (p + 1)2 × 1. El+1 is the number of elements at level l + 1.
55 The row values of Sl,l+1 are obtained from:
56
57 ClRi = [Cl+1 ]T Sil,l+1 for i = 0, · · · , nl − 1 (14)
58
59 8
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15 Active T-spline bases at level 0 Active T-spline bases at level 1 Hierarchical T-spline bases
16 (a) Active basis functions at level 0 and level 1 and corresponding hierarchical T-spline bases
17
18
19
20
21
22
23
24
25
26
27
28
29 Active T-spline bases at level 0 Active T-spline bases at level 1 Truncated hierarchical T-spline bases
30 (b) Active basis functions at level 0 (already truncated) and level 1 and corresponding truncated hierarchical T-spline bases
31
32 Figure 7: Hierarchical and truncated hierarchical T-spline bases
33
34
35 With the subdivision operator, the coordinates and weights of anchors in a refined T-spline mesh at hierarchy level i
36 can be determined recursively [26]:
 i−1 T
37 Y l,l+1  0
i i,0 T 0
38 Pw = [S ] Pw =  S  Pw (15)
39 l=0
40  
where Piw are the weighted control points at level i. Each weighted control point is defined as Piw, j = wij xi1 j , wij xi2 j , wij .
41
42 With the subdivision operator defined in Equation (12), and the weight w from Equation (15), the T-spline bases
43 can also be written in a rational form [17]:
44 P l,l+1 l+1
l l (ξ) wlI S IJ N J (ξ) X S l,l+1
45 w N J IJ
RlI (ξ) = I l I = = w l
I Rl+1 (ξ)
l+1 J
(16)
46 W (ξ) W l (ξ) w
J J
47
48 Using the rational T-spline bases, the subdivision operator must be modified as below:
49
50 l,l+1 wlI l,l+1
S̄ IJ = S IJ (17)
51 wl+1
J
52
53 to replace the standard subdivision operator Sl,l+1 .
54
55 4.2. Multi-level implementation of hierarchical T-splines
56 We next take univariate bases to illustrate the construction of hierarchical basis functions. Due to the nested
57 structure of T-spline basis function space, the multi-variate case can be derived straightforwardly. To construct the
58
59 9
60
61
62
63
64
65
 1
2
3 hierarchical basis function space H, the active elements and basis functions in the multi-level hierarchy must be
4 identified. The active element is chosen by a certain marking criterion, for example a posteriori error estimator [16].
5
No overlap or gap may exist between the active elements from different hierarchy levels, Figure 5(a). In the figure,
6
the parameter domain of active elements is plotted in green, where the parameter domain of active elements is defined
7
8 as the union of three parameter domains associated with hierarchy level l:
9 [
l−1 [
P−1 [
10 Ωd = Ωl− l l+
E Ai + E Al + E Ai E Ai = Ωe,i
d + E A + Ωd = with d (18)
11 i=0 i=l+1 e
12
13 where P is the number of hierarchy levels: E Ai represents the parameter domain of active elements at hierarchy level
14 i, Ωe,i l−
d denotes the parameter domain of element e at hierarchy level i; Ωd is the parameter domain of active elements
15 l+
at coarser hierarchy levels; and Ωd represents the parameter domain of active elements at finer hierarchy levels.
16 Herein, an element-based selection approach is employed to define the hierarchical basis function space H [17].
17 On the basis of Ωd , Ωl+ l−
d and Ωd , one defines three sets of basis function spaces:
18
19  l n l \ o


 A = N ∈ T l
: supp N l
E l
, ∅
20 


j j A



21 

 [
P−1 \

 n o
22 
 A+ = Al+ with Al+ = N lj ∈ Al : supp N lj Ωl+
 d ,∅
23 
 (19)



l=0
24 


25 

 [
P−1 n \ o


26 

 A − = Al− with Al− = N lj ∈ Al : supp N lj Ωl−
d ,∅

27 l=0
28
29 Al denotes the union of basis functions with support over active elements at hierarchy level l, Figure 5(a). Al+ denotes
30 the basis function set in Al with support over active elements at finer hierarchy levels, plotted as dashed lines, see
31 Figure 5(a). Al− represents basis functions in Al which have a support over active elements at coarser hierarchy levels,
32 as indicated by dotted lines in Figure 5(a). Finally, the space of hierarchical basis functions H is defined as:
33
[
P−1
34
H= Ala with Ala = Al \ Al− (20)
35
l=0
36
37 where "\" is the logic NOT, Ala denotes active basis functions at hierarchy level l, see Figure 5(b), and H denotes
38 standard hierarchical basis function space [6].
39 From the linear combination between basis functions at hierarchy level l and l+1, a space of truncated hierarchical
40 basis functions is obtained [7], see Figure 5(c):
41
42 [
P−1 n o
43 HT = AlT,a with AlT,a = τli ∈ Ala : supp τli * E Al+1 (21)
44 l=0
45 n P o
46 where τli is defined as τli = τli ∈ T l : τli = S il,l+1
j Nj
l+1
, see Equation (11).
47 Considering the active elements and basis functions, one can implement hierarchical basis functions to obtain the
48 stiffness matrix in a multi-level, adaptive manner. First, using Bézier extraction, the stiffness matrix of active ele-
49 ments at each hierarchy level is obtained, without consideration of possible interaction of multi-level basis functions.
50 Assembling the stiffness matrix at each hierarchy level, a global system of equations is obtained:
51
52 KU = F (22)
53
54 where U contains nodal degrees of freedom at each hierarchy level. The force vector is given by F and K is the stiffness
55 matrix, with the submatrices Ki along the diagonal, Ki being the stiffness matrix of active elements at hierarchy level
56 i, a square sparse matrix of size 2nic × 2nic . The number of control points at hierarchy level i is given by nic . It is noted
57 that the submatrices Ki are also highly sparse.
58
59 10
60
61
62
63
64
65
 1
2
3 To enforce the interaction between multi-level hierarchical bases in Equation (22), a hierarchical subdivision
4 operator Mh is introduced, resulting in the following hierarchical system of equations [30]:
5
6 K h Uh = F h with Kh = Mh KMTh and Fh = Mh F (23)
7
8 Again, Kh is a very sparse matrix. The hierarchical subdivision operator Mh is defined as:
9 0 
10 I M̂0,1 M̂0,2 . . . M̂0,P−1 
 1 1,2 1,P−1 

11  I M̂ . . . M̂ 
 2 2,P−1  
12 Mh =  I . . . M̂  (24)
13  .. 
14  0 . 
15 IP−1
16
in which 
17 
 NIl ∈ Ala
l 1 for I=J and
18 IIJ =
 (25)
19 0 else
20
The matrix M̂l,k in Equation (24) is defined as follows:
21
 l,k
22 

 S IJ for NIl ∈ Al+
l,k
23 Standard hierarchical bases: M̂IJ =

 0 else
24
 l,k (26)
25 

 S IJ for NIl ∈ Al+ and N Jk ∈ Ak−
l,k
26 Trunctated hierarchical bases: M̂IJ = 
 0
 else
27
28
29 where S l,k
IJ has been defined in Equation (11).
30 In this contribution, standard hierarchical T-splines and truncated hierarchical T-splines have been considered in
31 a unified framework, see Equation (23). Explicit basis function operations are avoided, as only matrix manipulations
32 have to be carried out. Hence, the efficiency of the proposed approach is superior to that of recent implementations of
33 hierarchical T-splines [8]. We note that, in particular, truncated hierarchical T-splines can be applied without explicit
34 truncation of the basis functions. Therefore, the innovative aspect of this work is the multi-level implementation of
35 hierarchical T-splines by using Bézier extraction in an efficient way without basis function operations. The extension
36 of the method to three dimensions case is straightforward by employing Bézier extraction to analyse cases as in
37 [31, 32].
38 Solution of Equation (23) yields the displacements Uh for the control points associated with hierarchical bases.
39 However, a non-linear solution scheme requires the displacement vector U rather than Uh from previous iteration, cf.
40 Equation (22):
41 U = MTh Uh (27)
42
43 4.3. Implication of truncation mechanism
44
When constructing hierarchical T-spline bases, the hierarchical and truncated hierarchical basis functions are never
45
46 computed explicitly. Instead, the hierarchical subdivision operator Mh builds the hierarchical T-spline bases in an
47 implicit manner. For the hierarchical basis functions, only the contributions of active basis functions in the set A+ are
48 considered, see Equation (26). During the (multi-level) Bézier extraction of T-spline bases, the numerical integration
49 is performed separately for each active element, at each hierarchy level, regardless of whether the basis functions,
50 which have a support over the active element, are part of hierarchical bases or not. The hierarchical subdivision
51 operator Mh accounts for the activity of basis functions and recovers the correct support of active basis functions, see
52 Figure 6. For the truncated hierarchical T-spline bases, the truncation mechanism is implemented by the hierarchical
53 subdivision operator Mh in Equation (26). The contribution of basis functions in the sets A− and A+ is considered
54 next. In the calculation of the stiffness matrix Kh , the mapping of the contribution of basis functions in A−l+1 to that
55 of basis functions in Al+ is included [17]. Eventually, the stiffness matrix elements resulting from basis functions in
56 A−l+1 are mapped onto those related to Al+ and stored afterwards in Al+ . Thus, the support domain of Al+ is truncated.
57 This is explained graphically in Figure 7.
58
59 11
60
61
62
63
64
65
 1
2
3 5. Adaptive hierarchical refinement
4
5 We now proceed to adaptive mesh refinement, and we will present the steps and algorithms for element refinement
6 as well as element coarsening:
7
8 S1 Solve Equation (23) to obtain the displacement Uh and compute U using Equation (27).
9
10 S2 Estimate the approximation error. The H1 norm and the energy norm of the element residual are employed
11 herein.
12
13 S3 Mark elements for refinement and coarsening on the basis of S2.
14 S4 Refine or coarsen the marked elements. If no refinement or coarsening is required, stop the calculation, other-
15 wise return to S1.
16
17
5.1. Element marking
18
19 For element refinement, element markingncan be done
o either using quantile marking or Dörfler marking [29]. For
20 a domain Ω with E elements, element errors εQ  Q ∈ Ω ⊂ R are obtained from Step S2. Then, we define a marking
21 parameter η ∈ [0, 1] to determine elements which must be refined. Let Ω = {Q1 , · · · , QE } and εQ1 > · · · > εQE , the
22 list of elements to be refined then reads:
23
24 Quantile marking: M = {Q1 , · · · , Qk } with k = ceil (ηE)
25 X
k−1 X
E X
k X
E
(28)
26 Dörfler marking: M = {Q1 , · · · , Qk } with εQ j < η εQ j and εQ j ≥ η εQ j
27 j=1 j=1 j=1 j=1
28
29 where ceil stands for the ceiling function, which rounds up to the nearest integer of ηE.
30 For element coarsening, the rule for element marking is problem-dependent. For example, in crack propagation
31 analysis, elements marked for coarsening are those with large crack opening [30].
32 Remark 1: To obtain a well-conditioned stiffness matrix Kh in Equation (23) the elements adjacent to marked
33 elements must be from the same, or at most from two consecutive hierarchy levels [17].
34
35
5.2. Refinement procedure
36
37 With the data structure in Section 4.1 at hand, we proceed to adaptive hierarchical refinement. Two sets of logical
38 vectors are defined to indicate the state of elements – active or inactive – at each hierarchy level. They are initialised
39 as false:
40
41 (1) Ea : indicator of active elements. Eia = {true : element i is active}.
42
43 (2) Eac : indicator of active child elements. Eiac = {true : child elements of element i are active}.
44
45 From Ea and Eac , three sets of logical vectors are obtained, which indicate the state of basis functions – active or
46 inactive – at each
n ohierarchy level. They are initialised as false. Here, we define all basis functions at P hierarchy
47 levels as: N = N i , (i = 1, 2, · · · , nbT ). nbT being the total number of basis functions at P hierarchy levels:
48
(1) Aa : indicator
n of basis functiono in the hierarchical basis function space H or HT , Equations (20) and (21).
49
50 Aia = true : N i ∈ H or HT .
51 n o
52 (2) A− : indicator of basis function in set A− , Equation (19). Ai− = true : N i ∈ A− .
53 n o
54 (3) A+ : indicator of basis function in set A+ , Equation (19). Ai+ = true : N i ∈ A+ .
55
A pseudo code to obtain Aa , A− and A+ can be found in [17]. The procedure for adaptive hierarchical refinement is
56
given in Algorithm 1.
57
58
59 12
60
61
62
63
64
65
 1
2
3
4 Algorithm 1 Adaptive hierarchical refinement
5
6  
7 A1 Read the geometry data to obtain the initial local knot vector Ξ10 , Ξ20 and the initial control points P0 .
8
   
9 A2 Carry out successive cell subdivision to generate Ξ1I , Ξ2I and PI for each hierarchy level I from Ξ10 , Ξ20 and
10 P0 .
11
12
A3 Compute the subdivision operator Sl,l+1 between two consecutive hierarchy levels l and l + 1.
13
14 A4 Obtain the list of active elements and active child elements. For the first load step, the active elements are
15 provided by initial T-spline mesh.
16
17
A5 Compute the logic vectors Aa , A+ , A− and the subdivision operator Mh .
18
19
20 A6 Solve Equation (23) and employ Equation (27) to obtain the displacement vector U.
21
22 A7 Check if elements should be refined or coarsened and mark them accordingly, see Algorithm 2. If there is no
23 marked element for refinement or coarsening, stop the calculation for current load step and go to next load step.
24 Otherwise obtain new list of active elements and active child elements on the basis of marked elements and
25 return to A4.
26
27
28 Algorithm 2 Element refinement and coarsening in adaptive hierarchical refinement.
29
30 B1 Compute the element error εQ of each active element. If coarsening is required, compute the corresponding
31 error indicators.
32
33
34 B2 Mark elements for refinement and coarsening.
35
36 B3 Refine the elements in order to obtain new list of active elements and active child elements. Here, the elements
37 to be refined are represented as Er and all child elements of Er as Erc .
38
39 B3.1 Get the old list of Ea and Eac .
40 B3.2 Set Ea (Er ) = false, Eac (Er ) = true.
41 B3.3 Set Ea (Erc ) = true, Eac (Erc ) = false.
42
43
44 B4 Coarsen the elements in order to obtain the updated list of active elements and active child elements, Ea and
45 Eac . Elements to be coarsened are denoted as Ec , the parent elements of Ec as E p , the parent elements of E p as
46 Egp , and all child elements of E p as E Ac .
47
48 B4.1 Get the list of Ea and Eac after element refinement.
49 B4.2 Set Ea (E p ) = true, Eac (Egp ) = true.
50
51 B4.3 Set Ea (E Ac ) = false, Eac (E p ) = false.
52
53
54
55
56
57
58
59 13
60
61
62
63
64
65
 1
2
3 5.3. Update process
4
5 During mesh refinement and coarsening, elements are introduced in the set of active elements. For non-linear
6 problems, it is then necessary to transfer the displacement from previous time step t to provide initial values for newly
7 activated elements at t + ∆t. The displacements that result from a transfer from coarse elements to fine elements are
8 exact. However, in the reverse process, some information is generally lost.
9 When the displacement vector t U is mapped to generate a new initial state vector during mesh refinement, t+∆t 0 U,
10 the process is exact:
 T
11 t+∆t l+1
= S̃l,l+1 t Ul
0 U (29)
12
13 where l is the hierarchy level, and S̃l,l+1 denotes the modified subdivision operator from Equation (12)
14  l,l+1
15 

 S IJ for N Jl+1 ∈ t+∆t Al+1 or t+∆t ATl+1
l,l+1
16 S̃ IJ =   (30)
0 else
17
18
with t+∆t A or t+∆t AT the space of hierarchical basis functions at t + ∆t.
19
For coarsening, a global least-square fit is employed to perform the mapping, and we minimise:
20
21 Z Z
ψ= t+∆t u − t u dΩ = t+∆t N t+∆t U − t u dΩ (31)
22 0 A 0
Ω Ω
23
24 with respect to t+∆t
0 U, yielding:
25
26 Z  T Z  T
t+∆t t+∆t t+∆t t+∆t t
27 MN 0 U =p with MN = NA NA dΩ; p= NA udΩ (32)
Ω Ω
28
29 where u contains the displacements, and t+∆t NA the basis functions associated with active elements at t + ∆t. MN can
30 be computed directly through the Gaussian quadrature on each active element at t + ∆t. However, the integration of p
31
needs to be modified as below:
32 Z  Z 
33 T T  
t+∆t t t+∆t t
34 p= NA udΩ = NA NA t UdΩ (33)
Ω Ω
35
36 where the second integration is carried out on each active element at t. t NA contains basis functions associated with
37 active elements at t.
38
39
40 6. Examples
41
42 We will now present four examples, aimed to validate the method in terms of accuracy and assess its performance.
43 Truncated hierarchical basis functions will be used throughout to describe the geometry of domain and to approximate
44 the field variable.
45 In the examples, only element refinement will be considered, no coarsening. The error of each element is computed
46 using the H 1 -norm and the energy norm kek [33], which are defined as:
47 sZ Z !
48
49 ku − ūkH 1 (Ωe ) = (u − ū)T · (u − ū) dS + (u − ū)′ T · (u − ū)′ dS
Ωe Ωe
50 s Z (34)
51
52 keke =
1
2 Ωe
σ̄)T ·
σ − σ̄
(σ C−1 · (σσ − σ̄σ̄) dS
53
 2
54
55 where u and σ stand for the analytical solution, ū and σ̄ denote the approximate solution, ∀u, ū ∈ H 1 (Ωe ) , and C
56 is the elastic stiffness tensor. (u − ū)′ is the first derivative of (u − ū) with respect to x and y, respectively. These error
57
58
59 14
60
61
62
63
64
65
 1
2
3 measures are employed for convergence studies, where mesh refinement is considered. The domain error is obtained
4 by summing up the element error.
5
6 r qX
X  2
7 ku − ūkH 1 (Ω) = ku − ūkH 1 (Ωe ) kek = kek2e (35)
e e
8
9 In marking elements for refinement and coarsening, the relative error of each element is employed to define the
10 element-wise error εe :
11
12 ku − ūkH 1 (Ωe ) keke
εe = qR R  εe = q R (36)
C−1 · σ dS
13 1
14 Ωe
uT · u dS + Ω u′ T · u′ dS 2 Ωe σT ·
e
15
16 In general, T-spline meshes are generated by adaptive local refinement of NURBS meshes [28]. To provide a good
17 illustration of hierarchical T-spines, we will directly give the initial mesh of the first two examples as a T-mesh with
18 T-junctions. For the rest examples, the initial mesh will be defined by quadratic NURBS surface, since it is a special
19 case of T-spline mesh. For all examples the geometry is modelled with same polynomial degree p in both parametric
20 directions.
21
22
23
24
25
26
27
28
29
30
31
32
33 Figure 8: Cantilever beam subjected to a parabolic traction: problem definition.
34
35
36
37 6.1. Cantilever beam under a parabolic traction
38 First, we will test the method to construct a T-spline mesh T α from an initial T-spline mesh T 0 , see Sections 3.1
39 and 4. A cantilever beam of length l and height h is considered, see Figure 8. It is fixed at x1 = 0 and subjected to a
40 parabolic traction at the free end x1 = l with a resultant force P. The geometry and material parameters read: length
41 l = 50 m, height h = 10 m, Young’s module E = 100 Pa, Poisson’s ratio ν = 0.3, and P = 1 N. The analytical solution
42 of the displacement field can be found in [34].
43 The domain is discretised by cubic T-splines. The initial T-spline mesh T 0 is given in Figure 9, which shows the
44 index domain, the parameter domain and the physical domain. A hierarchy of 3 levels is constructed from T 0 . The
45
T-spline mesh T 2 is shown in Figure 10(a).
46
In the analysis we apply Dirichlet boundary conditions at x1 = 0 and the traction t̄ = [0, σ12 ]T at x1 = L. For
47
48 the Dirichlet boundary condition, we could not directly impose the exact displacement at control points due  to the
49 non-interpolatory nature of T-splines. They are computed from the displacement and the parameter value ξ1 , ξ2 at
50 Gauss integration points.
51 For the computation of stiffness matrix K and right-hand force vector F, three-point Gaussian quadrature is em-
52 ployed to avoid shear locking effect in the simulation. Since we utilize cubic T-splines, the numerical solution should
53 be exact up to machine precision. The stress components σ11 and σ12 are given in Figures 11. The error in the stress
54 value, defined as the difference between numerical solution and analytical solution, is shown. The values are indeed
55 around the machine precision, which implies the proposed method of constructing hierarchical T-meshes works well.
56
57
58
59 15
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22 (a) T-spline mesh in the index and the parameter domain (b) T-spline mesh in the physical domain
23 0
24 Figure 9: Initial T-spline
 mesh  T for the cantilever beam. Here, the beam is given in the index domain (i, j), in
1 2
25 the parameter domain ξ , ξ , and in the physical domain (x1 , x2 ). The anchors are indicated by circular dots. The
26 polynomial degree of T-splines is p = 3 and the weights of anchors are w = 1.
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
Figure 10: T-spline mesh T 2 and the displacement distribution for the cantilever beam. (a) T-spline mesh T 2 in the
51
index and the parameter domains. The anchors are indicated by circular dots. Black denotes the mesh and the anchors
52
53 of initial T-spline mesh, T 0 , red stands for the mesh and the anchors generated for T 1 , and green represents the mesh
54 and the anchors generated for T 2 ; (b) contour plot for u1 ; (c) contour plot for u2 .
55
56
57
58
59 16
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10 (a) σ11 contour plot (b) error in σ11
11
12
13
14
15
16
17
18
(c) σ12 contour plot (d) error in σ12
19
20 Figure 11: Stress distribution for the cantilever beam.
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39 (a) problem setting (b) energy norm
40
41 Figure 12: Bending of a curved beam by a force at the end.
42
43
44 6.2. Bending of a curved beam by a force at the end
45 A curved beam with a unit thickness is considered next, see Figure 12(a). The beam is subjected to shear tractions
46 at the free end x2 = 0, with resultant force P. The geometry parameters are: inner radius Rmin = 5 m and outer radius
47 Rmax = 10 m. A linear elastic material is considered with: Young’s module E = 100 kPa, Poisson’s ratio ν = 0.3,
48 while P = −1 N. The analytical solution of displacements in polar coordinate system can be found in [34].
49 The problem has been solved using cubic T-spline bases with an initial T-spline mesh T 0 , see Figure 13. A
50 hierarchy of four levels is constructed from T 0 . The Dirichlet boundary conditions are weakly imposed at the boundary
51 x1 = 0. The Neumann boundary condition t̄ = [σ12 , 0]T is applied at x2 = 0 in an exact manner.
52 During adaptive refinement, elements are selected for refinement using Dörfler marking with η = 0.4. The conver-
53 gence in the energy norm, kek, is shown in Figure 12(b) for cubic NURBS and cubic T-splines. For uniform refinement,
54 the convergence rate will approach the optimal rate of convergence k = −p/2 = −1.5. The error level of the solution
55
for T-spline bases is higher than that for NURBS bases. A possible explanation is that the pre-defined T-splines do
56
not efficiently model the problem localisation. For the case of adaptive refinement, the convergence plot exhibits a
57
58
59 17
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27 (a) T-spline mesh in the index and parameter domain (b) T-spline mesh in the physical domain
28 0
29   mesh T for the curved beam. The beam is shown in the index domain (i, j), in the
Figure 13: Initial T-spline
30 parameter domain ξ1 , ξ2 , and in the physical domain (x1 , x2 ). The anchors are indicated by circular dots. The
31 polynomial degree of T-splines is p = 3. The local knot vectors and the coordinates of all anchors are given in
32 Appendix A.
33
34
35 zig-zag behaviour with a convergence rate which approximates the optimal rate k = −p/2 = −1.5. In general, the
36 error level for adaptive mesh refinement is between those of uniform T-spline and NURBS mesh refinement.
37 Figure 14 gives the contour plot of exact solution and errors in the stress component σ22 obtained numerically. In
38 the figure, Bézier meshes at refinement steps s = 5, 9, 14 are indicated by solid lines. They show a local refinement of
39 the mesh, giving an improved resolution of stress gradients, especially in the part with T-junctions in initial T-spline
40 mesh T 0 .
41
42 6.3. Infinite plate with a circular hole
43
The third example considers the adaptive hierarchical refinement using only NURBS. We have employed the cell
44
45 subdivision approach to construct hierarchical basis functions, see Sections 3 and 4. An infinite plate with a circular
46 hole is considered, with a radius R = 1 m, see Figure 15. The material properties are: Young’s modulus E = 100 N/m2 ,
47 Poisson’s ratio ν = 0.0 and thickness h = 1 m. The exact solutions of radial and tangential displacement are:
48 " # " #
T x rcos (2θ) R2 R4 T xr R2
49 ur = (1 + ν) + 4 2 − (1 + ν) 4 + (1 − ν) + (1 + ν) 2
50 2E r r 2E r
" 2 4
# (37)
51 T x rsin (2θ) R R
52 uθ = − (1 + ν) + 2 (1 − ν) 2 + (1 + ν) 4
2E r r
53
54
55
56
57
58
59 18
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17 (a) exact σ22 contour plot (b) error in σ22 at refinement step s = 5
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33 (c) error in σ22 at refinement step s = 9 (d) error in σ22 at refinement step s = 14
34
35 Figure 14: Stress distribution for the curved beam. Bézier meshes of active elements at each refinement step are
36 indicated by solid lines.
37
38
39
40
41
42
43
44
45
46
47
48
49 Figure 15: Infinite plate with a circular hole.
50
51
52 where θ is the azimuthal coordinate. The corresponding stress components read:
53 ! !
Tx R2 T x cos2θ 3R4 4R2
54 σr = 1− 2 + − + 1
55 2 r 2 r4 r2
! !
56 Tx R2 T x cos2θ 3R4
σθ = 1+ 2 − + 1 (38)
57 2 r 2 r4
58 !
T x sin2θ 3R4 19 2R2
59 σrθ = 4
− 2 −1
60 2 r r
61
62
63
64
65
 1
2
3 By virtue of symmetry only one quarter of the plate has to be modelled. Two different geometries are used to
4 represent the problem: a quarter of an annulus and a finite quarter plate. The exact traction from analytical solution is
5
imposed at the free boundary, see e.g. [35, 36].
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20 (a) problem definition of the quarter of an annulus (b) Bézier physical mesh and control points
21
22 Figure 16: Infinite plate with a circular hole: a quarter of an annulus
√ representation and initial T-spline mesh T 0 in the
23 physical domain. The weights of control points 1, 2 and 3 are 2/2; the weights of other control points are 1.
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41 (a) exact σ11 contour plot (b) H 1 norm
42
43 Figure 17: Quarter of an annulus representation: exact solution for σ11 and H 1 norm.
44
45
46 6.3.1. Quarter of an annulus representation
47 The geometry and boundary conditions of the annulus are shown in Figure 16a. The domain Ω is initially discre-
48 tised by NURBS with a polynomial degree p = 2, with knot vectors Ξ1 = [0, 0, 0, 1, 1, 1] and Ξ2 = [0, 0, 0, 1, 1, 1].
49 The coordinates of control points P are given in Figure 16(b). In this example, we consider NURBS with polynomial
50 degrees p = 2 and p = 4. For the fourth-order NURBS, the knot vector and the control points are updated by order
51 elevation from second-order NURBS. The NURBS discretizations for p = 2, 4 are taken as initial T-spline mesh T 0 .
52 In the analysis, a hierarchy of six levels is constructed from initial mesh T 0 . In adaptive refinement, elements
53 chosen for refinement are decided by quantile marking with η = 0.2. The convergence in the H 1 norm is shown in
54 Figure 17(b). Uniform mesh refinement can achieve an optimal convergence rate k = −p/2. Local mesh refinement
55
can reduce the error, but not the convergence rate. The reduced error level may be due to the effect of the resolution of
56
stress gradient by local mesh refinement. The exact solution of σ11 points to a stress concentration at (x1 , x2 ) = (0, 1),
57
58
59 20
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17 (a) error in σ11 at refinement step s = 1 (b) error in σ11 at refinement step s = 4
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32 (c) error in σ11 at refinement step s = 10 (d) error in σ11 at refinement step s = 15
33
34 Figure 18: Bézier meshes and error in σ11 for active elements at each refinement step for fourth-order NURBS bases.
35 The Bézier meshes of active elements are indicated by solid lines.
36
37
38 see Figure 17(a). Figure 18 presents the error of σ11 at each refinement step for the fourth-order NURBS. The figures
39 show a local refinement of the mesh around the hole, where smooth stress gradient is achieved.
40
41 6.3.2. Finite quarter plate representation
42 For this case the geometry and boundary conditions are shown in Figure 16(a). The domain Ω is discretised
43 by NURBS with a polynomial degree p = 2, with knot vectors Ξ1 = [0, 0, 0, 81 , 41 , 38 , 12 , 85 , 34 , 87 , 1, 1, 1] and
44 Ξ2 = [0, 0, 0, 41 , 12 , 34 , 1, 1, 1]. These knot vectors are obtained through h-refinement of open knot vectors, see [1].
45 Accordingly, the number of control points P is adapted. The corresponding physical mesh and anchors are shown
46 in Figure 16(b). Herein, we consider second-order as well as fourth-order NURBS bases to discretize the domain.
47 For the fourth-order NURBS bases, the knot vector and control points are obtained by order elevation
48  from second-
order NURBS bases. For the construction of initial T-spline mesh, T , the initial local knot vectors Ξ10 , Ξ20 and the
0
49
50 coordinates of anchors can be derived in a straightforward manner from Ξ1 , Ξ2 and P.
51 We consider a hierarchy of four levels to construct hierarchical basis functions. It is constructed from initial
52 T-spline mesh T 0 . Elements are refined by adaptive hierarchical refinement with quantile marking (η = 0.2). The
53 convergence plot of the H 1 norm is shown in Figure 20(b). As expected, the optimal convergence rate k = −p/2 is
54 achieved for uniform and for adaptive mesh refinement. The error level for adaptive mesh refinement is lower than
55 that for uniform mesh refinement. This is due to the fact that adaptive mesh refinement smoothens the stress gradient.
56 The phenomenon of smoothing the stress gradient is illustrated in Figure 21. The mesh around the hole with stress
57 concentration is refined gradually until the lowest hierarchy level. With adaptive mesh refinement, the error level is
58
59 21
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20 (a) problem definition for the plate (b) Bézier mesh and anchors
21
22 Figure 19: Infinite plate with circular hole: problem definition and initial T-spline mesh T 0 in the physical domain.
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40 (a) exact σ11 contour plot (b) H 1 norm
41
42 Figure 20: Representation of a finite quarter of a plate: exact solution of σ11 and H 1 norm.
43
44
45 reduced for the whole domain, which indicates that adaptive mesh refinement could not only efficiently model the
46 localisation, but also improve the accuracy globally.
47
48 6.4. Poisson problem on an L-shaped domain
49
As final example we consider a Poisson problem, which solves for the temperature u on an L-shaped domain, Fig-
50
ure 22(a). The definition of the L-shaped domain and the analytical solution of the problem can be found in [17]. The
51
52 L-shaped domain has been modelled by a single C 1 continuous quadratic B-spline patch, Figure 22(b). To define the
53 patch, the knot vectors are given as Ξ1 = [0, 0, 0, 18 , 14 , 83 , 12 , 58 , 43 , 78 , 1, 1, 1] and Ξ2 = [0, 0, 0, 41 , 21 , 34 , 1, 1, 1].
54 The corresponding Bézier physical  mesh and anchors are shown in Figure 22(b). For the construction of initial T-spline
55 mesh, T , the local knot vectors Ξ10 , Ξ20 are derived from Ξ1 and Ξ2 . A hierarchy of five levels is constructed from
0

56 the initial mesh T 0 , which constitutes the hierarchical T-spline basis function space. During the adaptive refinement,
57 elements are selected for refinement using quantile marking with η = 0.2.
58
59 22
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17 (a) error in σ11 at refinement step s = 1 (b) error in σ11 at refinement step s = 4
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33 (c) error in σ11 at refinement step s = 8 (d) error in σ11 at refinement step s = 10
34
35 Figure 21: Bézier meshes and the error in σ11 for active elements at each refinement step for second-order NURBS.
36 The Bézier meshes of active elements are indicated by solid lines.
37
38
39 Due to the singularity at the re-entrant corner (x1 , x2 ) = (0, 0), the rate of convergence k of the H 1 norm with
40 respect to the total number of degrees of freedom is given as:
41 ! !
1 π 1 2 1
42 k = − min p, = − min p, =− (39)
43 2 2π − β 2 3 3
44 Figure 23(b) presents a comparision of the convergence of adaptive mesh refinement and uniform mesh refinement.
45 For uniform refinement, the corresponding rate of convergence is k = −1/3, Figure 23(b). The optimal convergence
46
rate k = −1 is recovered by adaptive refinement, see Figure 23(b). It shows that the error level for adaptive refinement
47
is smaller than that for uniform refinement. This is because adaptive refinement better captures the gradient around
48
49 the re-entrant corner, Figure 24. From these figures it is observed that the mesh around the re-entrant corner is refined
50 gradually until the lowest hierarchy level.
51
52 7. Concluding remarks
53
54 We have developed a Bézier extraction framework for hierarchical T-splines, treating standard hierarchical T-
55 splines and truncated hierarchical T-splines in a unified and straightforward manner. Explicit basis function operations
56 are avoided, and only matrix manipulations have to be carried out. In particular, no explicit truncation of basis
57 functions is needed for the application of truncated hierarchical T-splines. The use of an element-wise point of view
58
59 23
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
(a) problem definition of the plate (b) Bézier physical mesh and anchors
22
23 Figure 22: Poisson problem on an L-shaped domain: problem definition and initial quadratic T-spline mesh in the
24 physical domain.
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
(a) exact u contour plot (b) H 1 error norm
42
43 Figure 23: Poisson problem on an L-shaped domain: exact solution of u and H 1 error norm.
44
45
46 facilitates the implementation in existing finite element codes. Moreover, adaptive refinement can be incorporated in
47 the analysis directly.
48 Algorithmically, a multi-level T-spline mesh is generated by successive cell subdivisions of an initial coarse T-
49 spline mesh. Subsequently, on each hierarchy level, Bézier extraction is applied to obtain the stiffness matrix. Initially,
50 this is done without consideration of any multi-level interaction, and this interaction is then enforced by a subdivision
51
operator. It has been detailed how the algorithms can be implemented. Numerical examples illustrate the accuracy of
52
the proposed method. Optimal convergence rates are obtained for all cases. However, the error level for adaptive mesh
53
54 refinement is generally lower than that for uniform mesh refinement, as to be expected. Upon local mesh refinement,
55 the error in the stress is reduced in the entire domain, which indicates that adaptive mesh refinement also improves
56 the accuracy globally.
57
58
59 24
60
61
62
63
64
65
 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17 (a) error in u at refinement step s = 10 (b) error in u at refinement step s = 22
18
19 Figure 24: Bézier meshes and error in u at each refinement step for quadratic T-spline bases. The error is given as the
20 difference between numerical solution and exact solution.
21
22
23 Acknowledgement
24
25 Financial support from the European Research Council (ERC Advanced Grant 664734 PoroFrac) is gratefully
26 acknowledged.
27
28
29 References
30 [1] T. J. R. Hughes, J. A. Cottrell, Y. Bazilevs, Isogeometric analysis: CAD, finite elements, NURBS, exact geometry and mesh refinement,
31 Computer Methods in Applied Mechanics and Engineering 194 (2005) 4135–4195.
32 [2] J. A. Cottrell, T. J. R. Hughes, Y. Bazilevs, Isogeometric analysis: toward integration of CAD and FEA, John Wiley & Sons, Chichester,
33 2009.
34 [3] T. W. Sederberg, J. Zheng, A. Bakenov, A. Nasri, T-splines and T-NURCCs, ACM Transactions on Graphics 22 (2003) 477–484.
[4] T. W. Sederberg, D. L. Cardon, G. T. Finnigan, N. S. North, J. Zheng, T. Lyche, T-spline simplification and local refinement, ACM Transac-
35 tions on Graphics 23 (2004) 276–283.
36 [5] M. A. Scott, X. Li, T. W. Sederberg, T. J. R. Hughes, Local refinement of analysis-suitable T-splines, Computer Methods in Applied Mechan-
37 ics and Engineering 213 (2012) 206–222.
38 [6] E. J. Evans, M. A. Scott, X. Li, D. C. Thomas, Hierarchical T-splines: Analysis-suitability, Bézier extraction, and application as an adaptive
39 basis for isogeometric analysis, Computer Methods in Applied Mechanics and Engineering 284 (2015) 1–20.
[7] X. Wei, Y. Zhang, L. Liu, T. J. R. Hughes, Truncated t-splines: Fundamentals and methods, Computer Methods in Applied Mechanics and
40 Engineering 316 (2017) 349–372.
41 [8] D. D′ Angella, S. Kollmannsberger, E. Rank, A. Reali, Multi-level bézier extraction for hierarchical local refinement of isogeometric analysis,
42 Computer Methods in Applied Mechanics and Engineering 328 (2018) 147–174.
43 [9] T. Dokken, T. Lyche, K. F. Pettersen, Polynomial splines over locally refined box-partitions, Computer Aided Geometric Design 30 (2013)
44 331–356.
[10] A. Bressan, Some properties of LR-splines, Computer Aided Geometric Design 30 (2013) 778–794.
45 [11] K. A. Johannessen, T. Kvamsdal, T. Dokken, Isogeometric analysis using LR B-splines, Computer Methods in Applied Mechanics and
46 Engineering 269 (2014) 471–514.
47 [12] L. Chen, R. de Borst, Locally refined t-splines, International Journal for Numerical Methods in Engineering (2018) DOI: 10.1002/nme.5759.
48 [13] A. V. Vuong, C. Giannelli, B. Jüttler, B. Simeon, A hierarchical approach to adaptive local refinement in isogeometric analysis, Computer
49 Methods in Applied Mechanics and Engineering 200 (2011) 3554–3567.
[14] C. Giannelli, B. Jüttler, H. Speleers, THB-splines: The truncated basis for hierarchical splines, Computer Aided Geometric Design 29 (2012)
50
485–498.
51 [15] P. Bornemann, F. Cirak, A subdivision-based implementation of the hierarchical b-spline finite element method, Computer Methods in
52 Applied Mechanics and Engineering 253 (2013) 584–598.
53 [16] A. Buffa, C. Giannelli, Adaptive isogeometric methods with hierarchical splines: error estimator and convergence, Mathematical Models and
54 Methods in Applied Sciences 26 (2016) 1–25.
[17] P. Hennig, S. Müller, M. Kästner, Bézier extraction and adaptive refinement of truncated hierarchical NURBS, Computer Methods in Applied
55
Mechanics and Engineering 305 (2016) 316–339.
56 [18] E. M. Garau, R. Vázquez, Algorithms for the implementation of adaptive isogeometric methods using hierarchical b-splines, Applied Numer-
57 ical Mathematics 123 (2018) 58–87.
58
59 25
60
61
62
63
64
65
 1
2
3 [19] X. Li, J. Deng, F. Chen, Surface modeling with polynomial splines over hierarchical T-meshes, The Visual Computer 23 (2007) 1027–1033.
4 [20] J. Deng, F. Chen, X. Li, C. Hu, W. Tong, Z. Yang, Y. Feng, Polynomial splines over hierarchical T-meshes, Graphical Models 70 (2008)
5 76–86.
6 [21] N. Nguyen-Thanh, H. Nguyen-Xuan, S. P. A. Bordas, T. Rabczuk, Isogeometric analysis using polynomial splines over hierarchical T-meshes
7 for two-dimensional elastic solids, Computer Methods in Applied Mechanics and Engineering 200 (2011) 1892–1908.
[22] R. de Borst, L. Chen, The role of beźier extraction in adaptive isogeometric analysis: Local refinement and hierarchical refinement, Interna-
8 tional Journal for Numerical Methods in Engineering 113 (6) (2018) 999–1019.
9 [23] Y. Bazilevs, V. M. Calo, J. A. Cottrell, J. A. Evans, T. J. R. Hughes, S. Lipton, M. A. Scott, T. W. Sederberg, Isogeometric analysis using
10 T-splines, Computer Methods in Applied Mechanics and Engineering 199 (2010) 229–263.
11 [24] M. A. Scott, M. J. Borden, C. V. Verhoosel, T. W. Sederberg, T. J. Hughes, Isogeometric finite element data structures based on Bézier
12 extraction of T-splines, International Journal for Numerical Methods in Engineering 88 (2011) 126–156.
[25] X. Li, J. Zheng, T. W. Sederberg, T. J. R. Hughes, M. A. Scott, On linear independence of T-spline blending functions, Computer Aided
13 Geometric Design 29 (2012) 63–76.
14 [26] S. May, J. Vignollet, R. de Borst, The role of the Bézier extraction operator for T-splines of arbitrary degree: linear dependencies, partition
15 of unity property, nesting behaviour and local refinement, International Journal for Numerical Methods in Engineering 103 (2015) 547–581.
16 [27] P. Morgenstern, D. Peterseim, Analysis-suitable adaptive T-mesh refinement with linear complexity, Computer Aided Geometric Design 34
17 (2015) 50–66.
[28] M. R. Dörfel, B. Jüttler, B. Simeon, Adaptive isogeometric analysis by local h-refinement with T-splines, Computer Methods in Applied
18 Mechanics and Engineering 199 (2010) 264–275.
19 [29] P. Hennig, M. Kästner, P. Morgenstern, D. Peterseim, Adaptive mesh refinement strategies in isogeometric analysis – A computational
20 comparison, Computer Methods in Applied Mechanics and Engineering 316 (2017) 424–448.
21 [30] L. Chen, E. J. Lingen, R. de Borst, Adaptive hierarchical refinement of nurbs in cohesive fracture analysis, International Journal for Numerical
22 Methods in Engineering 112 (13) (2017) 2151–2173.
[31] G. Beer, B. Marussig, J. Zechner, C. Dünser, T.-P. Fries, Boundary element analysis with trimmed nurbs and a generalized iga approach,
23 arXiv preprint arXiv:1406.3499.
24 [32] B. Marussig, J. Zechner, G. Beer, T.-P. Fries, Fast isogeometric boundary element method based on independent field approximation, Com-
25 puter Methods in Applied Mechanics and Engineering 284 (2015) 458–488.
26 [33] L. Chen, W. Dornisch, S. Klinkel, Hybrid collocation - Galerkin approach for the analysis of surface represented 3D-solids employing
27 SB-FEM, Computer Methods in Applied Mechanics and Engineering 295 (2015) 268–289.
[34] S. P. Timoshenko, J. N. Goodier, Theory of Elasticity, 2nd Edition, McGraw-Hill, New York, NY, 1951.
28 [35] S. Klinkel, L. Chen, W. Dornisch, A NURBS based hybrid collocation – Galerkin method for the analysis of boundary represented solids,
29 Computer Methods in Applied Mechanics and Engineering 284 (2015) 689–711.
30 [36] L. Chen, B. Simeon, S. Klinkel, A NURBS based Galerkin approach for the analysis of solids in boundary representation, Computer Methods
31 in Applied Mechanics and Engineering 305 (2016) 777–805.
32
33
34 Appendix A. Examples of anchors of T-spline meshes
35
36 Table A lists the local knot vector and the global coordinates of T-spline anchors in Figure 13.
37  
38 Table A: Local knot vector Ξ1 , Ξ2 , homogeneous coordinates (x1 w, x2 w) and weights (w) of T-spline anchors in
√
39 Figure 13; a = 2/2, b = Rmin , c = Rmax , d = (Rmin + 11Rmax )/12, e = (Rmin + 5Rmax )/6, f = (Rmin + 3Rmax )/4,
40 g = (Rmin + Rmax )/2, h = (3Rmin + Rmax )/4, m = (11Rmin + Rmax )/12, n = (5Rmin + Rmax )/6.
41
42
43 Anchor Ξ1 Ξ2 x1 w x2 w w
n o n o
44 1 ξ1 , ξ11 , ξ11 , ξ21 , ξ31 ξ2 , ξ12 , ξ12 , ξ22 , ξ32 0 −c 1
45 n 1 o n 1 o
2 ξ1 , ξ11 , ξ21 , ξ31 , ξ41 ξ2 , ξ12 , ξ12 , ξ22 , ξ32 ac
− (a+5)c (a+5)
46 n 1 o n 1 o 6
(10a+1)c
6 6

47 3 ξ1 , ξ21 , ξ31 , ξ41 , ξ51 ξ2 , ξ12 , ξ12 , ξ22 , ξ32 − (10a+13)c (5a+7)
n 1 o n 1 o 24 24 12
48 4 ξ1 , ξ31 , ξ41 , ξ51 , ξ61 ξ2 , ξ12 , ξ12 , ξ22 , ξ32 (14a+7)c
− (14a+3)c (7a+5)

49 n 2 o n 1 o 24
(a+2)c
24 12
(a+2)
5 ξ1 , ξ41 , ξ51 , ξ61 , ξ61 ξ2 , ξ12 , ξ12 , ξ22 , ξ32 − ac3
50 n 3 o n 1 o 3 3

51 6 ξ1 , ξ51 , ξ61 , ξ61 , ξ61 ξ2 , ξ12 , ξ12 , ξ22 , ξ32 c 0 1

n 4 o n 1 o
52 7 ξ1 , ξ11 , ξ11 , ξ21 , ξ31 ξ2 , ξ12 , ξ22 , ξ32 , ξ42 0 −d 1
53 n 1 o n 1 o
54 8 ξ1 , ξ11 , ξ21 , ξ31 , ξ41 ξ2 , ξ12 , ξ22 , ξ32 , ξ42 ad
− (a+5)d (a+5)
n 1 o n 1 o 6
(10a+1)d
6 6

55 9 ξ1 , ξ21 , ξ31 , ξ41 , ξ51 ξ2 , ξ12 , ξ22 , ξ32 , ξ42 − (10a+13)d (5a+7)

56 n 1 o n 1 o 24
(14a+7)d
24 12

10 ξ21 , ξ31 , ξ41 , ξ51 , ξ61 ξ12 , ξ12 , ξ22 , ξ32 , ξ42 24
− (14a+3)d
24
(7a+5)
12
57
58
59 26
60
61
62
63
64
65
 1
2
 
3 Table A: Local knot vector Ξ1 , Ξ2 , homogeneous coordinates (x1 w, x2 w) and weights (w) of T-spline anchors in
4 √
5 Figure 13; a = 2/2, b = Rmin , c = Rmax , d = (Rmin + 11Rmax )/12, e = (Rmin + 5Rmax )/6, f = (Rmin + 3Rmax )/4,
6 g = (Rmin + Rmax )/2, h = (3Rmin + Rmax )/4, m = (11Rmin + Rmax )/12, n = (5Rmin + Rmax )/6.
7
8 Anchor Ξ1 Ξ2 x1 w x2 w w
9 n o n o (a+2)e (a+2)
11 ξ1 , ξ41 , ξ51 , ξ61 , ξ61 ξ2 , ξ12 , ξ22 , ξ42 , ξ62 − ae3
10 n 3 o n 1 o 3 3

11 12 ξ1 , ξ51 , ξ61 , ξ61 , ξ61 ξ2 , ξ12 , ξ22 , ξ42 , ξ62 e 0 1

n 4 o n 1 o
12 13 ξ1 , ξ11 , ξ11 , ξ21 , ξ31 ξ2 , ξ22 , ξ32 , ξ42 , ξ52 0 −f 1
13 n 1 o n 1 o
14 ξ1 , ξ11 , ξ21 , ξ31 , ξ41 ξ2 , ξ22 , ξ32 , ξ42 , ξ52 af
− (a+5) f (a+5)
14 n 1 o n 1 o 6 6 6
(10a+1) f
15 15 ξ1 , ξ21 , ξ31 , ξ41 , ξ51 ξ2 , ξ22 , ξ32 , ξ42 , ξ52 − (10a+13) f (5a+7)
n 1 o n 1 o 24 24 12
16 16 ξ1 , ξ31 , ξ41 , ξ51 , ξ61 ξ2 , ξ22 , ξ32 , ξ42 , ξ52 (14a+7) f (14a+3) f
− 24 (7a+5)
17 n 2 o n 1 o 24 12

18 17 ξ1 , ξ11 , ξ11 , ξ21 , ξ31 ξ2 , ξ32 , ξ42 , ξ52 , ξ62 0 −g 1

n 1 o n 2 o
19 18 ξ1 , ξ11 , ξ21 , ξ31 , ξ41 ξ2 , ξ32 , ξ42 , ξ52 , ξ62 ag
− (a+5)g (a+5)

20 n 1 o n 2 o 6
(10a+1)g
6 6

19 ξ1 , ξ21 , ξ31 , ξ41 , ξ51 ξ2 , ξ32 , ξ42 , ξ52 , ξ62 − (10a+13)g (5a+7)
21 n 1 o n 2 o 24
(14a+7)g
24 12

22 20 ξ1 , ξ31 , ξ41 , ξ51 , ξ61 ξ2 , ξ32 , ξ42 , ξ52 , ξ62 − (14a+3)g (7a+5)
n 2 o n 2 o 24
(a+2)g
24 12
23 21 ξ1 , ξ41 , ξ51 , ξ61 , ξ61 ξ2 , ξ22 , ξ42 , ξ62 , ξ72 − ag (a+2)

24 n 3 o n 1 o 3 3 3

22 ξ1 , ξ51 , ξ61 , ξ61 , ξ61 ξ2 , ξ22 , ξ42 , ξ62 , ξ72 g 0 1

25 n 4 o n 1 o
26 23 ξ1 , ξ11 , ξ11 , ξ21 , ξ31 ξ2 , ξ42 , ξ52 , ξ62 , ξ72 0 −h 1
n 1 o n 3 o
27 24 ξ1 , ξ11 , ξ21 , ξ31 , ξ41 ξ2 , ξ42 , ξ52 , ξ62 , ξ72 ah
− (a+5)h (a+5)

28 n 1 o n 3 o 6
(10a+1)h
6 6

25 ξ1 , ξ21 , ξ31 , ξ41 , ξ51 ξ2 , ξ42 , ξ52 , ξ62 , ξ72 − (10a+13)h (5a+7)
29 n 1 o n 3 o 24
(14a+7)h
24 12

30 26 ξ1 , ξ31 , ξ41 , ξ51 , ξ61 ξ2 , ξ42 , ξ52 , ξ62 , ξ72 − (14a+3)h (7a+5)
n 2 o n 3 o 24 24 12
31 27 ξ1 , ξ11 , ξ11 , ξ21 , ξ31 ξ2 , ξ52 , ξ62 , ξ72 , ξ72 0 −m 1
32 n 1 o n 4 o
33 28 ξ1 , ξ11 , ξ21 , ξ31 , ξ41 ξ2 , ξ52 , ξ62 , ξ72 , ξ72 am
− (a+5)m (a+5)
n 1 o n 4 o 6
(10a+1)m
6 6

34 29 ξ1 , ξ21 , ξ31 , ξ41 , ξ51 ξ2 , ξ52 , ξ62 , ξ72 , ξ72 − (10a+13)m (5a+7)

35 n 1 o n 4 o 24
(14a+7)m
24 12

30 ξ1 , ξ31 , ξ41 , ξ51 , ξ61 ξ2 , ξ52 , ξ62 , ξ72 , ξ72 − (14a+3)m (7a+5)
36 n 2 o n 4 o 24
(a+2)m
24 12
(a+2)
37 31 ξ1 , ξ41 , ξ51 , ξ61 , ξ61 ξ2 , ξ42 , ξ62 , ξ72 , ξ72 − am
n 3 o n 2 o 3 3 3
38 32 ξ1 , ξ51 , ξ61 , ξ61 , ξ61 ξ2 , ξ42 , ξ62 , ξ72 , ξ72 m 0 1
39 n 4 o n 2 o
33 ξ1 , ξ11 , ξ11 , ξ21 , ξ31 ξ2 , ξ62 , ξ72 , ξ72 , ξ72 0 −n 1
40 n 1 o n 5 o
41 34 ξ1 , ξ11 , ξ21 , ξ31 , ξ41 ξ2 , ξ62 , ξ72 , ξ72 , ξ72 an
− (a+5)n (a+5)
n 1 o n 5 o 6
(10a+1)n
6 6
42 35 ξ1 , ξ21 , ξ31 , ξ41 , ξ51 ξ2 , ξ62 , ξ72 , ξ72 , ξ72 − (10a+13)n (5a+7)

43 n 1 o n 5 o 24
(14a+7)n
24 12

36 ξ1 , ξ31 , ξ41 , ξ51 , ξ61 ξ2 , ξ62 , ξ72 , ξ72 , ξ72 − (14a+3)n (7a+5)
44 n 2 o n 5 o 24
(a+2)n
24 12
(a+2)
45 37 ξ1 , ξ41 , ξ51 , ξ61 , ξ61 ξ2 , ξ62 , ξ72 , ξ72 , ξ72 − an
n 3 o n 4 o 3 3 3
46 38 ξ41 , ξ51 , ξ61 , ξ61 , ξ61 ξ42 , ξ62 , ξ72 , ξ72 , ξ72 n 0 1
47
48
49
50
51
52
53
54
55
56
57
58
59 27
60
61
62
63
64
65