Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

Lecture 4

Identification, integer problems and topology optimization

Engineering design optimization


Lecture 4: Identification, integer problems and topology optimization

PSpreadheet demo revisited P Identification problems P Integer problems


< Branch and bound < Simulated annealing < Genetic algorithms

PTopology optimization PODESSY topology demo

To read:

PFrom my homepage, www.ime.auc.dk/~jr/lecturenotes.htm:


< Artikel om heltalsprogrammering < Note om topologioptimering

PVisit the topopt home page: www.topopt.dtu.dk

Exercises:
John Rasmussen, Institute of Mechanical Engineering, Aalborg University, 2001

PDefine your own topology optimization problem and solve it in ODESSY

Identification
- a useful application of optimization technology

Demo
Dysfunctional spreadsheet optimization from lecture 3

Identification can be used to find unknown properties of a product, a process, a material, or any other physical or algorithmic structure. It is a very useful technique.

Identification example
- to identify material parameters
Material parameters are often difficult to measure by non-destructive testing. Suppose you have a problem with, say, vibrations in a composite material structure and wish to make a finite element model, but you dont know the material properties. Finding them by tests would mean destroying the structure, and subsequent analysis would be pointless.

Identification (contd)
- to identify material parameters
We start from an initial gues of material properties. The frequencies are computed and subtracted from the measured frequencies, and a 2-norm is evaluated. Sensitivity analysis reveals which way the properties should be changed to reduce the difference. Optimization changes the properties, and the process is repeated until convergence.

Material properties
Measured freq. - Comp. freq.

Difference(mat. prop) Minimize Non-destr. test

FEM

FE analysis Difference 3(fj - fj*)2 FE model

Optimization

If the correct material properties are found, then the difference will be exactly 0. Minimized differences larger than 0 indicate model inaccuracy.

Identification of unknown loads Loads on a bicycle frame


- identification of unknown loads
Loads can sometimes be difficult to measure. This is the case for wind loads on a bridge, or pedal loads on a bicycle. The latter require very expensive instrumented pedals. Another option is to measure other properties and convert them to loads by optimization/identification techniques. Strain gauges are an easy and reliable way to measure material strain.

- in a bicycle frame
We define a residual as the squared sum of differences between measured and computed strains:

To minimize, we need the gradient:

An ordinary bicycle made of steel tubes can be modeled very easily and very reliably with beam elements in a finite element system. Deflections and strains away from tube joints will be determined with good accuracy. A bike model with variable loads and a real bike with strain gauges can be made to have the same strain by varying the loads on the model.
John Rasmussen, Institute of Mechanical Engineering, Aalborg University, 2001

In linear elasticity, the strains are proportional to the loads, so the derivative of the strains w.r.t. the loads can be computed once and for all.

Identification
- a useful application of optimization technology

Beware that
P You need more experimental data than the number of parameters you want to identify. Otherwise, your solution will not be unique. P If the residual is minimized to a small number, then you have found a solution with good accuracy. PMany problem, such as the bicycle load identification, can be formulated and solved analytically. This may provide more knowledge about the solution. P If you run into convergence problems, then it is likely that the dependency between your measurements and your parameters is not strong enough. Find a better experiment.

P To identify material properties P To find unknown loads on a system P Infinitely many other useful ideas such as < Finding optimal process parameters from experimental observations of the quality of products < Finding the algorithm behind an unknown control system < Development of algorithms for image analysis from sample pictures <. <.

Integer vs. continuous problems Integer problems


- are optimization problems where the variables are only allowed to take on a small number of discrete values. Only discrete combinations of the design variables are allowed in integer problems.

P Integer problems occur frequently in


< economics < ressource and production planning < logistics < scheduling < design with components or according to codes < design of steel structures, frames, trusses, bridges, masts, etc. < conceptual design

x2

This may seem like an advantage: we go from an infinite number of solutions to a finite number. However, the finite number of solutions is usually still enourmously large.
x1

P Integer problems are mathematically and algorithmically challenging

Size of a simple design space


- when working with discrete variables

Integer problems
- have awful mathematical properties

PDisjoint design spaces PNo differentiability PNo convexity PNo road from one point to another PNo check for optimality

11 cross-sectional areas 5 tube types to choose from 511 = 49 million combinations


John Rasmussen, Institute of Mechanical Engineering, Aalborg University, 2001

Integer problems
Solution method 1: continous relaxation
Solve the problem with continuous variables. Round the solution off to the nearest integer point. Investigate the neighbourhood of integer points for better solutions. This does not always work. It usually fails if there are few integer values to choose from.

Integer problems
Solution method 2: branch and bound

x2

x2

Step 1: Solve the problem disregarding the integer constraints.

x1

x1

Integer problems
Solution method 2: branch and bound
x2

Integer problems
Solution method 2: branch and bound
x2

x1

x1 x2 x2

Step 3: Branch the remaining solution. If only feasible, integer solutions left, keep the best one as the final solution to the problem.

Step 2: Branch the problem into two sub-problems. Keep the feasible solutions.

x2

x2

x1

x1

No feasible solution

Feasible solution with one integer value.

x1

x1

Feasible, integer solution Feasible, integer solution

Integer problems
Solution method 2: branch and bound
The branches keep multiplying as long as there are non-integer solutions present.
Branch 1-1 Branch 1 Start problem Branch 2 Branch 2-2 Branch 1-2 Branck 2-1

Integer problems
Solution method 3: simulated annealing
PSimulated annealing is based on the probabilistic arrangement of atoms into crystals in an annealing liquid. P If the liquid is cooled slowly, the atoms have a good chance of finding their right positions, i.e., the one that leads to minimum energy of the solidified liquid. P If the liquid is cooled faster, a less optimal configuration with interior stresses is reached. PSimulated annealing is able to solve integer optimization problems by this analogy. The global optimum is only found if the cooling is very slow, i.e., many iterations are used. PMany thousands of iterations are frequent in simulated annealing algorithms. P Implementation and use is easy - no sensitivities required. But convergence is slow.

A branch can be cut, if there is no feasible solution to the problem. A branch can be cut if another branch with an all-integer solution has a better objective function. This is because the objective function never gets better when you apply more constraints.

John Rasmussen, Institute of Mechanical Engineering, Aalborg University, 2001

Integer problems
Solution method 4: genetic algorithms
PGenetic algorithms simulate the development of species and survival of the fittest. PThe fittest combinations of design variables survive and are combined to form new designs PConcepts such as mutation and crossover are used to generate freaky designs that are often poor but in rare cases very good. This way, the algorithm can leap from one local minimum to another - from a monkey to a human being. PGenetic algorithms are sorrounded by an aura of magic, and they have yet to prove that they are better than random searches for general problems.

Natural structures
... are somehow richer than what we imagine can come out of a design optimization process

They can be very complex solids...

Natural structures (contd)

Natural structures (contd)

... or something in between. ... or beautifully designed frames and membranes... Is the microstructure shape or material? Perhaps we need more general design models to be able to optimize our way to designs like these...

Topology optimization
The ground structure approach: everything composed of elementary particles.

Topology optimization
... can lead to much better solutions than shape optimization

Design space

Shape optimization
Volume fraction = 40%

Topology optimization

... the most general design space imaginable


John Rasmussen, Institute of Mechanical Engineering, Aalborg University, 2001

Topology optimization
... basic idea

Topology optimization
Problem formulation

Continuous relaxation
We cannot solve the integer problem, so we try a continuous relaxtion and hope it brings us close ot the optimum.

Discretization
We would like a finite number of variables to work on rather than a density function over the domain.

Choice of objective function


It turns out that we can only solve the problem for some special objective functions. One of them is the total elastic energy of the structure.

Elastic energy as objective


This is closely related to stress minimization. For some examples, the shape and topology optimization methods give the same results, and we can compare them to stress minimization
Stiffness Stress

John Rasmussen, Institute of Mechanical Engineering, Aalborg University, 2001

What is the density?


In 2-D structures, we can interpret the density as a thickness, so that the stiffness of each element is proportional to the density.

Density as thickness
Consider the famous fillet problem with density interpreted as thickness (Bendse & Kikuchi):

Power-law stiffness Density as thickness


This generates a fuzzy solution with density in between 0 and 1 in the whole structure. This is not an integer solution! If we assume a nonlinear relationship between density and stiffness, then we can make is disadvantageous to use intermediate densities. Only densities 0 and 1 provide economical stiffnesses

Bendse & Kikuchi

Power-law stiffness
Using a power-law stiffness with no other measures will separate the structure into black and white, but it will tend to generate fibre materials as much as the discretization allows. It only converges if we use composite material properties in the model.

ODESSY topology demo


Using ODESSY to design a bridge
100

30 Another option os to add a constraint on the boundary length. This will reduce the number of holes (end fibres) possible and lead to 0-1 solutions for an element model of finite size. ODESSY can do both. 50
Thomsen

John Rasmussen, Institute of Mechanical Engineering, Aalborg University, 2001

Topology optim. on www


Ole Sigmund & Co, TheTechn Univ. of Denmark, http://www.topopt.dtu.dk

Topology optimization briefly


Designrum - Vrdigrundlag - Begrnsninger - Analyse

P Design space: the widest possible P Analysis: limited to weight, volume stiffness, strength, stability, termal properties, etc.

Input

Desired output

This is a very prospective but still immature technology

Exercises
.. for lecture 4/10

Download bridge.acd from my homepage and run it on your computer:


ods_femo bridge 0 0 0 oto_init bridge oto_fem bridge 2 10

Change the definition to a new problem of your own choice and try it out.

To read:

PFrom my homepage, www.ime.auc.dk/~jr/lecturenotes.htm:


< Artikel om heltalsprogrammering < Note om topologioptimering

PVisit the topopt home page: www.topopt.dtu.dk

You might also like