This document presents an artificial intelligence approach for exploring design parameters in a multi-objective setting. The approach uses both statistical and deep learning models to optimize two disc coupling designs based on mass minimization and stress minimization. Deep learning models are trained to predict mass, buckling load, and stress values, then used along with an optimization algorithm to iteratively search for design parameters that minimize the objectives. Results show the deep learning models achieve similar performance to statistical models in navigating the design space and finding optimal designs that meet the objectives.

1. Application of Artificial Intelligence Based
Approach For Design Parameter Exploration in a
Multi-Objective Setting
Presented by
Gehendra Sharma
2nd Annual Symposium: Artificial Intelligence and Machine Learning at OU
Optimizing Models Validation and Comparison
Optimization Criteria Objectives Design A Design B
Mass Minimization
Mass (g) 0.05277 0.079
Stress (MPa) 275.02 195.645
Stress Minimization
Mass (g) 0.2714 0.2751
Stress (MPa) 110.75 95.66
StatisticalModelsDeepLearningModels
Statistical Method
Performed 27 simulation experiments for each design.
Deep Learning
Performed 125 simulation experiments for each design.
Disc Design A Disc Design B
Mass = 0.001997 lbt – 0.003713 bt
Stress = 263.3 + 1065.3 t2 – 0.47 lb - 25.1 lt2
Buckling load = - 0.9950 l2bt3 + 2075.194 bt3
Mass = 0.001529 lbt + 0.016126 lt – 0.2621 t
Stress = 292.91 + 769.27 t2 – 5.167 l – 17.52 lt2
Buckling load = - 1.4287 l2bt3 + 2957.43 bt3
Normal Boundary Intersection (NBI) Method
Minimize β
x ϵ S
Subject to α + βn = f(x)
DesignBDesignAOptimization
Model1(X) Predict Mass at X
Model2(X)
Predicts Buckling load at XModel3(X)
Predicts Stress at X
Trained to
Trained to
Trained to
Minimize Mass
Objective = Model1(X) + 5000*min(0, Model3(X)-150)2
Minimize Stress
Objective = Model2(X) + 5000*min(0, Model3(X)-150)2
Objective
Update X
Initial X is picked from the simulation data which best
satisfies the objective.
Store
Give small
changes to X’s
Evaluate
Objective
Update X in a
direction of
improving
objective
Is objective better than
the best obtained?
Store the best
value of objective
Design Exploration Approach
Statistical Method Deep Learning
Simulation
Experiments
Simulation
Data
ResponseSurface
Model
Optimization
Simulation
Experiments
Simulation
Data
DeepLearning
ModelOptimization
Frame of Reference
Engineering design involves series of steps followed in
designing a product that meets certain criteria and
accomplishes a predefined objective.
Disc Coupling
Disc Coupling transmits torque between two rotating
equipment's by means of discs.
Design Process [adopted from NGSS]
How can the ML/AI based techniques support human designers
in exploring design parameters in a multi-objective setting?
Buckling
MisalignmentDisc
Two source of loads in discs, i.e., torque and
misalignment.
Torque = Buckling + Fatigue due to fluctuating stress
Misalignment = Fatigue due to fluctuating stress
Problem Description
2 discs designs selected based on topology optimization.
Disc Design A
Disc Design BOptimized Topology
Bounds
Disc Design Parameters
Response of
interest
Length
mm
Width
mm
Thickness
mm
Low 24 3 0.3 Mass
Average 32 6 0.6 Stress
High 40 9 0.9 Buckling Load
Minimize: Mass and Stress
Subject to: Buckling load > 150 N
Optimization
Model
Simulation
Rangesand
Interest
ProblemDesignMethodResults&Discussion
Outcomes and Closing Remarks
Contributions
➢ The applicability of statistical/deep learning methods
in predicting simulation results for designing the
engineering systems.
➢ Optimizing the design requirements by utilizing the
trained prediction models.
➢ Deep learning models utilized for searching design
parameters for optimal design requirements.
➢ Deep learning is as effective as statistical tools in
navigating design space for searching optimal designs.
Way forward
➢ Identification of the key parameters in deep learning
models that are most sensitive with respect to making
an effective prediction and optimization of engineering
design problems.
➢ Development of a robust deep neural network that
makes effective prediction and optimization for wide
range of design problems.
X + Δ X
Better
?
0.0464 0.0452 0.05277 0.056
0.2268 0.224
0.2714
0.2567
0.071 0.0754 0.079 0.0754
0.27 0.2737 0.2751 0.2665
0
0.05
0.1
0.15
0.2
0.25
0.3
Regression Actual Deep Learning Actual
Mass(g)
Validating Regression and Deep Learning Results
Design A Mass Minimization Design A Stress Minimization
Design B Mass Minimization Design B Stress Minimization
277.3
303.5
275.02 265.4
100.36
121 110.75 123.8
200.29 195.7 195.645 195.7
92.39 97.69 95.66 95.54
0
50
100
150
200
250
300
350
Regression Actual Deep Learning Actual
Stress(MPa)
Validating Regression and Deep Learning Results for Stress
Design A Mass Minimization Design A Stress Minimization
Design B Mass Minimization Design B Stress Minimization
0.0464, 277.3
0.27, 92.39
0.212, 103.82
0.071, 200.2891
0
50
100
150
200
250
300
0 0.05 0.1 0.15 0.2 0.25 0.3
STRESS(MPa)
MASS (g)
PARETO SOLUTIONS
Design A Design B