Upload
ariel-phelps
View
221
Download
1
Embed Size (px)
Citation preview
SINDA/FUINTFor
Advanced Design
에이블맥스㈜
Page 2CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
목차
1. Introduction to Optimization2. Automated Determination of Worst Case Scenario3. Automating Model Calibration to Test Data4. Statistical Design: Reliability Engineering Module
Introduction To
Optimization
에이블맥스㈜
Page 4CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Purpose of this Class
The Solver is a high-level analysis module It is a way of tasking SINDA/FLUINT and/or Thermal Desktop It goes beyond the traditional “point design evaluation” mode of
steady-states and transients Optimize (minimize weight, maximize performance) Correlate (calibrate a model to test data) Fine the worst-case design scenario Estimate the reliability of a design Optimize the design for reliability
A related module, Reliability Engineering, is available for statistically evaluating the uncertainties in a design
Page 5CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
What Kind of Problems can The Solver Solve?
Typical tasks:– What heater power and interface pad conductance will keep the
temperature of a component within its upper and lower limits during various design cases, while requiring minimal electrical power?
– What should the flow rate and coolant loop line diameter be such that a component is kept at 20°C, while keeping the peak pumping power less than 20W?
– Given three thermal balance tests and the transient changes in between, what are the best-estimate values for the bolted joint conductance, the thermal mass of the batteries, and the dissipation rate of the transformer?
The Solver ...– Aids in preliminary design sizing and selections– Allows a user to solve for input values given desired responses
■ Any parameter can become an output variable (an unknown)– Helps optimize existing point designs– Can be used to automatically correlate (calibrate) models to test data– Automates the search for worst-case scenarios: design case definition
Page 6CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Symbols & Registers: A Quick Review
Symbols are Thermal Desktop (TD) user variables Registers are SINDA/FLUINT (S/F) user variables Symbols and registers are independent by default. Otherwise:
– Symbols may be sent to S/F as initial values for registers– In the Dynamic Mode, S/F may send back register values as new symbol
values, and might command new calculations based on those changes
Page 7CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Symbols and Registers
TD Symbols S/F Registers
Properties,dimensions,trajectories,etc.
Set points,Dissipations,K-factorsetc.
COMMONTO BOTH
Sent as initial conditions(via Case Set Manager)
Optionally sent back indynamic mode(e.g., CALL TDSETREG)
Page 8CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
The Solver in a Nutshell
SINDA/FLUINT autonomously changes one or more registers until some analysis objective, as defined by the user, is met– Registers are a “control panel” of “knobs” for centralized model changes– The Solver assumes control of one or more “knobs” on the control panel– Using the Dynamic Mode, the registers to be changed might also be TD
symbols, and updated TD calculations might be required Goal seeking
– Find an input value given a response (the “reverse problem”)– Example: find the conductivity such that T10 = 100.0 ... T→k, not k→T
Optimization– Minimize or maximize something subject to arbitrary constraints– Example: find the minimum heater power such that T10 > -20.0 at all times
Calibration to Test Data– Find the values of uncertainties that best fit available test data– Example: find the contact conductance by comparing with test data
Worst-Case Scenario Definition– Find the values of uncertainties/variations that generate the hot case, cold case ...
Page 9CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Symbols and Registers… and Design Variables
TD Symbols S/F Registers
Properties,dimensions,trajectories,etc.
Set points,Dissipations,K-factorsetc.
Sent as design variables(via Dynamic->Design tab)
Sent back indynamic mode(CALL TDSETDES)
DesignVariables
Page 10CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Optimization Terms
Objective– What is to be maximized, minimized, or meet some target value?
What distinguishes a good design from a better one? Design Variables
– What inputs can be changed as needed to achieve the objective?– Or, what are the unknowns that need to be determined?
Constraints– What are the limits on the design variables?– What criteria distinguish a viable design from a useless one?
■ perhaps some complex function of the predicted performance Evaluation Procedure
– What analytic operations (steady states, transients, etc.) are required to evaluate a given design? What calculations are required to answer these questions:■ What is the value of the objective for a given set of design values?■ Did this design satisfy or violate the constraints? By how much?
Page 11CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Quick Verbal Example
Minimize the mass of a fin (extended surface) for a given root temperature and environment by varying its length and thickness, but don’t accept a fin efficiency of less than 80%.
Objective– the mass ( = length * thickness * width * density)
Design variables– two: the length and the thickness
Constraints– one: fin efficiency > 0.80 (a response calculated by SINDA/FLUINT
and/or Thermal Desktop) Evaluation Procedure
– given a length and a thickness ...■ calculate the current mass (the value of the objective)■ find the steady state, calculate the fin efficiency (the constraint)
Page 12CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Optimization Terms: Illustrated
Page 13CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Optimization Terms: Objective
The purpose of the analysis: that which is being maximized or minimized
■ The mass (to be minimized)■ The performance metric (to be maximized)■ The error between test data and predictions (to be minimized when
correlating)■ The difference between the desired (target) value of an input
parameter and the current value (to be minimized when goal seeking)■ The maximum/minimum temperature experienced (when seeking
hot/cold design cases) Examples:
■ Minimize the mass of a fin■ Objective: the current mass (of the design being evaluated)
■ Maximize the performance of the fin■ Objective: the fine efficiency of the current design
■ Minimize the number of cycles for a thermostatic heater■ Objective: the number of cycles required by the current design
Page 14CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Optimization Terms: Design Variables
The variables that can be changed to meet the objective■ These are the unknowns for which the Solver must estimate a value
Goal Seeking:The unknown value (a traditional input variable, for example)
Optimization:Anything that needs to be sized or selected
Calibration and Worst-case Definition:The unknown value (a traditional input variable, for example)
■ Contact conductance, bond line thickness, as-built insulation performance, optical properties, natural convection film coefficient, head loss coefficient
■ Environmental temperature, humidity, orbital beta angle, trajectory, angle of attack, altitude, etc.
■ Can be applied as a unit multiplying factor to retrofit a old model The objective might itself be a design variable!
■ Example: Fine the minimum length fin that is 80% efficient■ Or the objective might be a simple function of design variables(i.e., not
a function of predicted performance)
Page 15CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Optimization Terms: Constraints
Those criteria which distinguish a valid design from a useless one
■ Despite appearances, constraints are more similar to the objective than to design variables: like the objective, constraints help guide or restrict the solution
Side Constraints■ Fixed limits on the values of design variables
■ Example: the outer diameter cannot exceed 3cm Configuration Constraints
■ More complex (formula-based) limits on the values of design variables■ Example: the outer diameter must be at least 0.5cm greater than
the inner diameter Performance Constraints
■ Arbitrarily complex, user-defined limits, perhaps requiring an S/F and TD evaluation to measure the relative success or failure of a design
■ Example: the temperature of component must never exceed 100 degrees
Page 16CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Optimization Terms: Evaluation Procedure
The method by which the user tells the Solver how suitable a design is
At the start of evaluation procedure:■ The user is given a new set of design values: a point design■ This design will never violate fixed side constraints■ This design usually won’t violate configuration constraints, and if it does
it won’t by much, but the user must tolerate a certain degree of violation
■ This design will often violate performance constraints, since they are as yet unknown (they have not yet been evaluated)
By the end of the evaluation procedure■ The user must have provided a new value of OBJECT
■ Example: set OBJECT to the current fin efficiency■ The user must have similarly updated any constraint variables
■ Example: set the current value of a constrained temperature The procedure consists of whatever calculations are needed
to update the values of the objective and constraint functions “Calculations” can be made by S/F, TD, or user (logic, expressions)
Page 17CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Calling Hierarchy
Point Design Analysis (Traditional Usage):
– SINDA/FLUINT calls …■ OPERATIONS, which calls …
STEADY (for example)
With Solver:
– SINDA/FLUINT calls …■ OPERATIONS, which calls …
SOLVER, which calls …
» PROCEDURE, which calls …
STEADY (for example)
(Green denotes user choice)
Page 18CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
A Quick Demo: Heated Bar
Aluminum bar, 1m long with 5cm x 5cm cross section, painted white (e=0.8), heated at one end (100W). Radiates to deep space. Initially at 100°C (uniform).
Normal question: How hot does the heated end get in 1000 seconds?
Page 19CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
A Quick Demo: Heated Bar (Cont’d)
Possible Solver Questions: What power causes the hot end to reach 200°C in 1000 seconds?
Goal seeking, with the design variable power, and objective is the temperature reached at 1000 seconds (goal: 200°C).
Does not need the TD dynamic mode to solve since no geometric calculations are involved
Leaving power at 100W, what cross section (width, height) causes the same result (200°C hot end at 1000 seconds)? Goal seeking again (design variable is width=height), but this time
dynamic mode is needed What is the minimum mass bar that causes the hot end to reach
200°C (and no more) in 1000 seconds? True optimization: design variables are width, height, and length.
Objective is width*height*length (to be minimized) Constraint: final end temperature < 200°C
For a realistic answer, might add more constraints on shape, such as0.5 < width/height < 20.01 < width/length < 1
Page 20CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Heated Bar Goal Seeking(TD Demo)
Assuming height=width, what width causes the bar to reach 200°C in 1000 seconds?
Design variable: widthEvaluation procedure: transient simulationConstraints: noneObjective:
Method 1: Run transient for 1000 seconds.Objective=end temp., Goal=200
Method 2: Run transient until end reaches 200°C.Objective=end time, Goal=1000
Either way, the Solver changes widthuntil the Objective equals the Goal
Page 21CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Explaining Optimization: The Hill Climbing Analogy
The Solver is like a hill climber, trying to find the highest (or lowest) point on a hillside covered with fences and potentially rough terrain
OBJECTIVE■ A hill to be climbed to find the summit (if maximizing)
Design Variables■ The map coordinates the climber may follow■ If there are N design variables, the “design space” is N dimensional
Constraints■ Fences that may be encountered along the way; there may be many■ These often prevent the summit from being reached
Evaluation Procedure■ Altimeter: what is the current elevation?
The climber is nearly blind: it cannot see very far and must stop often and test the local shape of the hillside, and the shape of any fences it finds.
Page 22CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Hill Climbing, Illustrated
Page 23CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
How the Solver Works
Finding a search direction■ Perturb each of the N design variables
■ Requires N evaluations of the objective and the constraints■ Find the shape of the local “terrain” by calculating the partial
derivatives of the objective and all active constraints to each design variable
■ If at an infeasible point, go towards the nearest feasible region■ Otherwise, initially head in the direction of steepest ascent
Searching■ Move linearly along the chosen search direction until:
■ Either a new constraint becomes active■ Or a local maximum (“ridge”) is found
■ Typically requires 3 to 10 more evaluations Perturb and fine a new search direction (varies by method
selected) New direction may be nearly tangential to a constraint, but is usually not
in the direction of steepest ascent even if no constraints are active
Page 24CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Summary so Far(for reference)
SOLVER is a high-level solution routine that may be called from OPERATIONS When called, SOLVER does the following:
Changes one or more design variables subject to limits and user constraints until OBJECT is as close as it can be to GOAL Design variables are a subset of registers Registers may also be TD symbols
To evaluate a particular design (e.g., a set of design values), Solver calls updates the model and then calls PROCEDURE
PROCEDURE: Arbitrary set of instructions, including BUILD/BUILDF statements and calls to solution
routines By the time PROCEDURE returns, the user is expected to provide fresh values of
OBJECT and any named constraint variables To pass design variables back to TD and update: CALL TDSETDES Then, to refresh the current case in TD before an S/F solution: CALL TDCASE After calculating OBJECT, send back to TD for display: CALL TDOBJ More detailed controls and communications are available: see TD manual
Anything that can go in OPERATIONS can go in PROCEDURE Usually OBJECT and constraint variables are updated at the end of PROCEDURE, but
they may be calculated initially, or cumulatively, or as part of an expression, etc.
Page 25CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Design Space Scanning
Explore design space (range of possible design variable values) and/or find good starting point for Solver All utilities return best values found (per objective and
constraints) Each design variable must have both upper and lower limits for
these utilities DVSWEEP
Parametric sweep of a single design variable. DSCANFF
Full factorial scan of multiple design variables. Example: 4 design variables sampled 3 values each (low,
medium, high): 3*3*3*3=81 calls DSCANLH
Latin hypercube scan of multiple design variables: each design variable uniquely sampled NLOOPO times
Recommended prescan for Solver
Page 26CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Example Latin Hypercube
Two dimensional design space with sampling resolution of five In other words: two design variables A and B ranging from AL to AH, BL to
BH, and NLOOPO=5 Two possible samplings (DSCANLH chooses randomly each time it is
called)
Each parameter is therefore sampled NLOOPO times: it doesn’t cost much to try different values of a variable that doesn’t make much difference
Page 27CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Prescanning: Find a good starting point
Call DSCANLH before SOLVER in OPERATIONS
NLOOPO = 20 $ test each design variable at NLOOPO values CALL DSCANLH $ prescan! NLOOPO = 200 $ allow Solver more procedures to do its work CALL SOLVER $ optimize!
Helps avoid various problems, including poor initial conditions Almost mandatory for model calibration and worst-case scenario seeking Cost of prescan usually pays for itself in reduced Solver work
However … Requires each design variable to have both upper and lower limits Equality constraints or tightly limited constraints can yield no valid
solution Use OFFCST call to temporarily disable selected constraints during the prescan Or use temporary tolerancing on tight constraints during the prescan (see
manual)
Automated Determination of Worst Case Scenario
에이블맥스㈜
Page 29CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Automating Each Phase of the Design Process
Page 30CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Worst-case Design Scenarios
The first step in a design process is to identify the worst-case scenarios The design will be developed and tested against these scenarios:
their revision often forces a design change. For thermal: one “hot case” and one “cold case” as a minimum
Currently: Margins and uncertainties are stacked up Conditions that can’t possibly happen or co-exist (e.g., BOL
properties combined with EOL dissipations, or steady-state at the subsolar point or within a planetary shadow)
This is called the “Coffin Corner” approach It is often unclear what stack-up or combinations yield the
worst case, especially with articulating components and complex dissipation
Page 31CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
The Problem
Despite the criticality of the results, cost of searching for the worst case scenarios can be prohibitive The number of cases grows geometrically Most older software does not facilitate repeated runs nor take
advantage of previous solutions In complex missions, the search must be repeated many times
during design development Approaches are informal (since no standards exist) and rarely
efficient. Common approaches: Full factorial (FF) search (all possible combinations of discretized
uncertainties) Monte Carlo (MC) search (hundreds to thousands of randomized
samples: a “shotgun” approach)
Page 32CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
New Technology
Parametric Software and APIs (e.g., MS Excel) Repeated runs can be scripted and searches automated Special effort spent minimizing recalculation costs
Latin Hypercube (LH) Scan : SINDA/FLUINT DSCANLH Requires fewer samples than full factorial or Monte Carlo
NLP (Gradient-based Optimization) Search: SINDA/FLUINT Solver
Directly seeks the worst case with minimum evaluations Hybrid LH/NLP Method Future: Elimination of search-then-design; the elimination of
worst-case scenarios altogether
Page 33CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Demonstration Problem
Simple Sample Problem 3-axis stabilized LEO (300km) nadir-facing box 2-axis tracking solar panel on leading side (+X) 1-axis scanning (+/- 30o) paraboloid dish on trailing side (-X) 60W “payload” with 600W 10minute pulse on the +Z face SPV/CPV NiH “battery” on the –Z face, realistic
charge/discharge/trikcle-charge profiles vs. shadow +Y and –Y faces are fully utilized as radiators Thermal Desktop model available upon request
What is the hot case beta angle, dish position, and start time for the power pulse?
Page 34CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Sample Problem Definition
Page 35CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Sample Problem Definition
Page 36CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Tools Used
Thermal Desktop/RadCAD for thermal/radiation model 15 orbit points, steady state plus 2 transient orbits per evaluation
for cyclic convergenceThermal Desktop “Dynamic Mode:” SINDA/FLUINT commands changes and recalculations as geometry/orbits change Per SINDA/FLUINT statistical analysis and optimization routines:
DSCANLH and SOLVER Total time to evaluate one case (all radiation and conduction
recalculations, steady/transient simulations): 45 seconds on a 2.2 GHz Pentium 4.
Page 37CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Full Factorial Scan (4x3x4=48 evaluations)
4 beta angles: 0, 30, 60, 90 3 scan angles: -30, 0, 30 4 pulse start times: 0, 1600, 3200, 4800 sec. from sobsolar
point
Page 38CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Latin Hypercube Explained
For N samples made, each parameter uniquely sampled 1/N times
For 2 variables A and B, if N=5:
Page 39CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Latin Hypercube Results
N=20 Samples (usually <20% of FF method) Example: b = 2.25, 6.75, 11.25, … 87.75
Found hotter temperatures in less evaluations: (note: this method is statistical: your mileage will vary slightly.)
Page 40CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Nonlinear Programming (optimization) Approach
Instead of “What is the best design” ask “What is the worst case?” Best design: vary A, B, C to minimize cost Worst case: what combination of A, B and C yield the maximum
temperature (hot case)?Good news: finds the worst point, not just nearby point
Bad news: sensitive to initial conditions Number of evaluations unknown (usually 20 to 100) Requires one search per component Might ‘stall’ at a local minimum
This isn’t serious for design optimization, but is more troublesome for test data calibration and acute for worst-case seeking
Page 41CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Using the Solver for Worst-case Seeking
Uncertainties Use design variables as environmental, mission uncertainties Easy to retrofit to a model as unit multiplying factors
Objective Maximize and/or minimize the design concern, for example:
Maximum temperature excursion Maximum heater power required Minimum battery power remaining
Evaluation Procedure Whatever solutions or series of solutions to stress design
The final set of such conditions is what is sought Can be just SINDA/FLUINT, or can include Thermal Desktop
calculations (e.g.: fine the worst-case beta angle)
Page 42CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Hybrid Method
Find good starting point with quick (say N=10) LH scan Finish off with NLP (Solver) Overcomes both initialization sensitivity of NLP and
discretization limitation of LH. The cost of LH “prescan” usually pays for itself in reduced NLP evaluations
Page 43CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Results Discussion: Sample Model
Results Battery peaked at intermediate beta angle: too low and the –Z
face doesn’t get much sun, too high and the battery isn’t used Payload peaked at fuller sun (high beta, but less than 90!) and
when pulse began near the subsolar point In retrospect:
Beta angle was the most important Pulse start time was of intermediate importance Scan angle for the dish was not important
FF and MC waste time resolving unimportant parameters. Discrete sampling like LH preserves the resolution of important parameters.
Page 44CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Battery’s Hot Case
Page 45CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Payload’s Hot Case
Page 46CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Conclusions
Existing statistical analysis and optimization tools can significantly reduce the cost (and improve the accuracy) of worst-case searches
Just like model calibration to test data, another nasty task has been automated
Automating Model Calibration to Test Data
에이블맥스㈜
Page 48CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Model Calibration to Test Data (aka “correlation”)
Thermal/fluid models are dominated by uncertainties■ Contact conductance■ Natural convection coefficients■ Surface optical properties■ Pressure losses in complex components■ As-built insulation performance
Calibration is critical■ Thermal/fluid models are often just intelligent extrapolations of known
test data to untestable conditions
Yet without the Solver, test data correlation is:■ Extremely time consuming■ Unstructured and informal
■ No procedures, methods, tools, or acceptance criteria■ Little mathematical basis for resulting “correlation”
■ Universally despised by analysts
Page 49CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Calibration as “Optimization?” Yes: Minimize Error
The Solver can find the best design…■ What dimensions, materials, etc. result in the least mass, the best
performance, etc.?
Or it can find the best model of a fixed design…■ What values of uncertainties result in the best fit to test data?
User has complete control over:■ The uncertainties to vary
■ By how much, in what order (if not simultaneous)■ The comparison procedure
■ A single steady state run■ A single transient run (end-point or cumulative comparison)■ A complex series of runs or test cases
■ The determination of a best fit■ Least squares, minimized maximum error, etc.
Page 50CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Using the Solver for Model Calibration
Correlation Parameters■ Use design variables as correlation parameters■ Choose the degree of uncertainty in each, apply as limits■ Easy to retrofit to a model as unit multiplying factors
Objective■ Minimize the difference between test and predictions■ Least squares or RMS: minimize the square root of the mean of squared
differences■ MINIMAX: minimize the maximum deviation
Comparison Procedure■ Whatever solutions or series of solutions to produce comparison■ Comparison can be singular (At the end of the solution), or cumulative
(accrued during the solution), weighted based on relative importance of comparison points, etc.
■ Can be just SINDA/FLUINT, or can include Thermal Desktop calculations (e.g.: find the best-fit emissivity)
Page 51CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Quick Verbal Example
Consider a fin (extended surface) heated at the root with a specified room air temperature, cooled by a fan
■ Root temperatures are known at three different base heater settings
What single convection coefficient best explains all test data points? Objective
■ Minimize the least square error between measured and predicted root temperatures:
OBJECT = SQRT ( ( Tm,1 – Tp,1)2 + (Tm,2 – Tp,2)2 + (Tm,3-Tp,3)2 )
“Design Variables” (aka Correlation Parameters)■ One: the convection coefficient “HCONV”
Constraints■ None (other than maybe lower and upper limits on HCONV)
Evaluation Procedures■ Given a current value of HCONV:
■ Run three steady-state at each of the three power levels to predict root temperature
■ After each steady-state, sum the squared error in OBJECT (see above)■ After all three solutions are completed, take the square root of OBJECT
Page 52CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Correlation Support Routines
Correlation involves handling and comparing a lot of data COMPARE
■ Compare two arrays (one test data, one predictions)■ Generate a report, or return values for the Solver
PREPLIST■ Prepare a COMPARE array given a list of node/lump/path IDs
PREPDAT1■ Prepare a COMPARE test data array from interpolations of singlet or
bivariate arrays
PREPDAT2■ Prepare a COMPARE test data array from tabular data (such as might
come from Excel or an electronic data logger)
Page 53CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
TD Demo:Find the Properties of a Rod Given Axial Temperature Profile
A 1m long, 1cm diameter rod of unknown conductivity and surface emissivity is used to suspend a 40K LH2 tank from a 300K chamber wall
The radiation environment can be considered an effective “sink temperature” but its value is unknown
Test data is taken for the temperature profile along the rod■ Miraculously, these points correspond to the node locations when the resolution is 20
and edge nodes are used:
What is the rod conductivity, surface emissivity, and effective chamber radiation temperature that corresponds to (best explains) the above temperatures?
Page 54CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
TD Demo:Using MINIMAX
Define “errmax” on Symbol Manager
Page 55CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
TD Demo:Using MINIMAX
Page 56CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
TD Demo:Using MINIMAX
Page 57CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
TD Demo:Using MINIMAX
Statistical Design: Reliability Engineering
Module
에이블맥스㈜
Page 59CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Reliability Engineering: Introduction
Treat uncertainties and unknowns statistically, not deterministically.
What is the resulting reliability?
Page 60CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Symbols and Registers… and Random Variables
TD Symbols S/F Registers
Properties,dimensions,trajectories,etc.
Set points,Dissipations,K-factorsetc.
Sent as design variables(via Dynamic->Design tab)
Sent back indynamic mode(CALL TDSETRAN)
RandomVariables
Page 61CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Using Reliability Engineering
Identify which registers are to be used as random variables These are the uncertainties in dimensions, environment, properties, etc. Define the way in which each random variable varies: its distribution
function Uniform: upper and lower bound only Normal (Gaussian): mean and standard deviation (or coef. of variation) Array: table of probability vs. value (truncated Gaussian, triangular “Witch’s Hat,”
Weibull, Chi-square, log normal, test data, etc.) Optionally, provide a list of arbitrarily complex reliability constraints
(responses and the limits on those responses that define failure: failure limits is perhaps a better term) Example: “The design fails if this temperature exceeds 100°C. Will it
happen?”Paraphrasing: “What are the chances that this temperature will exceed 100°C?”
These are cheap. Add lots of them, or use postprocessing for hindsight. Provide an arbitrarily complex evaluation procedure
Use any solution within SINDA/FLUINT to calculate how well any one design instance behaves relative to the reliability constraints
Reliability estimation routines perturb the random variables to determine the chances of each reliability constraint being violated
Page 62CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Using Reliability Engineering, Cont’d
Three different methods are available to predict reliability: SAMPLE: Monte Carlo sampling. (Typical: 1000 evaluations)
Simple random perturbations. Slow but powerful. DSAMPLE: Descriptive sampling. (Typical: 100 evaluations)
Breaks distributions into NLOOPR equal chunks. Much faster than SAMPLE, but was NLOOPR enough?
RELEST: Gradient method. (Typical: 5 to 10 evaluations) Makes a few assumptions: not applicable in all cases. Much faster than DSAMPLE, but were approximations OK?
Routine SAMPLE DSAMPLE RELEST
Method Monte Carlo Sampling Descriptive Sampling Gradient Method
Speed Slow Intermediate Fast
Convergence Detected? Yes No No
Fixed Execution Cost? No Yes Yes
Finds Overall Reliability?
Yes Yes No
Cumulative? Yes Somewhat No
Applicability? Unlimited Unlimited Limited. Assumes: - Gaussian random variables - Continuous and linear responses - Fixed failure limits
Page 63CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Data Flow for All Three Methods
Evaluation Procedure (user provided)
SAMPLE,Given these values of design variables,perform SINDA/FLUINT analysesor other calculations to determine: - the values of any constraints (if any)
New values of random variables
current values of reliability constraints
DSAMPLE,RELEST
Reliability Calc.(RELEST only)
Done
Convergence?(SAMPLE only)
Page 64CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Example Latin Hypercubes (DSAMPLE method)
Two dimensional variational space with sampling resolution of five In other words: two random variables A and B ranging from AL to AH, BL to
BH, and NLOOPR=5 Two possible samplings (DSAMPLE chooses randomly each time it is
called)
Each parameter is therefore sampled NLOOPR times: it doesn’t cost much to try different values of a variable that doesn’t make much difference
Page 65CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
A Quick Demo: Heated Bar (TD Demo)
Aluminum bar 1m long 5cm x 5cm cross section Painted white (e=0.8)
Radiates to deep space
Heated at one end (100W)
Normal Question: How hot does the heated end get at steady state? Nominal steady hot end
temperature is about 115°C
Page 66CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Example Statistical Problem (TD Demo)
What if some properties or boundary conditions weren’t known exactly?
Uncertainties: Width(=height): 5cm ± 0.1cm (= 1s,
normal distribution) Emissivity: Any value between 0.75 and
0.85 equally Power: 100W ± 5W (= 1s, normal
distribution)
New Question: What are the chances the hot end will be less than 100°C?
0.75
pro
babili
ty
5cm
pro
babili
ty
0.1cm
100W
pro
babili
ty
5W
0.85
Page 67CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Example Problem, ResultsRANTAB (mostly useful to echo inputs)
NUMBER OF RANDOM VARIABLES = 3 NUMBER OF REL. CONSTRAINTS = 3 RELPROCEDURE CALLS (LOOPCR) = 0 VS. MAXIMUM (NLOOPR) = 100 OVERALL RELIABILITY TALLY = -1.00000 (UNAVAIL.) TOTAL CUMULATIVE CALLS = 0 NSEED = 94740687, AERRR = 1.000000E-05, RERRR = 1.000000E-03, LAST ROUTINE: NONE
RANDOM VARIABLE TABULATION
NAME TYPE MEAN STD DEV COEF VAR LOWER LIM 1% 50% 99% UPPER LIM
WIDTH NORMAL 5.00000E-02 1.00000E-03 2.00000E-02 -1.00000E+30 4.76737E-02 5.00000E-02 5.23263E-02 1.00000E+30 EMISS UNIFORM 0.80000 2.88675E-02 3.60844E-02 0.75000 0.75100 0.80000 0.84900 0.85000 POWER NORMAL 100.00 5.0000 5.00000E-02 -1.00000E+30 88.368 100.00 111.63 1.00000E+30
RCSTTAB (main output routine)
NUMBER OF RANDOM VARIABLES = 3 NUMBER OF REL. CONSTRAINTS = 3 RELPROCEDURE CALLS (LOOPCR) = 100 VS. MAXIMUM (NLOOPR) = 100 OVERALL RELIABILITY TALLY = 0.03000 ( 3.000%) TOTAL CUMULATIVE CALLS = 10 NSEED = 94740687, AERRR = 1.000000E-05, RERRR = 1.000000E-03, LAST ROUTINE: DSAMPLE
RELIABILITY CONSTRAINT TABULATION
NO. NAME MEAN STD DEV COEF VAR LOWER LIM REL: TALLY REL: NORM UPPER LIM REL: TALLY REL: NORM
1 UNNAMED 114.80 8.2696 7.20329E-02 130.00 0.96000 0.96695 2 UNNAMED 114.80 8.2696 7.20329E-02 115.00 0.51000 0.50953 3 UNNAMED 114.80 8.2696 7.20329E-02 100.00 3.00000E-02 3.67271E-02
3% chance by tally
3.7% chance by fit to Gaussian curve
Page 68CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Example Problem, Histograms (EZXY Demo)
■ EZXY also allows the results to be queried in hindsight
Page 69CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Reliability Engineering for Users of the Solver
Like the Solver, it is an Advanced Design Module
High-level design solutions beyond steady state and transient
Many parallels to the Solver even re-uses a few of the control parameters like NERVUS easier to learn and use
Can be used instead of the Solver Can be used with the Solver: Reliability-based Optimization
Page 70CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Similarities Between the Solver and Reliability Engr.
Page 71CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Nesting Reliability Engineering within the Solver (Advanced)
Page 72CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
High-level Operations:Robust Design (Advanced)
Reliability engineering plus optimization: design for reliability What is the best design that is at least 99% reliable? What tolerances are acceptable? Avoid both under-design (risk) and over-design (cost)
Example: Reinterpreting the intent of MIL-STD 1540d Traditional approach: Keep the junction temperature under 125°C
(qualification) Apply 10°C margin (acceptance) plus 11°C (analysis/environ.
uncertainty) Stack up worst case dimensions, properties, environments, and then
hope might use the Solver to help find the worst cases, then again to synthesize
such a design applying 21°C margin as an optimization constraint might use Reliability Engineering to evaluate such a design
Robust Design approach to the same problem Apply 10°C margin (acceptance) as a reliability constraint (failure limit) Convert the 11°C into uncertainties in as-built properties, orbit, etc. Maximize performance, minimize weight using the Solver Use the reliability allocated to the thermal subsystem as an
optimization constraint
Page 73CAE ENGINEEREING TOTAL SOLUTION http://www.ablemax.co.kr
Robust Design, Illustrated
감 사 합 니 다 .CAE TOTAL SOLUTION 에이블맥스 ( 주 )
Page 75CAE TOTAL SOLUTION www.ablemax.co.kr
A Quick Demo: Heated Bar
Aluminum bar, 1m long with 5cm x 5cm cross section, painted white (e=0.8), heated at one end (100W). Radiates to deep space. Initially at 100°C (uniform).
Normal question: How hot does the heated end get in 1000 seconds?
Page 76CAE TOTAL SOLUTION www.ablemax.co.kr
A Quick Demo: Heated Bar – Goal Seeking
What power causes the hot end to reach 200°C in 1000 seconds? Goal seeking, with the design variable power, and objective is the
temperature reached at 1000 seconds (goal: 200°C). Does not need the TD dynamic mode to solve since no geometric
calculations are involved