SPRINGBACK COMPENSATION IN SHEET METAL FORMING

USING A SUCCESSIVE RESPONSE SURFACE METHOD

Nielen Stander, Mike Burger, Xinhai Zhu and Bradley Maker

Livermore Software Technology Corporation, 7374 Las Positas Road, Livermore, CA

Email: [email protected], [email protected], [email protected], [email protected]

Abstract corrections to compensate for springback are made by

modifying the shape of the stamping tools.
The purpose of this study is to demonstrate a
methodology for springback compensation in sheet Before the computer simulation became acceptable,
metal stamping operations. An optimization method trial-and-error methods were used for springback
based on the D-optimality criterion for experimental compensation in industry. This approach requires many
design and successive linear response surface years of die-shop experience and is also very time
approximations is employed to minimize the difference consuming – to make a modified die set could take
between the simulation results and the intended design. months. In addition, several trial-and-error corrections
A subdomain reduction scheme is applied to determine are frequently required before adequately compensated
the subdomain size for each iteration. The procedure, parts are obtained. Accordingly, the trial-and-error
which has been implemented in the program LS-OPT, process is very expensive, often requiring on the order
is interfaced with the simulation package LS-DYNA, of a million dollars to make a die that produces “good”
and the parametric preprocessor TrueGrid. These tools parts. When new materials are used or when a new
are used to input original tool geometry, material, and design is adopted, previous experience cannot be
process parameters, identify design variables, perform applied directly. The difficulties associated with the
springback simulations and output optimized tool trial-and-error method can result in uncertainty in
geometry. The standardized NUMISHEET’96 S-Rail is predicting costs and lead-times.
used as a benchmark example in this study. A
converged optimized design is obtained in four or five Computer simulation has gained popularity in the
iterations. It is found that springback trends are stamping industry due to its speed and low cost, and it
consistent with changes in the die shape, thereby has been proven to be effective in prediction of
suggesting other effective strategies for springback formability and springback behavior2. However, no
compensation. effective direct method is available to modify the die
based on the predicted springback. Rather, an inverse
Introduction analysis requiring an iterative scheme seems to be a
requirement. Despite this drawback, approximate
Springback is an elastic deformation which occurs at methods exist which can, in some cases, improve the
the end of a sheet metal stamping process, as the tool design.
stamped part is removed from the stamping tools. This
phenomenon has the effect of changing the finished The spring-forward method is based on numerical
shape of the part so that it no longer matches the simulation by Finite Element Analysis (FEA). The
forming tools. If this shape deviation is large, it can method begins by performing a stamping simulation,
cause difficulty during a subsequent assembly process, from which information for the deformed part is
or render the assembled part unusable. Accordingly, it obtained while it is still positioned in the closed dies.
is important to produce parts of which the finished This information includes the geometry, and material
shapes closely match the designed surface. Usually stress and strain data. The method then assumes that
subsequent springback deformation will be driven by
material stress, and that if the stress distribution
Copyright © 2002 by Livermore Software Technology through the material thickness is reversed, the resulting
Corporation. Published by the American Institute of springback deformation will also be in the reversed
Aeronautics and Astronautics, Inc., with permission.

1
American Institute of Aeronautics and Astronautics
direction, as compared to the actual part. Based on this Methodology
logic, the geometry, which is obtained by springback
analysis with reversed stress, can be used to predict Consider the general nonlinear optimization problem:
modifications to the dies. This method is very simple to
apply, and it is the most popular numerical method. Minimize
However, the method suffers from two major f ( x ), x∈R n (1)
shortcomings that prohibit use in many practical
applications. These are under-cut (interference between subject to the inequality constraints
the parts during the stamping process) and accuracy.
L j ≤ g j ( x )≤U j ; j =1,2,K,m (2)
The purpose of this study is therefore to investigate an
effective iterative algorithm for springback
compensation. An optimization method is applied to and simple bounds on the design variables
solve the inverse problem. Several component tools are
required. TrueGrid 9 is used as a pre-processor to xil ≤ xi ≤ xiu ;i =1,K,n (3)
parametrically define geometry of the rigid tools, LS-
DYNA4 software is used as a Finite Element based where Lj and Uj refer to the upper and lower bounds on
solver and LS-OPT6 is used to guide the solution to an each of the inequality constraints, and xil and xiu the
optimum. The NUMISHEET’96 S-RAIL1 example is lower and upper bounds on each of the design
used as a benchmark. variables, n is the number of design variables, and m
the number of inequality constraints. Note that equality
Approach constraints can be written as two inequality constraints
in the form of Equation 2 with Lj equal to Uj.
The die compensation analysis presented here uses an
optimization method based on an iterative response References 5, 7 and 8 can be consulted for a detail
surface scheme. Many stamping/springback simulations description of the Successive Response Surface Method
are performed during the iterative process. Using the (SRSM). The method, as implemented in LS-OPT5
parametric preprocessor TrueGrid, the user can contains a number of features that makes it robust and
introduce a set of design variables. These may include suitable for the solution of practical problems:
geometric variables, such as critical locations where
tool elevation should be modified, important tool radii,
• The D-optimal experimental design is used to best
and starting blank geometry. Process variables may
utilize the number of available runs. Over-
also be specified, such as drawbead restraining forces
sampling of 50%5 is used to maximize the
and binder and pad loads.
predictive capability of the response surfaces.
• Linear approximations are constructed using
Using this design variable information, LS-OPT
linear regression on all the points of the current
automatically creates a set of stamping/springback
iteration. Unit weighting is used for the
models (according to an experimental design method),
regression.
and submit several simultaneous simulation jobs. These
jobs are distributed through a computer network since • An adaptive domain reduction method8 is applied
they execute independently and will therefore exhibit as described below.
near perfect parallel efficiency. • An auxiliary problem that minimizes the
maximum constraint violation6 is solved to
After collecting and processing results from the first set enforce feasible designs. This formulation is
of simulations, LS-OPT predicts optimized values for effective in minimizing maximum discrepancies
each design variable, using a response surface based on such as those resulting from springback.
the result set for each response. Using these optimized
design variables, the next set of simulations The SRSM method uses a region of interest, a subspace
(experimental design) is automatically created and of the design space, to determine an approximate
submitted. This iterative process continues until each optimum. A range is chosen for each variable to
variable has been determined within a specified determine its initial size. A new region of interest
tolerance, or until a limiting number of iterations have centers on each successive optimum. Progress is made
been completed. by moving the center of the region of interest as well as
reducing its size. Figure 1 shows the possible
adaptation of the subregion.

2
American Institute of Aeronautics and Astronautics
 ( 0)
The starting point x will form the center point of the Therefore a new subregion, which is centered on the
first region of interest. The lower and upper bounds current point, does not change its size. This is called
panning (Figure 1(a)). If the optimum point coincides
( xirL,0 , xirR,0 ) of the initial subregion are calculated with the previous one, the subregion is stationary, but
using the specified initial range value ri( 0 ) so that reduces its size (zooming) (Figure 1(b)). Both panning
and zooming may occur if there is partial movement
( k +1)
(Figure 1(c)). The range ri for the new subregion
xirL ,0 = x i(0) −0.5ri( 0)
in the (k + 1)-th iteration is then determined by:
xirU , 0 = xi( 0 ) + 0.5ri ( 0 ) i =1,Kn (4)

ri( k +1) =λ i ri( k ) ; (9)

where n is the number of design variables. The
modification of the ranges on the variables for the next i =1,K,n;k =0,K,niter
iteration depends on the oscillatory nature of the
solution and the accuracy of the current optimum. where λi represents the contraction rate for each design
variable. To determine λi, d i( k ) is incorporated by
A contraction parameter γ is firstly determined based
on whether the current and previous designs x ( k ) and scaling according to a zoom parameter η, typically 0.5,
( k −1)
that represents pure zooming and the contraction
x are on the opposite or the same side of the
parameter γ to yield the contraction rate
region of interest. Thus an oscillation indicator c may
be determined in iteration k as
λi =η + d i( k ) (γ −η ) (10)
(k ) (k ) ( k −1)
c i =d i d i
(5) for each variable independently (see Figure 2).
where
Example: Numisheet 96 S-Rail
d i( k ) =2∆x i( k ) / ri( k ) ; The tools and sheet-metal blank of the Numisheet 96
∆x i( k ) = x i( k ) − x i( k −1) ; (6) springback benchmark1 problem are shown in Figure 3.
The punch is controlled at a constant 1m/s while the
d i( k ) ∈[−1;1] binder is driven by a piece-wise linear force curve as
shown in Fig. 4.
The oscillation indicator (purposely omitting indices i
and k) is normalized as ĉ where
Formulation of the optimization problem
The objective of the design procedure is to maximize
ˆ c sign(c) .
c= (7) the flatness of the flange pair as if the work-piece were
to be welded to a flat surface. To achieve this, a flat
surface is fitted through 24 selected flange points, using
The contraction parameter γ is then calculated as a linear regression analysis. These points are selected at
the flange inner and outer positions as shown in Figure
γ pan (1+cˆ)+γ osc (1−cˆ) 5. The offset of a point can be computed as ei = z i -
γ= . (8)
2
Z i where z i is the vertical coordinate of the point i
The parameter γosc is typically 0.5-0.7 representing and Z i is the vertical position of the point projected on
shrinkage to dampen oscillation, whereas γpan the plane. Using the selected points, two possible main
represents the pure panning case and therefore unity is approaches are available to formulate the design
typically chosen. problem.

The accuracy is estimated using the proximity of the 1. RMS: Compute a root mean square (RMS)
predicted optimum of the current iteration to the residual of the perpendicular offset of each
starting (previous) design. The smaller the distance point on the work-piece after springback,
between the starting and optimum designs, the more 24
∑ei / 24 , and use it as the objective for
2

rapidly the region of interest will diminish in size. If the i =1

solution is on the bound of the region of interest, the minimization.
optimal point is estimated to be beyond the region.

3
American Institute of Aeronautics and Astronautics
 2. Maximum: Constrain the offset of each point: shape is very close to a flat surface (Figs. 8a, b and c).
-E ≤ ei ≤ E, i=1,2,…24 and minimize the Figure 9 shows the optimal tool shape that will
auxiliary variable E, keeping E>0. The effect minimize the surface warp.
of this formulation is to minimize the greatest
offset after springback. Conclusions
Design variables From the above benchmark, we can draw the following
Nine design variables were chosen, namely the radius r conclusions:
which applies to all four corners of the cross-section as
well as the positions of 8 control points on the outer • As with all optimization procedures, the
perimeter of the die and binders. The model was effectiveness of the method for springback
parameterized with TrueGrid. The control points are compensation relies on the choice of a sufficient
connected by straight lines to hinge points at the number of suitable design variables. Even with the
tangent line to the radius. The control points define the choice of nine variables, the springback could
tool surfaces by controlling the z-coordinates of the only be reduced by about 70%.
selected points as shown in Figure 6. For the baseline • Using TrueGrid, the surfaces are mathematically
design all the x values are zero. defined (not simply a mesh perturbation), and
therefore the parametrization can be designed in
accordance with manufacturing requirements.
Note that the binder has been parameterized to assume • Judging by the optimum die geometry, the
the same shape as the die for a snug fit of the tools. At springback behavior is consistent with small to
t=0 the binders and die have to be sufficiently spaced to moderate perturbations in tool geometry. I.e. even
prevent interference with the sheet metal blank (see with the change of the tool geometry, the
Figs. 6 and 8). This condition is formulated in the springback still occurs in the same direction as
TrueGrid input file. before.
Simple bounds have been chosen for the variables so
that the optimization problem formulation becomes: Drawbacks of the optimization method are as follows:

• The choice of design variables depends heavily on

Min E the user’s experience, which makes it difficult for
complex part design. With more powerful
subject to
hardware and software it is conceivable that,
-E ≤ e i ≤ E; i = 1, 2, …24 initially, a large number of variables can be
chosen from which a suitable subset can be
-25mm < xj < 25mm; j = 1, 2, …8 selected by means of variable screening (for
3mm < R < 7mm instance using an analysis of variance (ANOVA)
method3).
where R is the corner radius.
• Optimization is expensive in terms of simulation
Results time. Large, realistic sheet metal forming models
typically have run times of days rather than hours.
LS-OPT, employing Formulation 2 (maximum offset)
was used to optimize the tool design. The problem was In spite of the drawbacks, the results are encouraging in
run on an HP V-class 16 processor server. 8 processors terms of the accuracy of the results obtained and the
were utilized. 16 simulations were conducted per robustness of the method. In the mean time other,
iteration. The time required for a full iteration is 3.5 mesh-based methods are being investigated as a means
hours. About 5 to 6 iterations were required for to accelerate the optimization phase of the procedure by
convergence. Further iterations were run to attempt reducing the number of variables.
finer convergence, but the maximum offset remained at
~0.8mm compared to the baseline 3.2mm (Figure 7a).

The dots represent the simulated results using LS-

DYNA, whereas the line represents the response
surface prediction. The history of variable x8 (lower
and upper bounded line in Figure 7b) suggests
convergence by iteration 5. Note that the optimal flange

4
American Institute of Aeronautics and Astronautics
 4. Livermore Software Technology Corporation
(2001), LS-DYNA User’s Manual, Version 960.
References 5. Roux, W.J., Stander, N., and Haftka, R.T. (1998),
1. Anon. S-Rail Benchmark Problem (1996). “Response Surface Approximations for Structural
Proceedings of the 3rd International Conference: Optimization”, International Journal for
Numerical Simulation of 3-D Sheet-Metal Forming Numerical Methods in Engineering, 42, 517-534.
Processes – Verification of Simulations with 6. Stander, N. (1999), LS-OPT User’s Manual
Experiments (NUMISHEET ’96). Eds. Lee, Kinzel Version 1, Livermore Software Technology
and Wagoner, September, 1996, pp. 612-799. Corporation, Livermore, CA.
2. He, N. and Wagoner, R.H (1996). Springback 7. Stander, N. (2001) “The Successive Response
Simulation in Sheet Metal Forming. Proceedings Surface Method Applied to Sheet-Metal Forming”,
of the 3rd International Conference: Numerical Proceedings of the First MIT Conference on
Simulation of 3-D Sheet-Metal Forming Processes Computational Fluid and Solid Mechanics, Boston,
– Verification of Simulations with Experiments June 12-14, 2001. Elsevier Science Ltd., Oxford.
(NUMISHEET ’96). Eds. Lee, Kinzel and 8. Stander N, Craig KJ (2002). On the robustness of a
Wagoner, September, 1996, pp. 308-315. domain reduction scheme for simulation-based
3. Khuri, A.I. and Cornell, J.A. (1996) Response optimization. To appear. Engineering
Surfaces, Designs and Analyses. Marcel Dekker, Computations. 2002. 19(4).
Inc., New York. p.43. 9. TrueGrid manual (2000), XYZ Scientific
Applications, Inc. Livermore, CA.

pan zoom pan & zoom

x1 rU,0
subregion(1) x1 range r1 (0)
x (0) x (0)

x1 rL,0 x (1)
x (1)
x (0) = x (1)
range r1 (1)
subregion(2)

(a) x2 (b) x2 (c) x2

Figure 1: Adaptation of subregion in SRSM: (a) pure panning, (b) pure zooming and (c) a combination of panning
and zooming

5
American Institute of Aeronautics and Astronautics
 λ
λ,γ
γpan

η
ĉ
1

0 γosc
-1

1 |d|

Figure 2: Domain reduction scheme: the sub-region contraction rate λ as a function of the oscillation indicator ĉ
and the absolute move distance |d |

Flat sheet
metal blank

Die

Binder

Punch
Binder

Figure 3: Numisheet 96 benchmark: Punch, die, binders and blank (baseline design)

6
American Institute of Aeronautics and Astronautics
 Force
1

0 1 Time 3

Figure 4: Binder Force as a function of time

Figure 5: Monitoring points on flanges (top view)

7
American Institute of Aeronautics and Astronautics
 x8
x7
x6 Spline
x5

x4
Radius x3
Die x2
Binder x1

Figure 6: Design variables ( x1 to x8 and Radius)

3.5 10
3 Simulation 5 Value

2.5 0 Bounds
Maximum offset

Response
2 Surface -5
X8

1.5 -10
1 -15

0.5 -20

0 -25
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7

Iteration Num ber Iteration Number

Figure 7: Optimization history of (a) maximum offset and (b) variable x8

8
American Institute of Aeronautics and Astronautics
 5.45mm

Figure 8a: Shape after springback (baseline tool design)

1.85mm

Figure 8b: Shape after springback (iteration 3)

1.67mm

Figure 8c: Shape after springback (iteration 10)

9
American Institute of Aeronautics and Astronautics
 -17.9mm
-2.9mm R=7.0mm
-4.8mm
-.64mm

-2.5mm

-7.2mm

-14.2mm

-10.9mm

Fig. 9: Optimum die geometry (iteration 10)

10
American Institute of Aeronautics and Astronautics