WorksComputer Science
Counting and Sampling Manufacturable Stacking Sequences
Abstract
Designing the stacking sequence of a composite laminate is a discrete optimisation problem: each ply takes one of a few orientations, each evaluation may cost a finite-element solve, and the design must obey manufacturing rules and design guidelines, which optimisers enforce by penalty, repair or rejection. We describe the admissible set exactly. The two manufacturing rules forbid a finite set of factors, so the sequences they admit form a regular language, recognised by an automaton that tracks the last orientation and the length of its run; balance and the ten per cent rule constrain the Parikh vector and are carried by a dynamic program over automaton state and counters. One table, polynomial in N, counts the admissible sequences without producing them, draws them uniformly, and tells whether a partial sequence can still be completed. We give pseudocode, exact counts for three orientation sets, closed-form ceilings on the composition, and four of the five single-rule entries a published table left uncomputed. At 48 plies, 99.8% of the nodes of a layer-wise search tree checking the rules locally have no admissible completion on eight orientations; the completion test removes them. A one-point crossover that cuts two admissible parents only where they share an automaton state and their child closes the counters returns admissible children by construction.
Introduction
A composite laminate is a stack of plies, each a sheet of parallel fibres laid at one of a finite set of orientations, and its stacking sequence lists those orientations from one surface to the other. We borrow the vocabulary of formal languages: the allowed orientations form an alphabet , a stacking sequence is a word over it, one letter per ply, and a set of stacking sequences is a language. The orientations determine the in-plane and bending stiffness of the laminate, hence its buckling load, natural frequencies and deflection under load, and the design problem is to maximise one of these responses, or to minimise subject to targets on them. It is a discrete optimisation problem over , and each evaluation of the objective may cost a finite-element solve. It has been searched with genetic algorithms since the early 1990s [10, 39], with permutation encodings [40, 30], particle swarms [13], scatter search [49] and ant colonies [6], and with branch and bound: layer by layer from the outer surface inwards [34], in a modified form that starts from “ideal” layers and keeps one branch per level [12 Sec. 2], or in the space of lamination parameters, where the admissible laminates form a fractal set [56], non-symmetric laminates included [42]. It has been extended to blending across panels [52, 60], to lamination parameters as a continuous relaxation [41, 4], to variable-stiffness laminates [1, 50], to strength as well as buckling [6, 5], to multiple objectives [2], and to structures other than flat panels: cylindrical panels [2], laminated grid plates [20], a wind-turbine blade [3] and an aero-engine fan-blade-like structure [11].
Not every word is a laminate that can be built or certified. Two manufacturing rules are local: consecutive plies must not differ by more than of orientation, and no more than three consecutive plies may share one. Designers add guidelines that are not local, namely mirror symmetry, balance between and , and a minimum share of ten per cent for each principal direction; symmetry and balance are dropped where a directional stiffness or a coupling is wanted, as in aeroelastic tailoring [51]. Each of these methods spends part of its effort on candidates outside the set these rules define: a genetic algorithm penalises the rules it does not build into its encoding [39] or repairs them [55], a sampler builds candidates and rejects the non-compliant, a branch and bound checks each partial sequence against the rules and learns only at the leaves that a branch had no admissible completion. The rules enter as a filter: as design rules to respect [27, 33] or to relax [57], as constraints on the fibre paths of tow-steered plies [9 Sec. 5], or as a coupled ply-order and ply-drop problem [3]. We describe the admissible set exactly and derive from the description algorithms that produce admissible sequences, and only those.
The set treated here is that of a single panel: blending between adjacent panels, ply drops [27] and damage tolerance are not part of it (§ 7).
Throughout the paper the reference size is plies. It is the ply count of the classical benchmark of Le Riche and Haftka [39], and it lies far beyond the ply counts at which we enumerate exhaustively (§ 3.6). Nothing in the construction depends on it: the counts are given from 24 to 54 plies, and every other quantity at 48.
Zein et al. [61], whose constraint-programming method searches a backtracking enumeration tree pruned as soon as a prefix breaks a rule, with rules of the same kind (no jump, at most four identical consecutive plies, prescribed ply counts per orientation), write of it that “it does not give compute the number of admissible stacking sequences” [sic] [61 p. 57], and frame the cost in terms of enumerating or candidates [61 Sec. 4.2]. Irsarri [28] tabulates the number of realisable stacking sequences rule by rule at four, eight and sixteen plies, and leaves several cells marked as not computed. Fedon et al. [21] note that the number of admissible laminates grows exponentially with the ply count, which makes complete enumeration impossible; that is true of enumeration and does not follow for counting. A granted family assigned to Boeing builds a library of admissible sublaminates by growing an integer tree and pruning the branches that violate the rules [54], the rules being stated there as ranges of ply counts per orientation, a cap on consecutive plies of one orientation, and a list of sequences that are disallowed, a finite antidisctionary in all but the name; a companion patent bounds the ply counts of a laminate per orientation to decide feasibility before any optimisation is run [53]. That tree is the unfolding of the automaton used here, and it is stored: one node for every admissible prefix. Its specification puts the size of the unconstrained space at sequences for a ten-layer sublaminate over four orientations and does not report how many of them survive the rules. Recent work maps the same design space onto a quantum state, or onto a tensor network, and carries the manufacturing rules as penalty terms in an objective [59].
Our results
We show that the two manufacturing rules forbid a finite set of factors, so that the sequences they admit form a regular language, recognised by an automaton whose state is the last orientation together with the length of the run it terminates (§ 2), while balance and the 10% rule constrain the Parikh vector and are carried exactly by a dynamic program over automaton state and counters. Symmetry reduces a laminate to its half-stack, and the mirror leaves only a condition on the last run. The construction therefore covers symmetric and non-symmetric, balanced and unbalanced laminates, over any finite set of orientations. We treat sets of four, eight and twelve orientations.
Algorithms. One dynamic program, built once per alphabet and ply count, supports three operations (§ 3, Algorithms 1–3): it returns the number of admissible sequences without producing any; it draws an admissible sequence uniformly at random in , with no rejection; and it decides whether a partial sequence can still be completed, by a lookup after a walk of its length. A fourth operation, recombination of two admissible parents, returns an admissible child on every call, a fresh exact draw when no admissible cut exists (Algorithm 4).
Counts. We give exact counts for the three alphabets, for symmetric laminates at ply counts from 24 to 54, and for non-symmetric ones, balanced and unbalanced, verified against exhaustive enumeration (§ 3.6). Under the rules of [28] our methods reproduce five of the six rule columns of its table at four and eight plies, and fill four of the five single-rule entries it leaves uncomputed at sixteen (§ 3.7). At 48 plies the guidelines divide the symmetric language by 13 and by 832 on the four- and eight-orientation sets, so the size advantage of the larger set falls from 51 794 to 801 (Figure 1). We also give closed-form ceilings on the composition.
Figure 1. The admissible set inside all stacking sequences, at . Each bar counts the half-stacks of 24 plies that survive one restriction more than the bar above it. In the righthand column, the extended set is larger by a factor of 51 794 on the manufacturing language and of 801 once the guidelines are imposed. Exact counts, as in Table 1.
Tree search. At 48 plies, 90.8% of the nodes of a branch and bound that assigns plies from the surface inwards and checks the rules at each node have, on four orientations, no admissible completion, 99.8% on eight and more than 99.99% on twelve (§ 4). The completion test of the program removes them, and since two prefixes in the same cell have the same completions, the tree of completable prefixes collapses onto the cells: of them against prefixes on eight orientations.
Recombination. Two identical parents are in the same automaton state at every cut, so a population that has converged recombines only into copies of its members, and the diversity of a search inside the language has to come from outside the operator (§ 6). The exact sampler cannot fail and reports an empty language as empty (§ 5).
The techniques are classical and we claim none of them. Counting the words that avoid a finite set of factors is done by a transfer matrix [22 Ex. V.11] or by string correlations [25]; the forbidden-factor description we use, and the automaton that recognises the words avoiding a finite antidictionary, are those of [14 Sec. 2]; and transfer matrices have been used to enumerate and generate constrained configurations in other fields [36]. The sampler of Section 5 belongs to the line that generates combinatorial objects uniformly by counting them first [58, 23, 26, 29, 46], to which Boltzmann samplers are the standard alternative [19], and which meets constraint programming where a sequence is constrained by an automaton [48, 16] or by a Markov model [47]; conditioning it on prescribed letter counts is classical within that method as well [18, 17], and also done by holonomic methods [7]. What is new is that this set of industrial constraints is of that kind, a regular language conditioned on thresholds and signed differences of its letter counts rather than on a prescribed composition, together with a table that carries only what those conditions need.
The paper is organised as follows. In § 2 we define the language and the two kinds of constraint, and in § 3 we give the algorithms, the counts and the bounds, and verify the counts. In § 4 we compare the construction with branch and bound, in § 5 we sample uniformly, and in § 6 we give the recombination operator. § 7 discusses what is due to whom and the limitations, and § 8 concludes.
The language of admissible stacking sequences
Words, forbidden factors, and the manufacturing rules
A stacking sequence of plies is a word of length over a finite alphabet of ply orientations. We consider the conventional set, its extension by and , and a finer one,
angles being measured in degrees; is also called the extended alphabet.
Recall that a word is a factor of if for some words , and that a set of words is an antidictionary for the language , the words that avoid [14 Sec. 2], [15]. When is finite, is regular [14 Sec. 2].
The disorientation between two orientations is
and the two manufacturing rules read as follows (Figure 2).
Figure 2. A stacking sequence and the two manufacturing rules. (a) An eight-ply symmetric laminate in exploded view, each ply carrying its fibre direction; the mid-plane is the axis of the mirror. (b) The laminate of (a) as a column of plies, admissible, and two sequences that are not, one against (M1) and one against (M2). Colour encodes ; the sign is carried by the fibres and by the lean of the stroke.
(M1) Disorientation. No two consecutive plies differ by more than , that is, for every .
(M2) Contiguity. No more than three consecutive plies share the same orientation.
Proposition 2.1. Rules (M1) and (M2) forbid exactly the finite set
and the set of stacking sequences satisfying both rules is . In particular for (four forbidden pairs, namely , , and , and four forbidden quadruples), for (twenty-eight pairs and eight quadruples) and for (sixty pairs and twelve quadruples).
Proof. Both rules are stated on windows of fixed length, two consecutive plies for (M1) and four for (M2), so a sequence violates them if and only if it contains one of the corresponding windows as a factor. Enumerating the violating windows gives (2). The cardinalities follow by inspection of on .
Theorem 2.2. The set of stacking sequences satisfying (M1) and (M2) is a regular language, recognised by a finite automaton with states besides the initial one, in which the state reached is a function of the prefix read.
Proof. By Proposition 2.1 it suffices to recognise . Take
where is the last orientation read and the length of the run it terminates, with transitions
Every state is accepting, and the states reachable by a single ply are those with ; prefixing a state for the empty word, with , gives a deterministic automaton in the usual sense. A transition is undefined exactly when it would complete a factor of , by (2): the second case of (4) rules out the forbidden pairs and the first the quadruples. Hence the automaton accepts .
Figure 3 shows the automaton for , where and there are 32 transitions; on it has 24 states and 100 transitions, on 36 and 240. It is, up to its sink state, the automaton that [14 Sec. 2] build from a finite antidictionary, whose states are the proper prefixes of forbidden words; here those prefixes are the empty word and the runs of one, two and three identical plies. In Figure 3 the missing cross transitions are exactly the forbidden pairs, and , so that what remains is a cycle on the four orientations.
Figure 3. The automaton of the manufacturing rules on . A state is the pair (last orientation, run length). Horizontal arrows repeat the last ply and advance the run, which stops at three; returning arrows change orientation and reset the run to one; the three states of an orientation are drawn as one dashed box, and each returning arrow stands for three transitions. Double circles mark the states with run one; the initial state is not drawn.
The design guidelines as Parikh constraints
The remaining guidelines do not forbid factors: they constrain how many plies of each orientation the laminate contains. Let denote the Parikh vector of , whose entry counts the occurrences of in . For a laminate of plies they read
(balance)
(10% rule)
and, where it is required, mirror symmetry about the mid-plane. The constructions below take each guideline as an option.
Proposition 2.3. The words of all lengths over that satisfy (M1), (M2) and (5) do not form a regular language, and neither do those that satisfy (6) as well.
Proof. Intersect with the regular language . In a word of consecutive plies differ by and no orientation is repeated, so (M1) and (M2) hold, and , for some . Balance holds exactly when , and then each of the four directions is a quarter of the plies, so (6) holds too. The intersection is in both cases, which is not regular by the pumping lemma; since regular languages are closed under intersection, neither set is regular.
Since (5) and (6) depend on only through , they are carried exactly by a dynamic program over pairs (automaton state, partial Parikh vector), developed in § 3. The same construction accommodates any constraint that depends on the word only through its Parikh vector, given counters that carry what that constraint needs.
Symmetric and non-symmetric laminates
A symmetric laminate of plies is written , where is the half-stack of plies and its reversal.
Lemma 2.4. Let and . Then satisfies (M1) and (M2) if and only if does and ends with a run of length one.
Proof. The factors of that are not factors of or of are those straddling the mid-plane, where the last ply of is immediately followed by its own copy. Rule (M1) is satisfied there because . Rule (M2) is satisfied there if and only if the run ending , of length , does not become a run of length , that is if and only if . Reversal preserves both rules, since is closed under reversal.
Remark 2.5. For odd the mid-plane carries a single ply that its own reflection does not duplicate, so the run condition of Lemma 2.4 is replaced by one on the run that ply terminates. The Parikh vector becomes , where is the central ply, so balance forces and the thresholds depend on . We treat only even below.
Lemma 2.6. With we have , so (5) and (6) hold for if and only if for every and for every .
Proof. Immediate, since is a permutation of .
Lemma 2.4 turns symmetry into a choice of accepting states, those with run one (the double circles of Figure 3), and Lemma 2.6 halves the counters.
A non-symmetric laminate is read whole; “non-symmetric” means here that symmetry is not imposed, so the class contains the symmetric laminates, a share below of it at every ply count tabulated. Rules (M1) and (M2) apply to the word of plies, balance and the 10% rule to its Parikh vector, and every state is accepting. We write for the length of the word the automaton reads, for a symmetric laminate and otherwise, and for the smallest value of a counter of that word that satisfies the 10% rule, that is the least integer with in the symmetric case and otherwise. Everything that follows holds for both unless stated otherwise.
Definition 2.7. Fix and the guidelines imposed, a subset of symmetry, balance (5) and the 10% rule (6). A laminate is admissible if it satisfies (M1), (M2) and every guideline in , and compliant if it is admissible with balance and the 10% rule in , the setting of § 5 and § 6. A word read by the automaton, or , is admissible or compliant when its laminate is. A completion of a word of length is a word of length such that is admissible. With empty the admissible words are the manufacturing language .
The statements of § 3 hold for every .
Counting
The transfer matrix
The term is meant in its combinatorial sense [22 Ex. V.11], unrelated to the transfer-matrix method of laminate mechanics.
Let be the adjacency matrix of the automaton of Theorem 2.2, if for some and 0 otherwise; since the state reached determines the ply read, walks and words correspond one to one. By Lemma 2.4 the admissible half-stacks of plies are the walks of length that start and end in a state with run one, so their number is
For a non-symmetric laminate the count is . (7) is an exact integer computation whose cost grows with the number of digits of the count and not with the count, and produces no sequence.
The dynamic program on the Parikh counters
We work on cells , where is an automaton state and a vector of counters holding what the guidelines need to know of the Parikh vector read so far:
for each principal direction , the count , since beyond the 10% rule is indifferent;
for each pair with , when balance is imposed, the difference .
Reading a ply updates to a vector that depends on and only. Let be the least number of plies that, appended to a word with counters , satisfy every imposed guideline, ignoring the automaton: the deficits of the thresholded directions that belong to no pair, for the pairs without threshold, and for when both are imposed, which the capped counts and determine; a guideline that is not imposed contributes nothing, and with the 10% rule off . A cell is accepting when and, for a symmetric laminate, has run one (Lemma 2.4).
Lemma 3.1. Two words of the same length that satisfy (M1) and (M2) and reach the same cell have the same completions (Definition 2.7). Moreover for every , so a word of length whose cell has has no completion, and neither has any of its extensions.
Proof. By (4) the continuations readable after a prefix depend on it only through , and whether a completion satisfies the guidelines depends on the prefix only through , which updates without reference to anything else. A ply raises one entry of the Parikh vector by one, so it moves a single term of need, that of its direction or of its pair, and lowers it by at most one; the remaining plies can therefore close at most units of deficit.
The program has two passes (Algorithm 2). The forward pass generates the reachable cells level by level, discarding by need; the backward pass assigns to each cell , the number of admissible completions of any prefix of length that reaches it. The number of admissible words is the sum of over the first level. The threshold is computed in integer arithmetic, as the least integer with , or for a half-stack.
Theorem 3.2. Let be fixed, with principal directions and sign pairs. Algorithm 2 computes the number of admissible words of length exactly, in
arithmetic operations on integers of bits, hence in time polynomial in . After it, Algorithm 3 draws a uniformly distributed admissible word in operations and decides whether a prefix of length is completable in .
Algorithm 1 Automaton(): the manufacturing rules (M1)–(M2).
for and do
if and then
▷ extend the run
else if and then
▷ open a new run ▷ otherwise undefined: completes a factor of
return
Algorithm 2 Table(): completion counts on the cells.
Require: options: symmetric or not, balance on or off, 10% rule on or off
; if symmetric, else
for do ▷ forward: reachable cells
for and with defined do
if then
add to
for do
if and has run one or not symmetric, else
for do ▷ backward: completions
for do
▷ over defined ; absent cells count 0
return , and
Algorithm 3 Sample() and Completable().
if then
return none ▷ the empty set, certified
draw with probability ; the ply of
for do
draw with probability , where
; append to
return ▷ uniform on ; mirror it if symmetric
function Completable() ▷ a prefix of length
if then return
walk through from and from (Theorem 2.2); return false if a transition is undefined
return for the cell reached
Proof. Correctness: by Lemma 3.1 the prefixes that reach a cell share their completions, so is well defined, the recursion of the backward pass counts each admissible completion exactly once by its next ply, and the discarded cells contribute nothing. Each level holds at most states times capped counts times differences, and each cell has at most successors. A count never exceeds . The walk of Algorithm 3 chooses each ply with probability equal to the share of completions it keeps, so the product telescopes to for every admissible word.
The operations counted are additions and comparisons of integers of that size; a draw takes a uniform integer below , and the cells are kept in hash tables with constant expected access.
With the discarding by need, the largest level at on holds 213 202 cells. When the distribution over compositions is wanted, as below, the counters are carried uncapped, at a cost in cells.
Cost of the design guidelines
At the guidelines divide the language by 13 on and by 832 on (Table 1). On balance acts on three pairs instead of one, and the 10% rule demands that 0, and 90 all be present in a laminate whose plies may also be spent on and , which do not help to satisfy it.
| 24 | 376 704 | 102 102 368 | |||
| 40 | 1 988 997 120 | 17 885 340 482 800 | |||
| 44 | 16 954 798 592 | 365 900 128 074 196 | |||
| 48 | 144 527 708 160 | 7 485 622 309 145 552 | |||
| 54 | 3 596 986 742 784 | 692 669 292 246 653 848 | |||
| 24 | 29 560 | 182 020 | 782 348 | ||
| 40 | 175 964 268 | 39 709 914 088 | |||
| 44 | 1 302 502 076 | 398 548 747 372 | |||
| 48 | 11 236 189 644 | 8 999 374 621 960 | |||
| 54 | 277 909 460 780 | 881 013 334 451 080 |
Table 1. Exact number of admissible symmetric half-stacks of plies on the three alphabets: (a) under the manufacturing rules alone, (b) under the rules together with the design guidelines, with the ratios between alphabets. In the columns values of and above are rounded to three significant figures; the deposited code returns the integers.
Hence the size advantage of the extended alphabet falls from on the manufacturing language to under the guidelines. On the guidelines divide the language by , and its advantage over falls from to .
Non-symmetric and unbalanced laminates
An unbalanced laminate couples extension with shear and strengthens the coupling of bending with twisting, a non-symmetric one couples extension with bending [51]. Algorithm 2 counts every class with the corresponding options (§ 2.3); for the non-symmetric balanced class, which carries a difference for every sign pair over the whole length, we count by meeting in the middle.
Proposition 3.3. Let be even and let be the number of words of plies satisfying (M1)–(M2) that reach the cell , obtained by carrying these counts through the forward pass of Algorithm 2 for non-symmetric words of length . Then the number of admissible non-symmetric laminates of plies is
over the pairs of cells such that the junction is admissible (if and , then when and otherwise), every sign-pair difference of is the opposite of that of , and for every principal direction .
Proof. Write the laminate as with . The antidictionary (2) is closed under reversal, and the Parikh constraints do not depend on the order, so is admissible exactly when its reversal is an admissible prefix of the same length; the state reached by is the pair (first ply of , length of its first run). The factors of that are factors of neither nor straddle the junction; they break (M1) exactly when , and (M2) exactly when and the two runs together exceed three. The Parikh vector of is the sum of those of and , so balance holds iff the differences are opposite, and a threshold holds iff the sum of the counts reaches , which the capped counts decide. The pruning of the forward pass removes only prefixes that no completion of length can save (Lemma 3.1), hence none that the sum needs.
Proposition 3.3 needs only the middle level, one table serving both halves. With the counts of each half exact and the sum taken modulo four primes of 61 bits and rebuilt by the Chinese remainder theorem, it reproduces all 56 non-symmetric counts obtained with Algorithm 2, and reaches on with cells at the middle level (Table 2).
Table 2: Exact number of admissible laminates of plies under the manufacturing rules and the 10% rule, for the three classes other than symmetric and balanced (Table 1): symmetric unbalanced, non-symmetric balanced, non-symmetric unbalanced. Values of and above are rounded to three significant figures; the deposited code returns the integers.
| symmetric unbalanced | non-symmetric balanced | non-symmetric unbalanced | ||
| 24 | 147 112 | |||
| 40 | ||||
| 44 | ||||
| 48 | ||||
| 54 | ||||
| 24 | 4 128 968 | |||
| 40 | ||||
| 44 | ||||
| 48 | ||||
| 54 | ||||
| 24 | ||||
| 40 | ||||
| 44 | ||||
| 48 | ||||
| 54 |
Table 2.
At the balanced non-symmetric laminates outnumber the symmetric ones by on , on and on . Dropping balance multiplies the non-symmetric set by 16, 691 and 5 648. The symmetric unbalanced class is 10.0, 147 and 619 times the symmetric balanced one. Every algorithm of this paper applies to these sets wherever their full table fits in memory: sampling and the completion test use it as it is, and the crossover of Section 6 is used without the run-one condition that a symmetric laminate imposes at its mid-plane (Lemma 2.4). The balanced non-symmetric class on is counted beyond that point, from , by Proposition 3.3 alone, and is not sampled here.
Ceilings on the composition
Proposition 3.4. Let be the smallest half-stack counter satisfying the 10% rule, and let be a compliant half-stack of plies. Then
Proof. By Lemma 2.6 the 10% rule reads for each of the four directions , and balance reads . Fix and bound the plies that cannot be spent on it. If , which have no distinct partner, the other three thresholded directions consume at least , whence . If , balance forces , so the pair consumes , while 0 and 90 consume at least ; hence . If belongs to a sign pair that carries no threshold ( and , and on also and ), the pair again consumes while all four thresholded directions consume at least , hence .
At , where and , (8) gives . That bound is attained: the dynamic program finds 840 compliant half-stacks with exactly 15, on and on .
It is not tight in general. At on the bound gives , while exhaustive enumeration shows the true maximum is 5. The bound ignores adjacency: a half-stack rich in 0° must break its runs every three plies, the only orientations adjacent to 0° are , and the 90° ply that the 10% rule demands cannot sit next to a 0° one. The exact maximum needs the joint state of automaton and counters, which the dynamic program carries.
Against exhaustive enumeration for every orientation at on and on , (8) is attained everywhere except at on , where the maxima fall short of it also for 90 (5 against 7) and for (3 against 4).
On at the compliant language has mode and mean 4.42, and 0.00058% of it lies at , half the half-stack.
Verification of the counts
Internally, we compared the counts with exhaustive enumeration filtered by rule checkers written separately from the dynamic program, on totals and on the whole distribution by Table 3: Table II-1 of [28] completed. The columns at four and eight plies are the published values, which our methods reproduce exactly except in the 10% row; at sixteen plies the symmetry row is published and the other entries are the cells the table leaves uncomputed.
| rule | 4 plies | 8 plies | 16 plies | closed form |
| Sym. | 16 | 256 | 65 536 | , even |
| Eq. | 70 | 12 870 | 601 080 390 | |
| Group. | 228 | 47 088 | 2 010 036 708 | |
| Des. | 108 | 8 748 | 57 395 628 | |
| TolDom. | 144 | 36 864 | 2 415 919 104 | , |
| 10% | 192 | 62 720 | not reproducible: see text |
Table 3. Table II-1 of [28] completed.
composition. The nine cases are on and on , and all coincide. The ten manufacturing counts of Table 1(a) on and agree with the matrix powers of (7), computed in exact integer arithmetic. We checked the general program of Algorithm 2, with capped counters, separately: against exhaustive enumeration on 80 cases that cover the three alphabets, symmetric and non-symmetric words, balance on and off and the 10% rule on and off; against every and entry of Table 1; by 2000 draws of Algorithm 3 in each of three settings, all compliant; and by Completable on every one of the prefixes at on , with no disagreement.
Externally, Table II-1 of [28] tabulates the number of realisable stacking sequences on rule by rule at 4, 8 and 16 plies. We recover five of its six rule columns at both 4 and 8 plies, and with them the definitions he used. Two differences forbid a column-by-column comparison with our own counts. His rules are symmetry (Sym.), balance of (Eq.), contiguity (Group.), disorientation (Des., our (M1)), no ply on either outer surface (TolDom.) and a 10% rule. His contiguity rule allows two identical consecutive plies where ours allows three; and his 10% rule is not a counting rule but a membrane-stiffness criterion, chosen to remain applicable to arbitrary orientations and, as he notes, less restrictive for thin laminates.
Completion of a published table
At sixteen plies Table II-1 leaves five of its six single-rule entries uncomputed [28 p. 59]: only the symmetry entry, which it deduces from the eight-ply enumeration, is filled. Four cumulative entries are also uncomputed; each combines rules with the 10% criterion. Each of the five combinatorial rules is counted by a transfer matrix or in closed form, so we fill four of the five blank cells in the sixteen-ply column of Table 3; the fifth, the 10% row, rests on a reading of a membrane-stiffness criterion that the thesis does not spell out and we cannot reconstruct.
Every entry at sixteen plies comes from a computation that first returns the published values at four and eight plies; the four other than symmetry, which is a formula, also agree with an independent exhaustive enumeration at eight plies.
Four of the five combinatorial rows are closed forms, and the fifth, contiguity, satisfies a two-term linear recurrence,
which returns the published values at four and eight plies [45]. Balance, for its part, is the central binomial coefficient:
It is the coefficient identity , and the values are those of [44].
The criterion behind that row requires the extensional stiffnesses , and of classical lamination theory [31] to be at least those of a virtual quasi-isotropic laminate of 40% the thickness. With plies of one thickness it depends on the sequence only through its counts and enters the program as three linear inequalities on uncapped counters; as stated, it gives 224 and 64 128 admissible sequences at 4 and 8 plies, against the published 192 and 62 720, for any carbon-epoxy ply.
Branch and bound
A branch and bound for the stacking sequence [34] assigns plies one at a time from the outer surface towards the mid-plane, since the outer plies weigh most in bending. Each node of its tree is a prefix, discarded when a bound on the objective over its completions cannot beat the incumbent [43]. The manufacturing rules are local, so a node that breaks one is discarded on the spot. The guidelines are not, and a prefix that obeys the rules may have no admissible completion; the tree learns it only after generating and rejecting every leaf below, unless the bound happens to discard the branch first for another reason.
The prefixes of length that obey the rules number , by the transfer matrix of § 3. Those with at least one admissible completion number over the cells of level with , where counts the prefixes that reach a cell and is obtained by carrying counts through the forward pass of Algorithm 2. Their difference is the part of a locally checked tree that leads nowhere. Table 4 gives it for the three alphabets; the values are exact and agree with an explicit tree at on , on and on a non-symmetric case.
Table 4: A layer-wise tree over symmetric half-stacks against the cells of Algorithm 2, rules and guidelines imposed; sizes summed over the depths . “Local” keeps every prefix that obeys the manufacturing rules, “exact” those with an admissible completion; “dead” is the share of the first not in the second; “cells” counts the cells with .
| admissible | tree, local | tree, exact | dead | cells | tree / cells | ||
| 24 | 29 560 | 836 388 | 85 556 | 89.8% | 1 650 | 52 | |
| 48 | 90.8% | 21 370 | |||||
| 24 | 182 020 | 614 124 | 99.63% | 8 726 | 70 | ||
| 48 | 99.76% | 1 257 774 | |||||
| 24 | 782 348 | 99.990% | 21 574 | 125 | |||
| 48 | 99.998% |
Table 4.
At on the dead share is zero at depth 6, 2% at depth 9, 14% at 12, 75% at 18 and 97% at 21 of 24: the dead branches sit near the mid-plane, where the tree is widest.
Three things follow.
First, counting with a tree means reaching every admissible leaf. By Lemma 3.1 a tree that never enters a dead branch is the unfolding of the cells: two prefixes in the same cell root identical subtrees, so the program keeps one copy of each. The tree grows exponentially with the depth, the cells polynomially (Theorem 3.2).
Second, a branch and bound that calls COMPLETABLE, a lookup, at each node never enters a dead branch, and its branching, its bound and its incumbent are untouched. On the eight- and twelve-orientation sets the test removes more than 99% of what a locally checked tree would visit before its bound acts.
Third, with the counters uncapped, the cells reachable from a node are the compositions its admissible completions can still attain, so a bound on a quantity that depends on the plies only through their counts, such as the membrane stiffness, can be taken over those compositions alone.
Exact uniform sampling
Algorithm 3 draws every admissible word with probability , in steps with no rejection (Theorem 3.2). The sampler never fails: it either returns a compliant laminate or reports that the language is empty, which it knows from the count.
A crossover valid by construction
Proposition 6.1 gives two conditions under which a one-point cut of two compliant half-stacks yields a compliant child. The Parikh vector of the child is a difference of prefix and suffix counters, so both conditions, equality of states and the Parikh test, are integer tests of cost per candidate cut, and a child costs once the prefix states and counters of the two parents are in hand.
Proposition 6.1. Let and be compliant half-stacks of plies and let . If the automaton state after equals the state after , and if the Parikh vector of satisfies the conditions of Lemma 2.6, then that word is a compliant half-stack.
Proof. The state of Theorem 2.2 determines which continuations are admissible, since (4) depends on the prefix only through it. The prefix is admissible because is, and it leaves the automaton in the state from which read its own suffix; that suffix is therefore readable, and the child satisfies (M1) and (M2). Its final state is the final state of , which has run one because is compliant, so Lemma 2.4 applies. The composition conditions hold by hypothesis. □
Algorithm 4 Cut(h_A, h_B): recombination inside the compliant language.
1: q_i^A, q_i^B ← automaton states after h_A[1..i], h_B[1..i], for 1 ≤ i < m
2: c_i^A, c_i^B ← Parikh vectors of the same prefixes
3: I ← { i : q_i^A = q_i^B and c_i^A + c(h_B) − c_i^B satisfies Lemma 2.6 }
4: if I ≠ ∅ then
5: choose i ∈ I; return h_A[1..i] h_B[i+1..m] ▷ compliant by Prop. 6.1
6: else
7: return Sample(F) ▷ no admissible cut: fresh exact draw
At an admissible cut exists for 49.0% of 3000 independent pairs on and 18.0% on . We verified compliance on the 2008 children with rule checkers written separately from the dynamic program, and the sufficiency of the condition by brute force over 3600 pairs and all their cuts, at and 20 on and and 14 on .
Proposition 6.2. Let be a compliant half-stack of plies. Every cut of with itself is admissible under Proposition 6.1, and every child it produces equals .
Proof. Take and any . The automaton state after equals itself, so the state condition of Proposition 6.1 holds at every ; and the child is , whose Parikh vector satisfies the conditions of Lemma 2.6 because is compliant. □
The fallback to the exact sampler, which is the only source of new material, fires exactly when no cut exists, and therefore never fires once the population has converged.
The proofs of Propositions 6.1 and 6.2 use only that the automaton state determines the admissible continuations and that acceptance depends on the final state and the counters; both therefore hold for any operator that joins a prefix and a suffix meeting in one state and closing the counters together. A diversity guard has to come from outside the operator.
Discussion
No result of this paper evaluates the response of a laminate: the counts, the sampler, the completion test and the recombination concern the admissible set. Closed-form responses exist for simple geometries under elementary loads, such as the buckling load and the natural frequencies of a simply supported rectangular plate of a symmetric and balanced laminate, exact when the bending-twisting terms and vanish. They do not exist for the geometries and load cases of an industrial component, nor, beyond a few antisymmetric lay-ups, for laminates whose couplings cannot be neglected, B in a non-symmetric laminate and in an unbalanced one; there each evaluation is a finite-element solve, and a search that spends evaluations on designs that cannot be built pays for them in solves.
Stacking-sequence optimisation is surveyed in [24]. Four things this paper uses are not ours. Treating the manufacturing constraints exactly is [61]. Expressing the manufacturing rules as forbidden factors and pruning a prefix tree by them is recited in a granted patent [54]. An evolutionary operator that preserves a design guideline is not new either: [28] has crossover and mutation preserving symmetry and balance, and what is new is preserving the whole modelled set in a single cut, each cut decided in . Nor is the sampling method ours. The backward walk proportional to the counts is the recursive method [23]; an automaton accepting under a semilinear condition on counters is a Parikh automaton [35]; and conditioning the recursive method on an exact composition is older still [18, 7], with a Boltzmann counterpart [8]. What is new is therefore narrow: exact conditioning on thresholds and signed differences of the letter counts, rather than on a prescribed composition, realised as one explicit product automaton of size polynomial in that serves every composition at once and decides completability.
Four limitations bound the scope. First, the guidelines modelled are symmetry, balance and the ten per cent rule: damage tolerance, ply drops and blending between adjacent panels are not represented, and blending is not a function of one laminate’s Parikh vector, so it would not fit this construction without extension. The patent family cited above handles blending by choosing each panel’s sublaminate from the pre-generated library so as to couple adjacent panels [32]. Second, the dynamic program tracks the sign pairs without a threshold, and and on also and , through their difference alone, which suffices for balance but not for constraints that weigh those orientations individually, such as the membrane-stiffness criterion of [28] on . Third, this paper measures no search: what the operations save a heuristic in evaluations and in generation time is not quantified here. Fourth, the program is polynomial for a fixed alphabet but its degree grows with the number of sign pairs: the non-symmetric balanced count on , which carries five differences, needs cells at the middle level, 69 GB and an hour and a half at .
Conclusions
The manufacturing rules and guidelines treated here form a regular language conditioned on thresholds and signed differences of its letter counts, so one table per ply count, polynomial in , counts the admissible sequences exactly, draws them uniformly and tells whether a partial sequence can still be completed. In a branch and bound it supplies the completion test without which 99.8% of a locally checked tree is dead at 48 plies on eight orientations, and in a genetic algorithm a recombination that returns admissible children by construction. Blending across panels, ply drops and damage tolerance are the constraints this construction does not yet carry.
Funding
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Data availability
The program of Algorithms 1–4 and the meet-in-the-middle counter are available on Zenodo [38]. The rule checkers of § 3.6, written separately from the dynamic program, are those of the open-source design workflow of [37], also on Zenodo. The scripts were written and run with the assistance of Claude (Anthropic), and every number reported was recomputed by a separate verification script.
Declaration of generative AI and AI-assisted technologies in the manuscript preparation process
During the preparation of this work the authors used Anthropic’s Claude, through the Claude Code interface, in order to draft and edit manuscript text, to write and run the counting, sampling and search scripts, and to check the cited sources against the statements that cite them. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.
References
- [1]Abdalla M. Abdalla, Shahriar Setoodeh, and Zafer Gürdal. Design of variable stiffness composite panels for maximum fundamental frequency using lamination parameters. Composite Structures, 81(2):283–291, 2007. doi:10.1016/j.compstruct.2006.08.018.DOI
- [2]M. Abouhamze and M. Shakeri. Multi-objective stacking sequence optimization of laminated cylindrical panels using a genetic algorithm and neural networks. Composite Structures, 81(2):253–263, 2007. doi:10.1016/j.compstruct.2006.08.015.
- [3]Alejandro Albanesi, Facundo Bre, Victor Fachinotti, and Cristian Gebhardt. Simultaneous ply-order, ply-number and ply-drop optimization of laminate wind turbine blades using the inverse finite element method. Composite Structures, 184:894–903, 2018. doi:10.1016/j.compstruct.2017.10.051.DOI
- [4]Mazen A. Albazzan, Ramy Harik, Brian F. Tatting, and Zafer Gürdal. Efficient design optimization of nonconventional laminated composites using lamination parameters: A state of the art. Composite Structures, 209:362–374, 2019. doi:10.1016/j.compstruct.2018.10.095.DOI
- [5]Haichao An, Shenyan Chen, and Hai Huang. Laminate stacking sequence optimization with strength constraints using two-level approximations and adaptive genetic algorithm. Structural and Multidisciplinary Optimization, 51(4):903–918, 2015. doi: 10.1007/s00158-014-1181-0.DOI
- [6]F. Aymerich and M. Serra. Optimization of laminate stacking sequence for maximum buckling load using the ant colony optimization (ACO) metaheuristic. Composites Part A: Applied Science and Manufacturing, 39(2):262–272, 2008. doi:10.1016/j.compositesa.2007.10.011.DOI
- [7]Alberto Bertoni, Paolo Massazza, and Roberto Radicioni. Random generation of words in regular languages with fixed occurrences of symbols. In Proceedings of WORDS’03, 4th International Conference on Combinatorics on Words, TUCS General Publication 27, pages 332–343, Turku, Finland, 2003. Turku Centre for Computer Science.
- [8]Olivier Bodini and Yann Ponty. Multi-dimensional Boltzmann sampling of languages. In DMTCS Proceedings, 21st International Meeting on Probabilistic, Combinatorial and Asymptotic Methods in the Analysis of Algorithms (AofA’10), 2010. Article 2793. doi: 10.46298/dmtcs.2793.DOI
- [9]Timothy R. Brooks and Joaquim R.R.A. Martins. On manufacturing constraints for tow-steered composite design optimization. Composite Structures, 204:548–559, 2018. doi:10.1016/j.compstruct.2018.07.100.DOI
- [10]Kelvin J. Callahan and George E. Weeks. Optimum design of composite laminates using genetic algorithms. Composites Engineering, 2(3):149–160, 1992. doi:10.1016/0961-9526(92)90001-M.DOI
- [11]Giacomo Canale, Stuart Andrews, Felice Rubino, Angelo Maligno, Roberto Citarella, and Paul M. Weaver. Realistic stacking sequence optimisation of an aero-engine fan blade-like structure subjected to frequency, deformation and manufacturing constraints. The Open Mechanical Engineering Journal, 12(1):151–163, 2018. doi:10.2174/1874155X01812010151.DOI
- [12]Giacomo Canale, Paul M. Weaver, Felice Rubino, Angelo Maligno, and Roberto Citarella. Lay-up optimization of laminated composites using a modified branch and bound method. The Open Mechanical Engineering Journal, 12(1):138–150, 2018. doi:10.2174/1874155X01812010138.DOI
- [13]Nan Chang, Wei Wang, Wei Yang, and Jian Wang. Ply stacking sequence optimization of composite laminate by permutation discrete particle swarm optimization. Structural and Multidisciplinary Optimization, 41(2):179–187, 2010. doi:10.1007/s00158-009-0417-x.DOI
- [14]M. Crochemore, F. Mignosi, and A. Restivo. Automata and forbidden words. Information Processing Letters, 67(3):111–117, 1998. doi:10.1016/S0020-0190(98)00104-5.
- [15]Maxime Crochemore, Filippo Mignosi, Antonio Restivo, and Sergio Salemi. Data compression using antidictionaries. Proceedings of the IEEE, 88(11):1756–1768, 2000. doi:10.1109/5.892711.
- [16]Sophie Demassey, Gilles Pesant, and Louis-Martin Rousseau. A cost-regular based hybrid column generation approach. Constraints, 11(4):315–333, 2006. doi:10.1007/s10601-006-9003-7.DOI
- [17]Alain Denise, Yann Ponty, and Michel Termier. Controlled non-uniform random generation of decomposable structures. Theoretical Computer Science, 411(40–42):3527–3552, 2010. doi:10.1016/j.tcs.2010.05.010.DOI
- [18]Alain Denise, Olivier Roques, and Michel Termier. Random generation of words of context-free languages according to the frequencies of letters. In Danièle Gardy and Abdelkader Mokdad, editors, Mathematics and Computer Science: Algorithms, Trees, Combinatorics and Probabilities, Trends in Mathematics, pages 113–125. Birkhäuser, 2000. doi:10.1007/978-3-0348-8405-1_10.DOI
- [19]Philippe Duchon, Philippe Flajolet, Guy Louchard, and Gilles Schaeffer. Boltzmann samplers for the random generation of combinatorial structures. Combinatorics, Probability and Computing, 13(4–5):577–625, 2004. doi:10.1017/S0963548304006315.DOI
- [20]Amir Ehsani and Jalil Rezaeepazhand. Stacking sequence optimization of laminated composite grid plates for maximum buckling load using genetic algorithm. International Journal of Mechanical Sciences, 119:97–106, 2016. doi:10.1016/j.ijmecsci.2016.09.028.
- [21]Noémie Fedon, Paul M. Weaver, Alberto Pirrera, and Terence Macquart. A method using beam search to design the lay-ups of composite laminates with many plies. Composites Part C: Open Access, 4:100072, 2021. doi:10.1016/j.jcomc.2020.100072.DOI
- [22]Philippe Flajolet and Robert Sedgewick. Analytic Combinatorics. Cambridge University Press, Cambridge, 2009. doi:10.1017/CBO9780511801655.DOI
- [23]Philippe Flajolet, Paul Zimmermann, and Bernard Van Cutsem. A calculus for the random generation of labelled combinatorial structures. Theoretical Computer Science, 132(1–2):1–35, 1994. doi:10.1016/0304-3975(94)90226-7.DOI
- [24]Hossein Ghiasi, Damiano Pasini, and Larry Lessard. Optimum stacking sequence design of composite materials. Part I: Constant stiffness design. Composite Structures, 90(1):1–11, 2009. doi:10.1016/j.compstruct.2009.01.006.DOI
- [25]L. J. Guibas and A. M. Odlyzko. String overlaps, pattern matching, and nontransitive games. Journal of Combinatorial Theory, Series A, 30(2):183–208, 1981. doi:10.1016/0097-3165(81)90005-4.DOI
- [26]Timothy Hickey and Jacques Cohen. Uniform random generation of strings in a context-free language. SIAM Journal on Computing, 12(4):645–655, 1983. doi:10.1137/0212044.DOI
- [27]F.-X. Irisarri, A. Lasseigne, F.-H. Leroy, and R. Le Riche. Optimal design of laminated composite structures with ply drops using stacking sequence tables. Composite Structures, 107:559–569, 2014. doi:10.1016/j.compstruct.2013.08.030.DOI
- [28]François-Xavier Irisarri. Stratégies de calcul pour l’optimisation multiobjectif des structures composites. Thèse de doctorat, Université Paul Sabatier — Toulouse III, Toulouse, France, 2009. HAL: tel-00514120. URL: https://theses.hal.science/tel-00514120v1.
- [29]Mark R. Jerrum, Leslie G. Valiant, and Vijay V. Vazirani. Random generation of combinatorial structures from a uniform distribution. Theoretical Computer Science, 43:169–188, 1986. doi:10.1016/0304-3975(86)90174–x.DOI
- [30]Zhao Jing, Xueling Fan, and Qin Sun. Stacking sequence optimization of composite laminates for maximum buckling load using permutation search algorithm. Composite Structures, 121:225–236, 2015. doi:10.1016/j.compstruct.2014.10.031.DOI
- [31]Robert M. Jones. Mechanics of Composite Materials. CRC Press, Boca Raton, 2nd edition, 2018. doi:10.1201/9781498711067.DOI
- [32]Laura Sumi Kang and Adriana Willem pje Blom. Optimization of ply orientations for multi-layer composite parts. U.S. Patent 10,467,350 B2, assigned to The Boeing Company, 2019. Filed 8 February 2016; granted 5 November 2019.
- [33]Jung-Seok Kim. Development of a user-friendly expert system for composite laminate design. Composite Structures, 79(1):76–83, 2007. doi:10.1016/j.compstruct.2005.11.030.DOI
- [34]Tae-Uk Kim and In Hee Hwang. Optimal design of composite wing subjected to gust loads. Computers & Structures, 83(19–20):1546–1554, 2005. doi:10.1016/j.compstruc.2005.02.002.DOI
- [35]Felix Klaedtke and Harald Rueß. Monadic second-order logics with cardinalities. In Automata, Languages and Programming (ICALP 2003), volume 2719 of Lecture Notes in Computer Science, pages 681–696, 2003. doi:10.1007/3-540-45061-0_54.DOI
- [36]A. Kloczkowski and R. L. Jernigan. Transfer matrix method for enumeration and generation of compact self-avoiding walks. I. Square lattices. The Journal of Chemical Physics, 109(12):5134–5146, 1998. doi:10.1063/1.477128.DOI
- [37]Alessio Langiu, Felice Rubino, and Giacomo Canale. composite-minply: an open-source validated workflow for minimum-thickness composite laminate design under buckling constraints, 2026. doi:10.5281/zenodo.22109868.DOI
- [38]Alessio Langiu, Felice Rubino, and Giacomo Canale. Counting and sampling manufacturable stacking sequences: code. Version 1.0.0 [software]. Zenodo, 2026. doi:10.5281/zenodo.23391635.arxiv.org/abs/2610.23146
- [39]Rodolphe Le Riche and Raphael T. Haftka. Optimization of laminate stacking sequence for buckling load maximization by genetic algorithm. AIAA Journal, 31(5):951–956, 1993. doi:10.2514/3.11710.DOI
- [40]Boyang Liu, Raphael T. Haftka, Mehmet A. Akgün, and Akira Todoroki. Permutation genetic algorithm for stacking sequence design of composite laminates. Computer Methods in Applied Mechanics and Engineering, 186(2–4):357–372, 2000. doi:10.1016/S0045-7825(99)90391–2.DOI
- [41]Xiaoyang Liu, Carol A. Featherston, and David Kennedy. Two-level layup optimization of composite laminate using lamination parameters. Composite Structures, 211:337–350, 2019. doi:10.1016/j.compstruct.2018.12.054.DOI
- [42]Ryosuke Matsuzaki and Akira Todoroki. Stacking-sequence optimization using fractal branch-and-bound method for unsymmetrical laminates. Composite Structures, 78(4):537–550, 2007. doi:10.1016/j.compstruct.2005.11.015.
- [43]David R. Morrison, Sheldon H. Jacobson, Jason J. Sauppe, and Edward C. Sewell. Branch-and-bound algorithms: A survey of recent advances in searching, branching, and pruning. Discrete Optimization, 19:79–102, 2016. doi:10.1016/j.disopt.2016.01.005.
- [44]OEIS Foundation Inc. Entry A000984: central binomial coefficients. The On-Line Encyclopedia of Integer Sequences, 2026. https://oeis.org/A000984.
- [45]OEIS Foundation Inc. Entry A123620: Expansion of (1 + x + x²)/(1 − 3x − 3x²). The On-Line Encyclopedia of Integer Sequences, 2026. https://oeis.org/A123620.
- [46]Johan Oudinet, Alain Denise, and Marie-Claude Gaudel. A new dichotomic algorithm for the uniform random generation of words in regular languages. Theoretical Computer Science, 502:165–176, 2013. doi:10.1016/j.tcs.2012.07.025.DOI
- [47]François Pachet and Pierre Roy. Markov constraints: steerable generation of Markov sequences. Constraints, 16(2):148–172, 2011. doi:10.1007/s10601-010-9101-4.
- [48]Gilles Pesant. A regular language membership constraint for finite sequences of variables. In Principles and Practice of Constraint Programming – CP 2004, Lecture Notes in Computer Science, pages 482–495, 2004. doi:10.1007/978-3-540-30201-8_36.DOI
- [49]A. Rama Mohan Rao and N. Arvind. A scatter search algorithm for stacking sequence optimisation of laminate composites. Composite Structures, 70(4):383–402, 2005. doi:10.1016/j.compstruct.2004.09.031.DOI
- [50]Ali Rashed and Eralp Demir. Design of variable stiffness composites for maximum fundamental frequency considering manufacturing constraints of tow steering. Composite Structures, 284:115151, 2022. doi:10.1016/j.compstruct.2021.115151.DOI
- [51]Michael H. Shirk, Terrence J. Hertz, and Terrence A. Weisshaar. Aeroelastic tailoring – theory, practice, and promise. Journal of Aircraft, 23(1):6–18, 1986. doi:10.2514/3.45260.DOI
- [52]Grant Soremekun, Zafer Gürdal, Christos Kassapoglou, and Darryl Toni. Stacking sequence blending of multiple composite laminates using genetic algorithms. Composite Structures, 56(1):53–62, 2002. doi:10.1016/S0263-8223(01)00185-4.DOI
- [53]Richard Joel Thompson and Adriana Willempe Blom-Schieber. Ply optimization feasibility analysis for multi-layer composite parts. U.S. Patent 10,322,792 B2, assigned to The Boeing Company, 2019. Priority 8 February 2016; granted 18 June 2019.
- [54]Richard Joel Thompson and Adriana Willempe Blom-Schieber. Sublaminate library generation for optimization of multi-panel composite parts. U.S. Patent 10,380,319 B2, assigned to The Boeing Company, 2019. Filed 5 January 2017 as a continuation-in-part of application 15/018,075 (priority 8 February 2016); granted 13 August 2019.
- [55]Akira Todoroki and Raphael T. Haftka. Stacking sequence optimization by a genetic algorithm with a new recessive gene like repair strategy. Composites Part B: Engineering, 29(3):277–285, 1998. doi:10.1016/S1359-8368(97)00030-9.DOI
- [56]Akira Todoroki and Yuichiro Terada. Improved fractal branch and bound method for stacking-sequence optimizations of laminates. AIAA Journal, 42(1):141–148, 2004. doi:10.2514/1.9038.DOI
- [57]G. Verchery. Design rules for the laminate stiffness. Mechanics of Composite Materials, 47(1):47–58, 2011. doi:10.1007/s11029-011-9186-x.DOI
- [58]Herbert S. Wilf. A unified setting for sequencing, ranking, and selection algorithms for combinatorial objects. Advances in Mathematics, 24(3):281–291, 1977. doi:10.1016/0001-8708(77)90059-7.DOI
- [59]Arne Wulff, Boyang Chen, Matthew Steinberg, Yinglu Tang, Matthias Möller, and Sebastian Feld. Quantum computing and tensor networks for laminate design: a novel approach to stacking sequence retrieval. Computer Methods in Applied Mechanics and Engineering, 432:117380, 2024. doi:10.1016/j.cma.2024.117380.DOI
- [60]Junbo Yang, Bifeng Song, Xiaoping Zhong, and Peng Jin. Optimal design of blended composite laminate structures using ply drop sequence. Composite Structures, 135:30–37, 2016. doi:10.1016/j.compstruct.2015.08.101.DOI
- [61]S. Zein, P. Basso, and S. Grihon. A constraint satisfaction programming approach for computing manufacturable stacking sequences. Computers & Structures, 136:56–63, 2014. doi:10.1016/j.compstruc.2014.01.016.DOI