Title: MULTISCALE MODELING OF SOLIDIFICATION OF MULTICOMPONENT ALLOYS
1 MULTISCALE MODELING OF SOLIDIFICATION OF
MULTICOMPONENT ALLOYS
LIJIAN TAN Presentation for Thesis Defense
(B-exam) Date 22 May 2007 Sibley School of
Mechanical and Aerospace Engineering Cornell
University
2ACKNOWLEDGEMENTS
- SPECIAL COMMITTEE
- Prof. Nicholas Zabaras, M A.E., Cornell
University - Prof. Subrata Mukherjee, T A.M., Cornell
University - Prof. Stephen Vavasis, C.S., Cornell University
- Prof. Doug James, C.S., Cornell University
- FUNDING SOURCES
- National Aeronautics and Space Administration
(NASA), Department - of Energy (DoE)
- Sibley School of Mechanical Aerospace
Engineering - Cornell Theory Center (CTC)
Materials Process Design and Control Laboratory
(MPDC)
3OUTLINE OF THE PRESENTATION
- Introduction alloy solidification processes.
- Micro-scale mathematical model
- Applications
- Interaction between multiple dendrites during
solidification - Multi-scale modeling of solidification
- Suggestions for future study
4Introduction and objectives of the current
research
5Introduction
Castings since 5500 BC
6Microstructure
Will it break?
Different microstructures
7Alloy solidification process
8Micro-scale mathematical model
9Mathematical model
Two main difficulties
- Applying boundary conditions on interface for
heat transfer, fluid flow and solute transport. - Multiple moving interfaces (multiple
phases/crystals).
10Complexity of the moving interface
- Jump in temperature gradient governs interface
motion
- No slip condition for flow
11Level Set Method
History Devised by Sethian and Osher (1988) as
a mathematical tool for computing interface
propagation.
Level set variable is simply distance to interface
We pay additional storage and extra computation
time to maintain the above signed distance by
solving
Advantage is that we get extra information
(distance to interface). This information helps
to compute interfacial geometric quantities,
define a novel model, doing adaptive meshing, and
etc.
12Present Model
Assumption 1 Solidification occurs in a
diffused zone of width 2w that is symmetric
around the zero level set. A phase volume
fraction can be defined accordingly.
This assumption allows us to use the volume
averaging technique.
(N. Zabaras and D. Samanta, 2004)
Dont need to worry about boundary conditions of
flow and solute any more!
13Extended Stefan Condition
Gibbs-Thomson condition has to be satisfied (one
of the major difficulties)
Do not want to apply this directly, because any
scheme with essential boundary condition is
numerically not energy conserving. Introduce
another assumption
Assumption 2 The solid-liquid interface
temperature is allowed to vary from the
equilibrium temperature in a way governed by
Temperature boundary condition is automatically
satisfied. Energy is numerically conserving!
Unknown parameter kN. How will selection of kN
affect the numerical solution?
14Numerical Solution For A Simple Problem
Steady state
Initial
If L1, C1
T0 Ice
kN0.001
kN1
kN1000
Conclusion Large kN converges to classical
Stefan problem.
15Stability Analysis
In the simple case of fixed heat fluxes,
interface temperature approaches equilibrium
temperature exponentially.
Stability requirement for this simple case is
16Convergence Behavior
With a grid of 64by64, we get
Results using finer mesh are compared with
results from literature in the next slide.
17Convergence Behavior
Benchmark problem Crystal growth with
initial perturbation.
Triggavason (1996)
Our method
Osher (1997)
Different results obtained by researchers
suggest that this problem is nontrivial.
All the referred results are using sharp
interface model.
Energy conserving makes the difference!
18Computation Requirement
Tracking interface makes the difference!
19Mesh Anisotropy Study
Normal surface tension
Rotated surface tension
20Mesh Anisotropy Study
Crystal shape mainly determined by the
anisotropy in surface tension not the initial
perturbation.
21Applications
- Pure material
- Crystal growth with convection
- Binary alloy
- Multi-component alloy
22Effects of Undercooling
(1) A small change in under-cooling will
lead to a drastic change of tip velocity.
(consistent with the solvability theory) (2)
Increased undercooling leads to sharper dendrite
shape and more obvious secondary dendrites.
23Extension to three dimension crystal growth
- Applicable to low under-cooling (at previously
unreachable range using phase field method, Ref.
Karma 2000) with a moderate grid.
24Crystal Growth with Convection
Velocity of inlet flow at top 0.035
Pr23.1 Other Conditions are the same as the
previous 2d diffusion benchmark problem.
25Crystal Growth with Convection in 3D
- Similar to the 2D case, crystal tips will tilt
in the upstream direction. - Distribute work and storage. (12 processors are
used in the below example)
Thermal boundary layer
26Alloy solidification
For alloys, uniform mesh doesnt work very well
due to the huge difference between thermal
boundary layer and solute boundary layer.
- Difference between thermal boundary layer and
solute boundary layer
- Tree type data structure for mesh refinement
27Simple Adaptive Mesh Test Problem
28Results Using Adaptive Meshing
Le10 (boundary layer differ by 10 times)
Micro-segregation can be observed in the crystal
maximum liquid concentration about 0.05.
(compares well with Ref Heinrich 2003)
29Effects Of Refinement Criterion
Interface position (curved interface) is the
solved variable in this problem. Carefully
choosing the refinement criterion leads to the
same solution using a full grid.
30Crystal Growth of Alloy in 3D
Ni-Cu alloy Copper concentration 0.40831
at.frac. Domain a cube with side length
35mm Difficulties in this problem High
under-cooling 226 K High solidification
speed High Lewis number 14,860
Simulation of crystal growth of alloy in 3D is
computationally very intensive. Our solution is
to use both techniques of domain decomposition
and adaptive meshing!
31Adaptive Domain Decomposition (Mesh Partition)
Mesh
Dual graph
32 Technique Issues about Mesh Partition
- Efficient
- Require mesh partition very frequently
(adaptive). Slow is unacceptable. - Maintain neighboring information using link list,
e.g. for a node, there is a link list for its
neighboring elements, and a link list for its
neighboring edges. - Still linear in storage greatly speed up the
mesh partition procedure. - Parallel
- Keep data distributed, work distributed. (Need to
handle huge data) - Defined a global address (process id pointer)
- Batch way
- (From To) Message Type Message Length
Message content - Put all messages in a link list, and send them
out together
33Demonstration of adaptive domain decomposition
Colored by process id
343D CRYSTAL GROWH (Ni-Cu Alloy)
3 million elements (without adaptive meshing 200
million elements)
353D CRYSTAL GROWTH WITH CONVECTION
Comparing with the pure material case, the growth
for alloy is much more unstable due to the
rejection of solution.
36Multi-component alloy system
Multi-phase system one liquid phase one or
more than one solid phases.
We use a signed distance function for each phase.
- Relation between the signed distances
- Exactly one signed distance would be negative
- The smallest positive signed distance has same
absolute value of the negative signed distance
37Compute Eutectic Growth with Multiple Level Sets
- Stable growth with 4 seeds
- Unstable growth with 2 seeds
- Unstable to stable growth with 10 seeds
38Compute Peritectic Growth with Multiple Level Sets
Solute concentration for peritectic growth of Fe
0.3wt C alloy at time 0.6s, 1.5s, 1.8s, and
2.4s.
39Interaction between multiple crystals
40Handle Multiple Interfaces
Method 1 A signed distance function for each
phase.
Method 2 Markers to identify different region
Each color (orientation of the crystal) is used
as a marker.
Efficient, appropriate for hundreds of crystals.
41Crystal orientation
Different crystal orientation leads to different
growth velocity.
42Extension of crystal orientation
As a feature of level set method, interface
velocity must be evaluated at nodes near
interface on both sides. Crystal orientation
needs to be extended a certain distance away from
the crystal to the liquid region.
43Simple numerical study
The purpose of this study is to verify the
accuracy of using markers.
Growth of 9 initial seeds (circular shape) with
different orientation.
44Comparison with method using multiple level sets
Dashed line method with multiple level set
functions. Solid line method with a single level
set function (using markers).
45Nucleation model
Crystals are not nucleated simultaneously. To
simulate nucleation, we use the following model
- Nucleation sites density ?, location of each
nucleation site totally random (uniformly
distributed in the domain). - Orientation angle orientation angle of each
nucleation site totally random (uniformly
distributed between 0 and 2p). - Each nucleation site becomes an actual seed iff
the required undercooling is satisfied. The
required undercooling is modeled to be a fixed
value or as a random variable. - We assume the nucleation sites fixed (do sampling
first and then run the micro-scale model
deterministically).
46Signed distance change due to nucleation
We update the signed distance function at each
node y, after a circular seed with radius R0 is
generated at location xi.
47CET transition study of Al-3Cu alloy
Follow conditions in Beckermann (2006). Relation
between microstructure and processing parameters
48Randomness effects
49Interaction between a large number of crystals
50Multi-scale modeling
51An example which requires multi-scale modeling
Material properties
4
0
3
0
2
0
Boundary conditions
1
0
Initial condition
0
0
1
0
2
0
3
0
4
0
52Computational results using adaptive domain
decomposition
Computation time 2 days with 8 nodes (16
CPUs). Cannot wait so long! Can we obtain
results in a faster way (multi-scale modeling)?
53What we can expect from multi-scale modeling
- Of course, we cannot expect microscopic details.
But - We want to know macroscopic temperature,
macroscopic concentration, liquid volume fraction.
- Microstructure features are often of interest,
e.g. 1st/2nd arm spacing, Heyns interception
measure, etc. Let us denote these features as
54Widely accepted assumptions
Assumption 1 Without convection, macroscopic
temperature can be modeled as
Assumption 2 At a reasonably high solidification
speed and without fluid flow, macroscopic
concentration constant.
Assumption 3 Microstructure depends on
macroscopic cooling history and thermal
gradient history.
Assumption 4 Volume fraction only depends on
microstructure, and temperature.
55Macro-scale model
Temperature
Liquid volume fraction
Microstructure features
First two equations coupled. Microstructure
features determined as a post-processing process.
56Relevant sample problems
Infinite number of sample problems can be
selected. How to select the ones related to our
problem of interest is the key! Use a very
simple model to find relevant sample
problems. Model M (1)
treat material as pure material (sharp and stable
interface) (2) do not model
nucleation
57Comparison of three involved models
58Solution features of model M
Define solute features of model M to be the
interface velocity and thermal gradient in the
liquid at the time the interface passes through.
59Selection of sample problems
Given any solution feature of model M, we can
find a problem, such that features of model M
for this problem equals to the given solution
feature.
Sample problem
Chose a domain (rectangle is used) with initial
and boundary condition form the following
analytical solution
60Multi-scale framework
61Solve the previous problem
Material properties
4
0
3
0
2
0
Boundary conditions
1
0
Initial condition
0
0
1
0
2
0
3
0
4
0
62Step 1 Get solution features of model M
Plot solution features of model M for all nodes
in the feature spaces
63Step 2 Fully-resolved solutions of sample
problems
64Obtained liquid volume fraction
65Use iterations to obtain temperature, volume
fraction, microstructure features
66Temperature at time 130
Data-base approach result
Macro-scale model result with Lever rule
Fully-resolved model results with different
sampling of nucleation sites.
Average
67Liquid volume fraction at time 130
Left temperature field and volume fraction
contours (0.95 and 0.05) Right volume fraction
contour on top of fully-resolved model interface
position
68Predicted microstructure features
Results in rectangle predicted microstructure
Results in the middle fully-resolved model
results Black solid line predicted CET
transition location
69Solidification of Al-Cu alloy
70Step 1 Solution features of model M
71Step 2 Fully-resolved solution of sample problems
72Periodic boundary condition for the sample problem
Top half results copied from below Bottom half
Computational domain Periodic boundary condition
to minimize effects of boundary on directional
solidification solution
73Lquid volume fraction for different
microstructure features
74Iterative process for convergence
Left half (black points) results after iter 0.
Right half (green points) results after iter
3.
75Comparison with Lever rule (temperature at
t12.7s)
Left Lever rule Right Database approach
76Microstructure in the domain
E
D
A
C
A
B
F
B
H
H
G
F
G
C
E
D
A (95mm,75mm) B (90mm,75mm) C (75mm,75mm) D
(60mm,80mm)
E (90mm,10mm) F (80mm,20mm) G (65mm,35mm) H
(50mm,50mm)
77Microstructure from side to center
A
D
B
C
A
B
C
D
Fine columnar ? coarse columnar ? Equiaxed
78Microstructure from corner to center
E
F
H
G
G
F
E
H
Fine equiaxed ? Coarse equiaxed
79Suggestions for future research
80Consider flow effects in the multi-scale model
The computationally efficient model we used to
identify relevant sample problems (with its
analytical solution) is not applicable for
problems with convection effects. Extension of
the current technique or other techniques would
be necessary to efficiently consider convection
effects in a multi-scale framework.
81Modeling fluid structure interaction in
micro-scale
In our current micro-scale model, the crystal is
assumed static after nucleation. In reality, the
crystals would move, rotate, compact and break
into fragments. Recently, there are lots of
advances in fluid-structure interaction. These
advances can be used to improve the micro-scale
model.
82Atomic scale computation
Our current micro-scale model relies on input
from phase-diagram and a few parameters to mimic
the crystal orientation anisotropy, surface
tension effects, kinetic under-cooling effects
and nucleation. Computation in the atomic scale
(not continuum any more) and related multi-scale
techniques to use atomic scale computation
results are of great significance.
83Solid-Solid phase transformation
In our current model, only liquid to solid phase
transformation is considered. After this phase
transformation, solid-solid transformation is
also very crucial to the final microstructure. M
odeling solid-solid phase transformation after
solidification and study of the properties of the
final microstructure is an open area.
84 THANK YOU FOR YOUR ATTENTION