Digital Design and Manufacturing Lab: Research

Curating Complex Datasets for Machine Learning Applied to Flexible Assembly Design

• Project Period: 2021 - 2024
• Sponsor: National Science Foundation (NSF)
• Investigators: Jami Shah (Lead), Yannis Korkolis ISE (Co-PI), Joseph Davidson, ASU (Co-PI)
• OSU staff/students: S. Ramnath, A. Adrian

Research Questions:

• Can Deep Machine Learning algorithms, such as ANNs, be formulated at a fine feature resolution to distinguish good and bad design characteristics of complex engineering artifacts?
• Can ANNs be architected to handle multi-stage decision processes?
• Can trained and validated ANNs support design exploration or optimization better than traditional methods, such as Response Surface Methodology?

Scope:

The primary emphasis of our research is on data curation for advancing research in data science for complex engineering systems. Data curation must take into account not only the modality and granularity of data needed for ML, but also the volume of each data set needed for training purposes. In addition, the level of data heterogeneity (balance) needs to be considered to allow algorithms to discriminate between characteristics of good and bad designs.

Major Tasks:

• Set up multi-stage simulation pipeline to include material characterization, component stamping and springback in explicit FEA, two-part welded assemblies (end-end, T-jt)
• Automate simulation pipeline
• Deploy simulation pipeline to produce i/o data for component stamping and 2-part assembly
• Devise math methods to extract key parameters from simulation data
• Investigate suitable data modality, granularity, heterogeneity, and volume for processing in parallel through federated ANNs, or individually by distributed machine learning networks,
• Publish data sets for each process on Amazon Cloud, each including appropriate subsets of mathematical data representations of artifact geometry abstractions, decompositions and partitions, in association with parametrically reduced performance metrics

Simulation Pipeline:

 Stamping Spring back:

• To generate training data for dimensional variations in stamped components spring back explicit FE simulations are being performed and validated, using various nonlinear elastic-plastic models representative of Dual Phase steels used in automotive frames
• Tooling is adapted from NUMISHEET Benchmarks, which also serves to validate our simulations
• Given the need for quantity and variety of training data, Python scripts will be developed to encode and parametrically vary tooling geometries to conform to component shape and size variants
• Simulation inputs: geometric parameters of each component, scalar parameters of the material models and tooling geometry parameters; potentially 12,600 data sets
• Output parameters: cross-sectional and transverse variations due to spring back, extracted with the aid of scripts

Two Part Sub-assembly:

• Spring back studies described above will produce free-state variations (size, form, orientation, position) with respect to a common datum
• Automate generation of training data for stages of assembly, a generalized input structure is proposed (assembly graph, free-state of spring back components geometry, clamp and joint locations)
• Simulation will produce the free state of the assembly
• Additional data will be extracted from the results for use in distributed neural nets
• The 12600 component deformed geometries obtained from stamping simulations will be grouped into compatible categories (straight-straight, straight-curved; straight-T) and ends labeled as suitable for lap or T based on shape. Additionally, nominal sizes need to be matched
• We will construct a DOE factorial table to pre-determine which Stage I component assemblies are to be simulated

Envisioned Neural Network Pipeline

The high complexity of engineering FEA and CAD geometric data is not conducive to be processed by a single ANN in their entirety. Instead, we envision a series-parallel distributed network that process abstractions, decompositions, partitions of each data sample into sub-sets and have subsets processed in parallel through federated ANNs, then pool the results to correlate the full sample inputs to relevant responses – a process we label as a distributed machine learning network.

Publishing Datasets for Advanced Data Science in Engineering

• As opposed to random unrelated object data sets (animals, furniture, etc.) that are found on the Internet today and which have no design performance attributes, our project is proposing to produce alternative design configurations for particular design objectives – inter-related, technologically verified sets with simulation
• Extant approaches to machine learning for complex spatial-geometric sets have not been successful because the level of resolution needed cannot be achieved by feeding data to a single algorithm. The research questions cannot be answered without such curated datasets
• There are five major aspects of data curation: application relevance, volume of data, modality, granularity and heterogeneity. The first two are self-explanatory
• Modality: Different forms of data (e.g., geometry of mechanical components can be represented as full 3D BRep, as multiple 2D images from different angles, as point clouds, and so on)

• Granularity: Refers to the resolution or level of detail in the data models. We need to find the right balance between size and meaningfulness of the data for ML

•  Heterogeneity: This refers to the need for having enough variety, a balanced representation of feature variants for effective training of the algorithm

• Relation between data and ML algorithms (common known cases):

Fuzzy Feature Recognition from Topology Optimization Meshes to Match Parametric CAD Models of Hoods

• Project Period: 2021 - 2022
• Sponsor: Honda Development and Manufacturing
• Investigators: Jami Shah
• OSU staff/students: S. Ramnath, K. Witek

Overview

Primary Goal: Support conceptual design of hood frames by finding closest match from 10,000 parametric CAD models to user generated Topology Optimization (TO) frame

Benefit: Starter designs will help getting off to a faster design  while retaining important load path information  obtained from TO

Major Tasks

• Generate fuzzy data set of hood frames:
  • From representative hood skins with variety of load cases
  • From frame designs in current frame DB
• Devise new feature recognition techniques for fuzzy TO  set to find closest match
• Train ML algorithms to match fuzzy TO to hood frame data set
  • Compare supervised vs unsupervised
  • Find optimal model modality and resolution
• Validate and test ML
• Implement simple UI to execute the process

Hood Dataset

This set was generated in a previous project for a totally different purpose

• Current data set contains ~ 11,000 designs
• 10 skins and 11 hood  feature patterns
• Each combination of skins  and features has 100  design variants
• Represented as parametrized CATIA models that can be edited by the user

Generate fuzzy dataset of hood frames:

• From representative hood skins with variety of load cases
• From frame designs in current frame DB

Skin based TO Domain	Frame based TO Domain
Produces Novel Designs	Limited by Pockets and Features
Poor match between TO Results and Existing Hoods	Match between TO Results and Existing Hood Models is High
Requires Additional Unsupervised Algorithms (e.g. clustering) to Assign Labels	Ease of Labeling 'Fuzzified' Dataset based on Initial Skin and Features
Easy to Introduce Starter Holes	Starter Holes are Difficult to Introduce or Not Needed
Locations of Loads Applied can be Varied Freely	Locations of Loads cannot Interfere with Pocket Features

Machine Learning for Similarity Detection

• Using serial ML networks vs parallel ML  networks for feature recognition
• Type of data required for ML methods:
  • 2D images (Projection views; Sector views; Section views)
  • 3D mesh
• Evaluate use of unsupervised vs supervised ML methods (preliminary studies in progress) 

 Developing Machine Intelligence Techniques for Automotive Body Frame Engineering Design

• Project Period: 2020 - 2022
• Sponsor: Honda R&D Americas, Ohio
• Investigators: Jami Shah (Lead), Satchit Ramnath (PhD Candidate)

Research Questions

• Can experiential and generative design techniques be coupled to produce novel structures with good performance?
• How to generate manufacturable car body frames using solid meshes of monolithic structures computed by topology optimization?
• Is it feasible to fit hollow fabricated component sections between inner and outer styling surfaces to match inertia properties of solid sections resulting from topology optimization?
• How to formulate joint design based on geometry of components to be joined, loading conditions and weldability?
• How to generalize the methodology to any type of structure?

Workflow

Computational Techniques

• Generalization: determine design space which must account for voids that come from TO
  • Machine learning for edge detection of TO results for design domain creation
  • Automated 3D vectorization for creating curves/lines from 2D edges for design domain creation
• Generalization: separate solid and potential beam segments
• Cross-section generation
  • Radial basis functions (RBF) based points fitting for adapting properties of TO results into X-sec
  • Graph grammar based rules to evolve cross-sections
  • Genetic algorithm to optimize cross-sections with manufacturing constraints to meet target fitness function
• Joint Design
  • Use cross-section from adjacent components to create a blend with skeleton as spine
  • Create hybrid BIW model (beams for components and voxels for joints)
  • Run TO simulation on hybrid BIW model
  • Apply rules of manufacturing to convert solid result to sheet metal
  • Create parameterized sheet metal model using surface blend

Analytical Solutions for Production Variability in Complex Assemblies (Aircraft Engines)

• Project Period: 2017 - 2019
• Sponsor: Digital Manufacturing & Design Institute, Chicago (one of Federally sponsored NNMIs - National Network of Manufacturing Institutes)
• Partners: OSU (Lead), Rolls-Royce, Siemens PLM, ASU
• OSU staff/students: S. Ramnath, P. Haghighi, A. Schraff, K. Putta

Problem:

• Many engineering organizations typically rely on traditional dimensional analysis techniques whose calculations are mostly 1-D in nature. Small or medium manufacturers may not even have such expertise and may rely on trial and error or past experience.
• Traditional method may be sufficient to handle relatively simple assembled geometries, it may be very difficult to understand dimensional variations of complex assemblies with several degrees of freedom.
• Goal: Creation of mathematical and statistical methods, and computer tools for supporting variability mitigation via inter-operable digital tools which enable designers and manufacturing engineers to account for full 3D geometric variations in complex assemblies ensuring accuracy, efficiency, ease of use, digital interoperability and scalability.

Research Questions:

• There are a variety of tolerance analysis methods and software. How can we determine which gives the best prediction?
• How can statistical results be verified with a small sample size?

Major Tasks:

We developed a two-pronged approach:

• Pre-manufacturing (feed forward) strategy to use these new predictive 3D capabilities to gain better understanding of the variability in assemblies, which in turn, enable the effect of variation
• Post-manufacturing (feedback) strategy to determine optimal use of the tolerance budget to minimize accumulated effect on assemblies

Feed-Forward: Predictive Analysis

• Selected assemblies from Rolls-Royce engines were used in this project (also similar surrogates due to Export Control restrictions)
• Four different analysis methods were used to predict critical clearances based on "un-informed" GD&T; Predictions were compared to each other and to measured data
• Because the Rolls-Royce engine (hidden) was Export Controlled, we can only publically talk about our methods in the context of a surrogate assembly (shown)

Comparison of Variability Predictions:

The surrogate assembly was taken from ASME's professional development course on tolerance stacks:

• This assembly consists of 13 parts, 137 mating conditions, 124 GDT specs
• Two gaps were identified for feed forward predictive study
• CAD models of all parts in this assembly were created in NX
• Geometric features overlaying CAD geometry, GD&T, assembly operations and measurements were defined in Siemens TeamCenter.
• Tolerance analysis was done on that model within VisVSA and its summary data (PDO) was translated to aCTF format for predictive studies.
• Analysis methods were compared to each other

Comparison:

• There are difference between the actual GD&T schema and the VisVSA model, when comparing the 1D manual results with the other methods due to some restriction in modeling in Teamcenter, e.g. some profile tolerances are replaced with size tolerances in the automated 1D chart and other methods using the automated loops.
• 1D automated has a smaller ranges compared to 1D manual because our automated software does not account for bonus tolerance in press fit.
• The range of values obtained in Monte Carlo based analysis is dependent on the number of simulations; generally, more simulations should give a more accurate mean value and a bigger range.
• For Gap 2, the difference comes from a well-known problem in tolerances analysis of cylindrical mating features; and that is the difference between the nominal and the extreme possible location of the mating FOS. In the worst-case analysis, the analyzer has considered the maximum clearance by moving the pin in the furthers extreme location, while the automated system, has done analysis based on the extracted loop and the mating condition and location of features as translated from the geometry in the minimum scenario

Software Limitations:

VisVSA

• Some tolerance types will not be available for the feature, depending upon the feature type. For instance, flatness is not available for a pin feature (ends).
• Diameter modifier and material modifier (MMC, LMC) may be added to a tolerance depending on the feature type, tolerance type and the types of the datum features.
• A material modifier can only be specified to a feature of size for a straightness, positional, or orientation tolerance.
• For a specific feature, the user can apply at most, one tolerance type from each group, even if the DRFs are different.

ASU/OSU software

• Translator:
  • Runout tolerances are replaced with corresponding Size/Position tolerances (aCTF post-processor limitation)
  • Profile tolerances are translated to position (VisVSA limitation)
• Loop detection:
  • It does not consider DRFs, while the original thought was that it won’t affect in T-map we have realized it does have an effect.
• Analysis code:
  • The 1D automated chart cannot take Press fit into account
  • Since all automated methods are feature based, it cannot handle tapered pins correctly
  • Patterns are modeled and analyzed as one feature with pattern tolerances

Comparative Study of In-process and Post-process Approaches to Convert Topology Optimization Results to Manufacturable Sheet Metal Assemblies

• Project Period: 2016 - 2017
• Sponsor: Honda R&D Americas, Ohio
• Investigators: Jami Shah (Lead), Alok Sutradhar (Post-doc)

Background

Much research and commercial development has been done in advancing topology optimization (TO) of structures    ….     yet:

• It is isolated from mainstream product development tasks
• Requires considerable manual work to produce “faired” CAD geometry
• Non-parameterized geometry is not suitable for size optimization
• Results typically do not conform to manufacturable designs by industrial manufacturing processes, such as sheet stampings or castings

• We studies two potential approaches to to convert TO result to manufacturable sheet metal assemblies that we labeled “in-process” and “Post-process”

In-Process Steps:

• Formulate shape constraints in canonical form corresponding to complex geometries producible by manufacturing processes (e.g. stamping, joining)
• Embed selected classes of constraints into TO using projection functions

Post-Process Steps:

• Split/decompose into individual stamped pieces
• Match X-sec and profiles for each segment
• Output parametrized CAD to CATIA

Study Conclusion

• We concluded that the in-process strategy would be far more difficult to implement and less practical not easily generalizable
• Subsequent projects in DDML only focused on the post-process approach

Integrated Structural Optimization of Automotive Assemblies: Post-Process of Topology Optimization - Phase 1

• Project Period: 2017 - 2019
• Sponsor: Honda R&D Americas, Ohio
• P.I.: Jami Shah
• OSU staff/students: J. Kresslein, S. Ramnath

Objective:

Conversion of topology optimized designs to parameterized CAD parts manufacturable through mainstream mfg. (sheet stamping)

Assumptions:

• Starting Topology: Branched
• Starting Geometry: Solid cross-sections; high length to cross-sectional area ratio
• Given Design Envelope
• Pre-defined and expandable Cross sectional libraries
• Styling: No styling constraints (included in design space)
• Materials: Single Material

Phase 1 encompassed Determination of workflow, Skeletonization, Segmentation and Cross-section extraction from TO results

Workflow:

Model clean-up and reduction: Used Meshlab and Starlab (open source)

 Get clean, connected, watertight surface mesh from voxels

Skeletonization of TO mesh (to extract load paths)

 Segmentation: Splitting the skeleton into separate components that may be manufactured as one piece

Cross-section extraction from segments (using planar cuts)

Computer inertia properties of extracted cross-sections (with irregular boundaries):

• We considered convex hull, Delauney Triangulation, Alpha Shape, and 'Concave Hull' (shrink wrap) algorithms
• Concave hull algorithm gave the best representation of the section boundary (ref: Moreira, Adriano, and Maribel Yasmina Santos. "Concave hull: A k-nearest neighbours approach for the computation of the region occupied by a set of points." (2007): 61-68)

​Integrated Structural Optimization of Automotive Assemblies: Post-Process of Topology Optimization - Phase 2

• Project Period: 2019 - 2020
• Sponsor: Honda R&D Americas, Ohio
• P.I.: Jami Shah
• OSU staff/students: LS Wang, S. Ramnath, P. Haghighi

In a continuation of the project we developed: Skeleton Editor, cross-sec libraries, cross-sec initialization into Design Space, and Parametric CAD (CATIA) geometry generation

Skeleton Editor

Cross-section Libraries

Cross-section Initialization into Design Space

Method for Orienting and Scaling Cross-sections to Fit inside Design Space

 Parametric CAD (CATIA) Geometry Generation

Geometric Algorithms Developed/Adapted

Hybrid Techniques Research for Joining Dissimilar Materials

• Project Period: 2016 - 2019 (*extended to 2020 with internal funds after end of sponsored period)
• Sponsor: Department of Commerce (partial funding #06-49-06019)
• P.I.: Jami Shah
• Technical Consultants: Dr. Ali Nassiri (OSU), Ed Kortis (Arnold Fastening), Pete Edwards (Honda EGA),  Dr. Ramirez and Dr. Benatar (OSU Welding Eng)
• OSU staff/students: J-H Kim

In order to reduce vehicle weight, automotive manufacturers are increasingly using non-ferrous materials. This poses new challenges for joining dissimilar materials, since spot welding (RSW) does not work.

Hybrid Joints Knowledge Based System (KBS)

• Objective: Deliver the knowledge of joining methods to structural designers in early design phases

 Functions

• Data mining: Get data on feasibility, cost, structural integrity of joining methods
• Collect any available data of the joining methods
• Structure & populate a knowledge repository of joining to store the data in a formal way
• Devise Search algorithm
• Apply optimal algorithm to search for the appropriate joining methods satisfying the criteria
• Develop Software UI; Provide user friendly interface to display the suitable joining methods under the constraints defined by the designers

Structure of KBS

Process Parameters: Any categories related to the joining process Unique attributes: Characteristics from joining process that are distinguished from other joining methods Common attributes: Processes in joining methods that can be found in common   Performance parameters Strength-wise attributes Design constraints that the designers may have

This knowledge base contains thousands of entities and relations characterizing key aspects: materials that can be joined, mechanical performance, process parameters, cost, weight and robustness for thermal, mechanical, adhesive and hybrid joining methods relevant for multi-material joining. This information has been collected from diverse open sources, supplier manuals, industry showcases, academic labs and research literature.

KBS works in two phases with increasing level of details

Initial Search:

Detailed Search:

Surrogate Models

Objective: Typical car bodies contain more than 5000 joints of various types. In order to simulate body structures for crash loads, it is necessary to simplify and idealize the joints (Surrogate FE models). This must be done to account or the joint stiffness contribution at the global level. For critical joints, a detailed FE analysis can be done at the local level with sub-modeling techniques, taking the actual joint geometry and material conditions into account.

Surrogate modeling methods are used to perform quick  simulations to check the performance of the joint.

Scope: In this project, we undertook the following tasks:

• Survey of FEA idealization methods and past work on common joint types
• For selected joining methods (FDS, adhesive bonding), FE parameters are ‘tuned’ to match the experimental data and detailed FE models
• Demonstrate use of tuned models in joint design

 Characteristics of FDS

Detailed FE Model of a Fastener 0.08mm element size gives over 1,000,000 elements

Potential Surrogates Geometry simplified into beams, solid elements, and MPC Simulated the 'global behavior' of the joint Simulation result is calibrated to  curve-fit the test results Reduces complexity drastically: Solving time reduces from days to hours

Calibration of Surrogate Models

• Minimize the difference between the simulation result and the test result
• LS-Opt parameter calibration was used to curve-fit the simulation result

Lap Joint: FDS Surrogate with MPC

Shear Load	Cross Tension Load

Adhesive Joint with Cohesive Elements:

Shear Load	Cross Tension Load

Uses Cases:

Surrogate Modeling of Hybrid Joints

Why use hybrid joints? Hybrid joints (adhesive + fastener) provide additional energy absorption and longer failure displacement The initial load is taken by the adhesive; when it fails the fastener holds the joint Surrogate results correlate well with the characteristics of the hybrid joints

Design of Joints (number of fasteners, locations, adhesive layup)

Automated Big Data Generation for Deep Learning for Automotive Hood

• Project Period: 2018 - 2019
• Sponsor: Honda R&D Americas, Ohio
• P.I.: Jami Shah
• OSU staff/students: S. Ramnath, P. Haghighi, Y. Jiang 

Objective:

Investigate the use of deep machine learning to extract feature attributes and patterns correlated to performance parameters; use such knowledge to develop future structural design and optimization tools.

Tasks:

• Study hood design process and data to devise automated workflow
• Automate and Generate hood geometry data set (~10,000 samples)
• Conduct FEA studies of hood lift and twist load cases for design variants

CAD-FEA Automation Workflow

Automated Data Generated

•  Started with 6 skin surfaces (exterior hood surface) and 6 feature patterns (support frame)
• Expanded to 10 skin surfaces and 11 feature patterns for hood frames
• Devised a generalized parametric method to map any feature pattern to any skin

•  About 30% parameter combinations were failing to produce CATIA models
• To overcome this problem, a constraint network and filtering scheme was devised for: Local validity: parametrization & constraint schema; Global validity: prevent interactions between features

• Generated and validated 10,478 models
• Model formats stored: All models have been saved as *.CATPart, *.stp, *.stl formats

Big Data Efficacy met for Machine Learning?

• Quantity: >10,000
• Variety: 110 topology variants x 100 size variants (can be further enhanced with multiple materials and thicknesses)
• Validity: idealized geometries validated to correspond to real hood performance (see next Task)
• Efficiency: re-formulated macros; one macro runs in 40-50mins to produce 100 models

Standardized Boundary Conditions (BCs) for Hood Lift Analysis

Refined idealization and standardized BCs to get better match between idealized and actual geometries:

​

Comparative Studies of Structural Optimization Methods for Automotive Hood Frames

• Project Period: 2019 - 2021
• Sponsor: Honda R&D Americas, Ohio
• P.I.: Jami Shah
• OSU staff/students: S. Ramnath, J. Ma, A Li.

Objective:

To enhance experiential design, we performed comparative studies of structural optimization methods for automotive hood frames with respect to component efficiency and practicality.

Scope: The methods included were:

• Traditional Design of Experiments (DoE) based sampling
• Non-traditional DoE for handling non-uniform parametric sets from alternative designs
• Topology Optimization
• Machine Learning (cursory demonstration)

Traditional DoE

• As in any DOE, it is good idea to reduce the number of parameters, both on input and response factors. The former was done with sensitivities and main effects study in MiniTab. The latter was doe with correlation studies between response variable
• Each hood frame and skin combination has different set of parameters, so the use of traditional DOE is limited

DoE Response Correlation:

• 20,000 simulations were run (2 load cases * 100 design points 100 idealized models) to determine if we can reduce the response of all the simulations to just one variable – either lift or twist load
• The correlation coefficient was found to be 0.836: hood lift and twist are positively correlated

Parameter Reduction Example

• Traditional DoE can only be done with parametric variants from the same skin and pocket combination. Feature parameters were varied with optimal space filling within practical ranges. The deflection result was sorted to find the optimal parameters

Non-traditional DoE

Three different approaches were used to deal with non-uniform parameter sets:

Method 1: Separate DoE Skin by Skin

• For each skin, data points for all pocket patterns were put in one table, and they were sorted based on the value of the output parameter
• Because skins are of different sizes, comparison across skins is only possible if the output parameter (deflection) is normalized; therefore, we used the product of the structure mass and the directional deformation

Method 2: Amalgamated Dataset (from all skins)

• Uses the same input, output parameters as method  1, but combines data points for all skins
• This method allows us to see if there is a dominant pattern, regardless of skin shape and size
• Done by looking at the frequency of each pattern appearing in the top n samples
• Method 1 practicality is limited because it does not align with design processes. Designers would be constrained by the skin geometry but it could serve as a general guideline
• Max-Min normalization: Since each skin has a different property, the output parameter (deflection*mass) must be further normalized before the analysis
• We can sort by response and look at frequencies in top n%.

Method 3: Homogenization of Input Parameters for Uniform Comparison

The question we asked: can we describe all skins and all pocket patterns with some key parameters in a uniform way?

• Skin attributes: Aspect Ratio, Axial curvature, Transverse curvature,
• Frame attributes: Net Area, pocket pattern types and the max transverse depth
• 11 pocket patterns were categorized into 5 types

•  In this manner, we can use Response Surfaces to see the effect of each attribute. Furthermore, the RS can be explored to find optimal points

•  Using GA, 100 candidate points were obtained for 11000 data points
• Since there is not a data point which has the exact same value of input parameters as candidate points, the closest data point for each candidate point was found
• The method of doing that is to get the absolute value of the subtraction between the data point and the candidate point for every candidate point

Topology Optimization

• The purpose of topology optimization study is to derive the non-parametrically optimal structure of pocket feature pattern that can minimize deflection and improve the hood performance

Procedure:

• Establish a uniform way to conduct Topology Optimzation
• Generate solid region from skin extrusion
• Improve realism with pre-defined holes and local loads
• Randomize hole locations, load locations, mass ratio, etc.

Machine Learning (Exploratory Study)

• Results from mix-match of six skin/features obtained from DoE studies
• Used clustering algorithm (k-means) to create 4 clusters based on deformations
• Cluster bounds (automatically) generated are:
  • 1mm < cluster 1 < 6.02mm (count: 1917)
  • 6.06mm < cluster 2 < 10.93mm (count: 2368)
  • 10.95mm < cluster 3 < 17.71mm (count: 662)
  • 17.8mm < cluster 4 < 24.89mm (count: 215)
• Based on the above table, the ‘best’ pocket/feature pattern for surface I is pocket E
• The method above can also be applied to stress and/or mass values

Archived Projects (ASU DAL): 2015 and earlier

The DESIGN AUTOMATION LAB was established by Prof. Jami Shah at Arizona State University and operated from 1991 to 2015. The lab focused on software technology development to support design and manufacturing. Several hundred MS, PhD, undergrad interns and post-docs passed contributed to cutting edge research in CAD, FEA, CAM, CAPP, GD&T and Cognitive Studies of Designers. DAL Graduates have achieved great success in academia and industry.

Sponsors

• Government: NSF, DARPA, NIST, Army Research Ofc, DMDII
• Industry: GE, TI, ICI, Boeing, Ford, AlliedSignal/Honeywell, HP, Siemens PLM
• Consortiums: CAM-I, USCAR, AIAG (Automotive consortium)

Research Areas

• Parametric & Features Technologies
• Mathematical modeling of Tolerances & Dimensional Metrology
• Cognitive studies of Design Ideation
• Design & Manufacturing Automation 

DAL Research Evolution

 Project Sponsors

Parametric & Features Technologies

Reference: Shan & Mantyla, "Parametric & Feature based CAD/CAM", John Wiley, 1995. 

Assembly Feature Characterization & Recognition

• Motivation: reverse engineering of design intent for replacement parts in legacy systems for the US Army
• Extension of NRep to assembly relations
• Definition of assembly feature templates and data models
• Automatic recognition algorithms for mating features, kinematic and structural function - towards machine informatics

Mathematical Modeling of Geometric Tolerances

• There is no mathematical model capable of capturing all GD&T ASME Y14.5 conventions; hence none consistent with engineering practice
• Many methods have been proposed for tolerance representation & analysis, based on elegant math models but failed to gain acceptance because they were not compatible with the standards
• Research Goal: Develop math models for GD&T consistent with the standard, i.e. retroactively fit a math model to the conventions in ASME Y14.5M

Bi-level Tolerance Math Model

Metric Model (T-Maps*)	Topological Model (CTF Graph)
Models individual tolerance zones for each class by mapping to hypothetical higher dimension space models the composite quantitative effect of all tolerances on a given feature	Similar to topology or control schema Models relationships between all feature control frames Models datum reference frames (DRF) and their precedence (datum flow chain)

*US Patent No. 6,963,824

More than 50 T-maps have been developed based on combinations of target feature, tolerance type and datum type; here is an excerpt from the T-map catalog

Worst Case Tolerance Analysis with T-Maps 

Tolerance Stacks: Fitting Accumulation Map inside Functional Map

ASU GDT Testbed (Includes Multiple Analysis Methods) 

• Automated Stacks analysis for any chosen method:
  • 1-D min/max charts
  • 3-D statistical & worst case with Monte Carlo simulation
  • 3-D statistical & worst case with ASU math model (T-maps)

Project funded by NSF through two different grants (1999 - 2006)

 Automatic Tolerancing of Mechanical Assemblies from STEP AP203

Objective: Automatically generate GD&T schemas and allocate tolerances to mechanical assembles in which only the nominal configured geometry is available in STEP AP203 files.

History: This project was proposed as a multi-phase plan by ASU under the DARPA AVM program. Only the first phase was completed when AVM shut down. The project got a new life under DMDII sponsorship.

 Auto-Tolerancing: Definitions

• First order tolerancing:  GD&T based only on geometric conditions for assemblability
• Second order: GD&T based on both assemblability and design intent/function
• Third order: GD&T based on all of the above, while optimizing manufacturing cost

In this project we implemented and delivered First Order Auto-tolerancing modules to DMDII membership.

Pre-processing Modules: 

 Auto-Tolerancing Modules:

Normative Algorithms for CMM Metrology

• If mechanical parts are produced based on drawings (CAD) in accordance with GDT standards and measured by different methods, will we get consistent results?
• Will shop floor manual methods be consistent with CMMs?
• Will different CMM tolerance verification software working with the same measured data points, agree with each other?
• How should CMM data be processed to be consistent with Y14.5?

Cognitive Studies of Design Ideation

"Innovation will be our ticket to a vibrant healthy economy with a competitive edge, and the creation of large numbers of high quality jobs" - NAE Report

Simulation Real World Design

Design Skills: Characterization & Standardized Testing

Objectives

• Enumerate design skills associated with design tasks
• Characterize skills in terms of measurable indicators
• Devise standardized tests for conceptual design skills
• Collect test data; Evaluate, Norm and Validate tests

Applications/Motivation

• Standardized methods to evaluate engineering students,  design courses, teaching methods,..
• Tools to form balanced design teams based on collective skill profiles where each designer is characterized in terms of design skills, sub-skills

Measures of design ideation effectiveness

Reference: Shah, J. J., Smith, S. M., Vargas-Hernandez, N., 2003, “Metrics for Measuring Ideation Effectiveness”,  Design Studies, V24 (2), pp. 111-134.

Outcome "End Goal" measures:

•  Novelty: how unusual or unexpected an idea is as compared to other ideas
•  Quality: feasibility and conformance to design specifications  

"Process" measures:

• How effective is the ideation method in expanding design space?
•  How effective is the ideation method in exploring design space?
  • ​​​​​​​Variety: how different concepts are from each other
  • Quantity: total number of ideas generated

Refinements to ideation measures:

• Best Quality and  Best Novelty
• “Efficiency” (above metrics divided by time)

Standardized Tests of Engineering Design Skills

Collaborators: R. Milsap (ASU), S. Smith (Texas A&M)

Research Questions

• What cognitive skills are involved in engineering design?
• What are the indicators of these skills?
• Is it possible to measure skill levels in individuals?
• Is it possible to create standardized tests for each skill?

Applications/Motivation

• Standardized methods to evaluate engineering students, design courses, teaching methods,..
• Tools to form balanced design teams based on collective skill profiles where each designer is characterized in terms of design skills, sub-skills
• Research tool for “calibrating” individuals and teams used in empirical studies

The scope of our study is limited to skills related to early design stages:

• Problem formulation (PD)
• Visual Thinking (VT)
• Divergent Thinking (DT)
• Qualitative Reasoning (QR)

Four tests were developed and validated:

Divergent Thinking Skill Characterization (Definitions and Measures)

Visual Thinking Skills and Definitions 

Design Problem Formulator (NSF CreativeIT Initiative: Major Grants)

• Problem Formulator is an interactive web-based tool that helps designers build a model of their understanding of a design problem
• How the Formulator helps designers
  • Documentation: remembering one’s problem definition in later stages
  • Communication: sharing one’s understanding of a problem with others
  • Organization: structure can help in expressing the many aspects that need to be considered

Computer Implementation

The FORMULATOR can also be used as a research tool; P-maps are stored in structured ways and can be compared to normative formulations. Steps can be recorded and replayed.

Creating P-Maps GUI

Detailed P-Maps 

Design Ideator

Multi-modal toolset for concept generation

• Intuitive strategies
• Experiential methods/knowledge bases
• Holistic framework

The long term goal of this research is two-fold:

• Develop a framework in which multiple ideation methods can be inter-laced for facilitating conceptual design 
• Create a research testbed which would facilitate the collection of large amounts of data to build models and theories of design cognition and effective tools for conceptual design, particularly for novices
• Passive CACD: supports documentation of models, maintains relations between models and facilitates navigation between tasks and applications;
• Creating domain independent languages and representations for function, behavior, embodiment would suffice for implementing CACD
• Active CACD: pro-actively guides the designer, “senses” creativity blocks, suggests strategies, provides access to relevant knowledge and data in repositories and KBs, supports ideation
• Both have human designer driving the process; not automated systems
• 50 years of research in cognitive science, design thinking and human problem solving have studied causes of impasse
• Many ideation methods, tools, strategies have been devised
  • …. But to what extent can these be incorporated into computer tools?

Software Tools for Design & Manufacturing Support

Developing Theoretical Foundations for DfM: In collaboration with Prof. Paul Wright, U.C. at Berkeley.

DfM System Architecture

• DFM Advisor, previously developed in the DAL, offers a robust and customizable framework  for manufacturability analysis.

Semantic Integration of CAD and CAE with UoFs (Units of Functionality)

DARPA project in collaboration with Boeing Defense and MSC.

PI: J Shah, Co-PI: Susan Urban

Legacy Systems Engineering

Re-constructing Tech Specs for Army devices and systems to facilitate manufacturing spare parts in Army Mobile Hospitals.

 DARPA TTO AVM (Adaptive Vehicle Make) Program

Crowd sourced design competitions based on resources provided by AVM contractors (part catalogs, simulation tools, test benches..)

ASU projects in AVM:

• META2 with Boeing (Complexity and Adaptability Metrics)
• ifab with PARC (Digital design and machining)
• ifoundry with PSU, RICARDO, Recon
  • ​​​​​​​Assemblability analysis
  • Predicting variability in alternative assembly sequences
  • Assembly feature recognition
  • Auto-tolerancing of submitted design

Complexity Quantification

• Aerospace and military cyber physical systems are intrinsically complex; what we need to do is determine the minimum complexity to meet given requirements
• Implication: explicitly consider the trade-off between performance and complexity
• Research Challenge: create metrics that are mathematically consistent, domain independent, applicable to all stages of product development and can be validated/calibrated with real data

• The development, production and operation of complex systems involves multiple phases, domains and elements
  • ​​​​​​​Artifact: Functional structure (A_f); Physical structure (A_p)
  • Development process: intrinsic (D_i) , extrinsic (D_e)
  • Manufacturing: intrinsic (M_i), extrinsic (M_e)
  • Service: operation (S_o) , maintenance (S_m)
  • The overall complexity X is the weighted sum of all contributions

• Consider each domain to be spanned by three basis vectors corresponding to Size, Coupling and Solvability
• Each design option can be represented by coordinate values a_ij in the basis directions

Conducted several case studies and compared to actual development time and peak labor hours at Boeing.

 Fractionated Satellite Design

• Design space exploration at conceptual level
• Multi-objective optimization: cost, time, adaptability
• Correlating value based analysis with product architecture characteristics

Automating Digital Fixture-setups for Machining Castings to Minimize Scrap

• Background: Sand castings have poor dimensional control; consequently, some surfaces and features need to be machined (e.g. bearing surfaces, sealing surfaces, mating surfaces)
• Problem: How to reduce setup time for machining yet ensure a good finished part from each casting?
• Research objective: Given point cloud scan of a sand casting and a CAD model of the finished part (including desired tolerances), machining fixture setup, can we devise a method to determine fixture adjustments automatically to ensure that all critical features will lie in their tolerance zones before machining operations begin

• The setup point represents a small displacement of the body with respect to the intersection coordinate system (GCS/FxCS)
• The small displacement is transformed into adjustments of the 6 adjustors pictured. The part will then be forced to make contact with all adjustors and clamped in place.