OPTIMAL DESIGN OF COMPOSITE FUSELAGE FRAMES FOR PROGRESSIVE FAILURE AND ENERGY ABSORPTION BY GENETIC ALGORITHMS.

Marshall B. Woodson*,and Eric R. Johnson** Department of Aerospace and Ocean Engineering Virginia Polytechnic Institute and State University Blacksburg, Virginia 24061-0203 Raphael T. ~aftka*** Department of Aerospace Engineering, Mechanics and Engineering Science University of Florida Gainsville, F1. 3261 1-2031

Abstract
This work presents a method for preliminary design of laminated composite fuselage frames with improved crashworthiness. The proposed method uses a progressive failure analysis and genetic algorithms to obtain frame designs optimized for maximum energy absorption. To demonstrate the method, a representative composite fuselage frame is designed for failure at two different load levels. The analytical results indicate that significant increases in crashworthiness can be achieved with this frame design methodology. Vehicle structural designs are considered crashworthy if occupants are protected from exposure to high accelerations and other crash hazards('"). A key to achieving such protection is controlled absorption of crash impact energy through progressive failure of the structure. The structural designer attempts to control the failure process so as to tailor the magnitude and duration of the crash acceleration pulse for increased survivability. One of the failure mechanisms commonly used for crash energy management of an aircraft structure is fuselage crushing below the main passenger deck'. This failure mode is governed primarily by fuselage frame design, Ideally a frames load response will have an elastic-plastic character. We seek a controlled failure load followel by a sustained crushing force as a function of displacement at the point of contact. The objective is to maximize the energy absorbed, without peak occupant accelerations

* ** ***

National Kesearch Council Associate. Correspondence author. Associate Professor. member AIAA and ASME Professor: Associate fellow AIAA

This paper is declmed a work of the 1J.S. Government and is not subject to copyright protection in the lJnited Statcs.

exceeding the limits of human survivability. To achieve this, fuselage frames must be designed for an optimal crushing response. Frame design requirements established by the normal flight and service loads, are constrained by the additional design requirements of a "potentially survivable" crash. Fuselage structures constructed from composite materials require a special attention to crashworthiness4. Conventional aluminum skin-stringer designs will absorb kinetic crash energy through a process that involves ductile crushing and yielding of the fuselage structure. This type of crushing behavior is attributed to the formation of plastic hinges in the frames at the locations where local bending moments are high. For a laminated composite skin-stringer fuselage design, the thin walled frames are predisposed to fail by brittle fracture at these same locations. Since less material is involved in the failure process, this brittle failure mode can result in lower energy absorption. As a consequence, designing composite fuselage structures to provide sufficient levels of crash energy absorption requires special attention. Maintaining or improving upon the existing levels of crashworthiness exhibited by conventional aluminum fuselage structures is an important design criterion for composite fuselage structures In the present work, the design of hselage frames for improved crashworthiness is approached as an optimization problem We seek to maximize the ener$y absorbed subject to a constraint on maximum failure load. Energy absorbed is equal to the area under the load response curve, and passenger accelerations are directly related to failure loads The optimum structural response for crashworthiness is a controlled failure load followed by a

sustained crushing force. (See Figure 1 ). Force f (8) Failure

1
I

Energy Absorbed = Area under f-S curve

-1

Displacement (6)

6=Ax

1;igui.e I Desired failure resporlse

The load response curves for each frame design are obtained from a progressive failure analysis The traditional approach to progressive failure analysis is to make a detailed branched shell model and perform the analysis using a materially and geometrically nonlinear finite element analysis methodology such as DYNA3D5 These analysis programs calculate the progressive failure response by coupling an incremental finite element analysis in space, with a finite difference analysis in time. Nonlinear material laws are followed by incremental changes in the loads and displacements at each time step With this analysis approach, a single progressive failure analysis of one structural component can require many hours of computer time Because of the high computational requirements, these analysis methodologies are impractical for optimization in today's computing environment The computational requirements of progressive failure analysis must be reduced before a practical optimization methodology can be developed The approach taken in this work is to use efficient structural models with predefined failure modes We maintain enough analytical sophistication to capture the character of the stnictural response, but keep it simple so that optimization remains feasible

For the present work, there is a focus on ingreasing progressive failure analysis speed. i his is viewed as a necessary step for development of a practical optimization methodology for crashworthiness. Two approaches to increasing analysis speed are used. First, the traditional branched shell finite element models for frames are replaced with more efficient beam finite element models, and second, the incremental finite element analysis is replaced with a stepwise linear solution. While it would be desirable to have the capability for an exact progressive failure analysis, this is not considered feasible given the degree of analytical simplification mandated by the computational requirements of optimization for crashworthiness. The beam finite element is based on an extension of Vlasov's6 beam theory to composite materials7. The Vlasov approach to beam theory was selected based on its superior performance under torsional loads. The use of a beam theory as opposed to the more traditional branched shell analysis represents a trade-off of accuracy for efficiency. Analytical experiments show that beam finite element models are typically about two orders of magnitude faster than equivalent branched shell models, however, their accuracy is not always acceptable. In particular, some curved beam cross sections are susceptible to analysis errors resulting from cross-sectional distortion. Extensive comparisons to branched shell finite element results show that the current composite curved beam finite element is suficiently accurate for preliminary design and analysis of frames for crashworthiness A detailed description of the theory as well as results of the comparisons to equivalent branched shell models can be found in Refs. [7], and [8]. In general, the beam analysis proved capable of producing reasonably accurate results for a variety of open cross-sections and laminate stacking sequences. Employing a reasonably accurate and efficient beam theory was viewed

as critical to the ultimate success of the optimization. The present stepwise linear model for progressive failure of the composite beam element was developed from an extension of Tsai's9 selective and progressive lamina degradation model. For fuselage frames with the characteristic brittle failure mode described earlier, the progressive failure response can be approximated using a series of linear failure analysis. First, the frame is modeled using curved beam finite elements and an analysis is performed. From the results of this analysis, the minimum strength ratio1 is calculated and a failure load for the critical frame element is determined. To limit the number of discrete failure events in the progressive failure analysis, all structural elements having a strength ratio within 2.5% of the minimum are failed as a group. Failure consists of modifying the strength and stiffness of critical lamina in accordance with Tsai's lamina degradation model. Other failure modes such as crippling are possible, but are not implemented at this time. After the modification cycle, beam stiffnesses are recalculated, and the finite element model is assembled and analyzed. If the calculated displacement at failure is increased from the previous analysis, the analysis is accepted and a new point on the load displacement curve is generated. Otherwise, the modification cycle is repeated until either the displacement is increased or ultimate failure occurs. Figure 2 illustrates how the load displacement response is constructed using

I. Strength ratio is the allowable load divided by the actual load. Strength ratios less than 1 indicate failure.

this methodology. Load Discrete load drop at each failure event (6 )

Where X, and X' are the tensile and compressive strength in the fiber direction, Y and Y' are the tensile and compressive strength in the transverse direction, and S is the in-plane shear strength. The basic geometry of the frame along with some usehl section terminology is shown in Figure 3 and . This frame is designated "lb" by the authors of Ref.[lo]

Displacement

/
1

Radial Load

Discrete stiffness change at each failure event (f: ) Iligure 2 Rrrildirlg the load response curve

radius = 36"

An evaluation of the stepwise method for progressive failure analysis was conducted by comparing the analysis results to the available data from a limited number of test cases. The analysis was applied to an "I" cross-section curved composite fuselage type frame originally tested by E.Moas et. a1. lo. Material for this frame is AS415208, graphite epoxy, with properties as given in Table 1, and Table 2

F@ire 3 Testkame geometry

Yable I StrfJi~ess properties - AS4 5208 graphite epoxy

Stiffness (Msi)

Section AA

I.'igtrre 4 (,'rosssection geomeby

Table 3 Tr.stJi.umelan~imte stacking sequences

Stacking Sequence

li~bie 2 Strength properties - AS4 5208 graphite epoxy

Strength (ksi)
-

Cap Web S Attachment Skin

[+45/-45/0/90]s [+45/-45/0/90]~
[+45/-45/0/90/90/0/+45/-451,

X'

Y'

[+45/-45/0/90/ t45/-45/0/90]s

The nominal lamina thickness of ,00525" was modified as required to match the average measured dimensions. The cross section was modeled using the natural junctions and thickness changes as section element boundaries, and 20 curved beam elements were used along the arc length. Clamped end conditions were assumed at each end of the frame. The experimental load response data for this frame is compared with analysis results in Figure 5. The results shown indicate reasonable agreement for this progressive failure analysis methodology.

(or stochastic) search methods are genetic algorithms, and s~mzr/ateu'arznea/~i~g They are alternatives to traditional optimization methods and have emerged as important tools for certain types of difficult optimization problems Probabilistic methods do not require derivative information, and easily handle a discrete and discontinuous nonlinear design space They frequently locate many optimal and near optimal designs simultaneously, giving the designer a choice. Considering these factors, the genetic optimization algorithm is considered a reasonable approach to the optimization problem Genetic algorithms (developed by Holland" ) use techniques derived from biology, and rely on the application of Darwin's principle of survival of the fittest. When a population of biological creatures is allowed to evolve over generations, individual characteristics that are useful for survival tend to be passed on to future generations. This occurs because the individuals carrying these traits get more chances to breed. In biological populations, these characteristics are stored as chromosomal strings. The basic operations that result in a structured yet randomized exchange of genetic information between the chromosomal strings of reproducing parents include crossover, and occasional m~rtatzon. Genetic algo-

Optimization for Crashworthiness

Two important factors influence the selection of an optimization methodology. First, energy absorption is nonlinearly related to the frame design variables ( Recall that laminate stacking sequence and cross-sectional geometry were selected as design variables), and second, there are discontinuities and local optima in the design space. If gradient based methods are applied to this problem the discontinuities are likely to cause convergence problems and the presence of local optima will make finding the global optimum more difficult. Probabilistic optimization methods overcome the problems just delineated. Two probabilistic

0.0

0.5

1 .O

1.5

2.0

2.5

Displacement (in) Figure 5 Comparison of the present progressive failure analysis to experimental data

rithms do not require derivative information, and easily handle a discontinuous or discrete valued nonlinear design space. These features make genetic algorithms attractive for optimization problems involving laminate stacking

on of C~enetic AlgxUhms to O~timization of Energy Absorption.

The example problem and geometry shown in Figure 3 will be used to illustrate the optimal design methodology. The cross section shape is simplified to reduce the number of independent lamina in the cross section to one.(see Figure 6)

Genetic algorithms mimic the mechanics of natural genetics with mathematical operations that are the counterparts of the natural ones. These operations involve a similar randomized exchange of numbers in strings that represent the design variables. Unlike many search algorithms that move from one point to another in the design space, genetic algorithms work with a population of designs. Keeping many solution points that may have the potential of being close to optimum (local or global), rather than singling out a point early in the process, reduces the risk of getting stuck at a local minimum. This aspect of the genetic algorithms is responsible for the increased chance for obtaining global or near global optima. The outcome of a genetic search is a population of good designs rather than a single design. This aspect can be very useful to the designer. Genetic algorithms are experimentally proven to be robust, and the reader is referred to Goldbergl"or a complete discussion of their theoretical properties. The drawback to genetic algorithms is that a population of designs evolving over generations towards an optimum requires a relatively high number of analyses. A typical optimization might require many thousands of objective function evaluations. This places a practical limit of the type of problems which are feasible to solve with this method. A fast analysis method is essential, and therefore the use of simple analysis models is important. The stepwise progressive failure analysis using the Vlasov type composite curved beam model is very efficient, and permits the use of genetic algorithms for the energy absorption problem at hand

Section AA Figure 6 Cross sectior~ g e o m e ~ f o oplir n~izaliorr Clamped end conditions are assumed and the load is applied radially inward at the apex A uniform cross section is assumed along the circumference of the frame, and the same laminate is used for each branch of the cross section. This laminate is arbitrarily set to a constant thickness of eight plies Material properties are as given in Table 1 and Table 2 Ply orientation and stacking sequence are the design variables These ply orientations are limited to eight discrete values ( 0, +30, -30, 145,-45, +60, -60, and 90 degrees) A constraint is placed on the maximum allowable crushing force, and laminate symmetry is enforced Raw fitness is calculated as the area under the load response curve The size of the design space for this example is 8' or 4096 feasible design configurations Two composite fuselage frames are designed for crashworthiness, and compared to a baseline composite frame. The first frame is

designed to achieve a 500 lb crush load, and the second is designed for a 1000 Ib crush load. The baseline case is a quasi-isotropic wall lay up. Load displacement response diagrams from optimizations at the two different crushing loads are shown in Figure 7, and Figure 8. The failure sequence for both frames is similar. Failure begins near the point of load application on the radially inboard flanges of the frame. The weakest plies fail, then failure progresses to inner flanges at the fixed ends (m degrees ). These are both tensile failures due to bending of the cross section. The process continues by alternating between these two locations until ultimate failure occurs. The total energy absorbed and the optimal laminate stacking sequences obtained are summarized

in Table 4

Table 4 h a m e op/~mrza/ror~ example 5 l

Design Load (Ibs) Energy Absorbed (in*lbs) Optimal Stacking Sequence

The optimal solutions were obtained relatively quickly. For the 500 Ib case, the optimal solution required 52 generations, which represents approximately 12.6% of the design space. For the 1000 Ib case, 43 generations were required, which represents 10.4% of the design space. For both load cases, designs were obtained which did not violate maximum crushing force constraint. The crushing load approached the limit, but remained just below it as a sustained force until ultimate failure. This behavior is

0.0

0.2

0.4

0.6 0.8 Displacement (in)

1.O

1.2

1.4

1.6

Figure 7 Frame optimizedfor 500 lb cru.~hrng force

0.5

1 .O Displacement (in)

1.5

2.0

I+'iprre8 firame optimizedfor 1000 lh cr~ishilg force

precisely what we had hoped to achieve A comparison to progressive failure analysis results for a quasi isotropic [ 45 / -45 / 0 190 Is frame is shown in Figure 8. This quasi isotropic frame represents a composite design which has not been optimized for crashworthiness, but could be used for reduced weight (Note that all three frames represented have the same weight). The maximum crushing load in this case reaches 1392 Ibs, and the total energy absorbed prior to ultimate failure is 8 16.4 in*lbs. At the first limit load level of 500 lbs, the quasi isotropic frame has absorbed approximately 98 in*lbs of crash energy. This represents less than 24% of the energy absorbed by the frame optimized for the 500 lb crush load. At the 1000 Ib limit load, the quasi isotropic frame has absorbed approximately 405 in*lbs of energy, or less that 45% of the energy absorbed by the frame optimized for that load level. It is important to remember that for survivability, the crash energy must be absorbed

before the occupants are exposed to lethal levels of acceleration. For both design cases, the optimized frames would significantly reduce the peak acceleration loads on the vehicle occupants After obtaining the results for laminate optimization, the design methodology was modified to allow for the optimization of cross-sectional dimensions. The basic cross section shape is specified ( I, J, channel,.. .), and dimensions are given by a set of cross sectional node coordinates. Degrees of freedom for these coordinates are then specified as either fixed, free, or constrained. The optimizer is allowed to move the "free" cross sectional degrees of freedom as necessary to improve crashworthiness. The "free" degrees of freedom become independent design variables in the optimization. The 500 lb load case was repeated with one laminate as before, but with three added independent cross section shape variables. The

[ 1000 Ib case ]
0.5 1.0 Displacement (in)

Figure 9 ('omparison to baseline qunsi-isotroprc case Table 5 Optimal design improvement over hnselitie qzinsi isotropic case
Design load case ( w Energy absorbed (in*lbs) Baseline energy (inybs)
% improvement

design space for this example is significantly larger (8%73= 1,404,928).The dimensions of the baseline six noded "I" cross section are given in Table 6 for reference. The free or independent degrees of freedom are shaded in Table 6. The remaining degrees of freedom are constrained such that symmetry of the cross section is preserved.

absorbed energy is increased 49.9%, compared to the equivalent case with no shape optimization (Figure 7) The optimized cross section shape is given in Table 8

Table 8 Optimal cross section dimensionsfor b case, with shape the 500 Z optimization.

7uble 6 Baseline cross section dimensionsfor shape optimization

-0.75 0.0 0.75 -1.75 0.00 1.75

1.3810 1.3810 1.3810 0.0585 0.0585 0.0585

2
3

5
6

The Optimal design obtained for this case is given in Table 7, and Table 8, and the optimal load response is shown in Figure 10 The

Even with the analytical simplifications and assumptions that were made, the computational requirements of optimization for crash-

".
0.0 0.5

1 .O
Displacement (in)

1.5

2.0

Figure I0 Comparison to baseline quasi-isotropic case absorbed for the 500 Ib case with shape optrmization. Table 7 Erierl~y
Energy Absorbed (in*lbs)
633.151

Optimal Stacking Sequence

1-45 145 145 /-601,

% increasea in Ea
49.9

% hcrease from b baseline Ea

546.07

a. % increase over optimized design without shape optimization (see Figure 7) baseline quasi isotropic design ( see Table 5) b. % increase

worthiness are large. An approximation of the solution time required for this optimization method using branched shell finite elements is compared to the actual requirements for the beam model in Table 9 As stated earlier, the conventional approach to this class of progressive failure analysis problems is to use a materially and geometrically nonlinear analysis. It is estimated that adopting this approach would increase the computational requirements for the beam model by at least an order of magnitude. Judging from the solution times presented in Table 9, this additional increase in computational requirements might push optimization for crashworthiness to the limits of practicality. In retrospect, it appears that the choice to use a stepwise linear analysis was appropriate. Successful optimization often requires approximate analysis.

involved bringing together contributions from the areas of crashworthiness, composite beam theory, finite element development, progressive failure analysis, and optimization.The progressive failure analysis methodology developed is specific to composite frames and preliminary comparisons to experimental data show a reasonable correlation given the level of simplification involved. The examples presented demonstrate the capability of the analysis and optimal design methodology. Significant increases in energy absorption are achieved using relatively few independent design variables.

Conclusions
The objective of this work was to develop a practical methodology for optimizing thinwalled composite fuselage frames for crashworthiness. Practical optimization for crashworthiness requires a focus on achieving high computational efficiency while maintaining reasonable accuracy. Developing the analysis Table 9 ('omparisoi~ of reqrrired sol~rtim time

This research is sponsored under NASA Grant NAG-1 -343, the NASA-Virginia Tech Composites program and Mr. Huey Carden of the Landing and Impact Dynamics Branch, NASA Langley Research Center, is the technical monitor.

Kindervater, Ch.. Georgi, 11.. Korber. 11.. "Crashworthy lksign of Aircraft Subfloor Structural Components", E t ~ r r a A hsorptiotz 0f~4it-ci.aji Structurvs as an Aspect of Crashworthit~ess. ACiAKD

Solution 'I'ime ( IBM RISC 6000) Element Assembly & Static Finite Element Solution

Model

Failure Analysis & Modification cycle

of one Progressive b'ailure analysis (20+ analysis) 12.0 s 33 min. 40 s

One Ckncration of 10 Individuals

(he Optimization of 1 0 0 Generations

Present Composite Heam

0.4 s

2 . 0 min

3 hrs. 33 min 23 days. 8 hrs. 52 min

EAI, Branched Shell

8.0 s

5 hrs. 37 min

Conference Proceedings No 443, Luxcnlbourg. Ma? 1988. pp. 12.1 - 12.24 Eianwn, J S . and Tennj son. R C , "The Dynamic Hel~avior of Stiffened Composite Fuselage Shell A~ bsorptron ) of.4 ~rcraft StrucStrr~ctures".F t l e ~ ; ture~ as ail Aspect of C r a s h w o r t h e s s AGARD Conference Proceedmgs No 443, Luxcnlbomg. May 1988.pp 15 1-15 12 Carper. C I1 and Burrows. I, T ."Evolvmg Crash\+orthmess I1cslgn Crlter~a",Energy Ahso~ptroti oj L4~lr;.rajt Structiwes as an .hpect of Crashorthrness. AGAR11 Confercncc Proceedmgs No 443. Luxembourg, May 1988, pp 3 1 - 3 8 Carden, I4 D.. Boitnott. R. I,.. and Fasanella. E . L, , "Behavior of Con~posite/Metal Aircraft Structural Elements and Components IJnder Crash Type Loads--What Are They Telling I Js?", presented at the 17th Congress of the International Cooncil of Aeronautical Sciences, Stockholm, Sweden. September 1990 Ifallquist. J . O . and Renson. 1). J.. "DYNA3D. a computer code for crashworthiness cugineering." Report IJCRL-95 152, IJniversity of California, Lawrence Livermore National Laboratory, Livermorc. CA. September 1986. Vlasov. V.%.,"Thiu-WalledElastic Reams". Second Edition, Moscow, 1959, Translated from Russian and published for the National Science Foundation, Washington. D.C.. 1961. Woodson. M R . Johnson. E K . and IlaftLa. K T , "A Vlasov I'heoq for Lammated Conlposlte C~rcular Beams w ~ t h Thm-Walled Open Sections", The 34th AH.4 ASZfK .4SCEL411S.It5'C Sttuctures, hYt~rctwal Dynatnrcs and Alaterrals Corfe~vnce (Scssion 50 Composite Beams, Plates. and Shells). FIyatt Regency. I aJolla, Californ~a, Apnl 19-21, 1993. (AIAA Paper 93- 1619-CP) Woodson. M.B.. Optimal Design (f?T('omposite Fuselage Fratnes for ('vushw orthiness. Ph.D. Dissertation. Virginia Polytechnic Institute and State 1Jniversity. Blacksburg. Va. December 1994 Tsal, S W . Theoty of ('onlposites Ilesrgn, 'l'hmk Composites. Dayton. Ohlo, 1992 Moas. I. Jr.. noitnott. R. I,.and Griffin. 0 .H. Jr. "An Experimental and Analytical Iwestigation on the Response of GrEp Composite I-Framcs",4nler.ican I felicoptet. Societv .Yational Technical hSprcimlists ' d feeting on Rotoct.a$ Structtrres: "Techt~ologv

for the 1990 :F and Beyond ". Williamsburg, Va. October 1991.

Holland, J.II..Adaptation in ,VaturaI and.4rt1fkial Systems. Universie of Michigan Prcss. Ann Arbor. MI., 1975 I,e Kiche, R. and Haftka. R.T.. "Optin~ization of 1,aniinate Stacking Sequence for B~cliling Load Maximization by Genetic Algorithm", .-1ljl,4 Jorrrnal. Vol 3 1. No. 5, May 1993. pp. 95 1-956. Goldberg. David E.. ( h e t i c .31goritllt~isin Search, Olptimization, and Machine 1,eatning. AddisonWesley Publishing Co.. Inc.. I989