SigOpt Summit Recap with Novelis: Streamlining Materials Design with Intelligent Experimentation with Novelis

Steven Stein

Overview

In this blog post we are reviewing the SigOpt Summit presentation by Novelis on how they are Streamlining Materials Design with Intelligent Experimentation with Novelis. During this talk, Dr. Vishwanath Hegadekatte of Novelis showed how they created an experiment environment to improve the outcome of R&D projects targets. Their goals were to run Intelligent Experiments to achieve better results with less simulations. Specifically, they wanted to experiment how they could design better aluminum materials for specific customers.

This blog will detail how Novelis used SigOpt in to manage four different experiments to find how to improve the material properties of aluminum for their customers. A brief overview of the use cases are below with associated results:

Use CaseDesign Improvement Using SigOptSimulations Reduced
#1 - Beverage Can Bottom DesignMax Buckling Pressure:
7.5 bar → 8.6 bar (15%)
Max Column Load:
1.5kn → 3.0kn (100%)
15,000 → 99
(99% Reduction)
#2 - Automotive Axial Crush AnalysisLoad Factor:
0.75 → 0.9 (20%)
250 → 26
(90% Reduction)
#3 - Can Lid Design OptimizationsMax Buckling Pressure:
112 MPa → 113 MPa
Not Measured
#4 - Aging Model OptimizationsLog Error:
Lowest error of all optimizers
~10,000 → ~250
(98% Reduction)

Results of all use cases

Novelis

Novelis is the world’s leading producer of flat-rolled aluminum products and the world’s largest recycler of aluminum. With businesses across the globe and 14,650 employees, Novelis generates $12.3B in revenue shipping 3,274 kt of products in 9 countries. Novelis has 4 main verticals: beverage cans, automotive, aerospace, and specialties. Their customers include some of the largest and best-known aerospace, automotive, beverage can, architecture, and consumer electronics brands in the world.

 

Experiment Design for Materials

Novelis wanted to create a design of experiments (DOE) environment to experiment how to create the right type of aluminum for their customer. The goal was to create fast and efficient experiments in an R&D setting. Each customer of Novelis requires a slightly different characteristic in the aluminum, may it be the strength, the thermal threshold, or manufacturing throughput. And thus Novelis needed a quick and effective method to run experiments yielding better designed materials. The problem is each iteration of design is costly and time consuming given the high powered computers systems that were required for each experiment. Therefore, SigOpt was used to reduce the total number of simulations and yield better results with less runs.

Why did Novelis Choose SigOpt?

  • Ease of use: The Enterprise level GUI is very important to make the experiment design simple. This enabled other departments in Novelis to easily replicate experiments and use SigOpt for their own needs.
  • Optimal designs: Novelis produced better designs which yield stronger structure for their customers’ use cases.
  • Reduced simulations: SigOpt achieved results faster than other optimizers. Meaning SigOpt ran less simulations and thus consumed less IT resources to find the optimal answers.

Case Study #1 – Beverage Can Bottom Design

This use case involves designing a can to withstand higher pressure and reducing the chance of the can buckling at the bottom. Novelis’ challenge is to modify the design parameters of the can to achieve the highest performance as measured by the pressure the aluminum can withstand. A beverage can is one of the most engineered objects you will ever hold in your hand, rivaling that of the Space Shuttle in complexity. A beverage can has a large amount of parameters that can affect its design. However, if you were to analyze every combination of these parameters to find an optimal design, you would need to run 2e12 experiments to evaluate the whole parameter space.. Evaluating all of these parameters using a Finite Element analyzer (FE) would take approximately 147 years. This is not feasible for just one design. So, Novelis required a more streamlined method to evaluate the parameter space and find the optimal solution. And thus Novelis used SigOpt to solve this problem.

Beverage cans are known as the most engineered products you will ever touch

Using a commercially available method, Novelis ran 15,000 simulations to find an optimal result. This result found a design that could withstand 7.5 bars of pressure and a load of 1.5kN (the column strength of the beverage can). Then using SigOpt, Novelis found a better design while only running 99 simulations. This design was able to withstand 8.6 bars of pressure and a maximum load of 3.0kN. This is a 15% and 100% improvement, respectively. Running less simulations also means less demand on IT resources for this task; allowing for these IT resources to instead be used on other projects.

Parameters used to design the bottom of a beverage can

Commercial optimized vs SigOpt optimized can design

Novelis was able to find a more optimal solution while consuming less IT resources by using SigOpt as compared to a commercially available option.

Commercial optimized vs SigOpt optimized can design

Case Study #2 – Automotive Axial Crash Simulation

Crash simulations are one of the most important applications where aluminum is used. Design safer cars are possible by modifying how the aluminum is created. Novelis’ task is to design the aluminum component to fold in a known, or “good”, way rather than in a “bad” way. Folding in a “bad” way can cause cracking and additional damage to vehicles than if the aluminum were to fold in the “good” way.

 LS-OPTSigOpt
Best Value~0.75~0.9
# simulations25026

Axial Crash Optimization Results

Case Study #3 – Beverage Can Lid Design

The objective here was to adjust the geometry of the can lid to find the highest buckle performance; the pressure at which the can lid would pop out. Novelis used SigOpt to run various experiments using a finite element analysis simulation and found a better design to improve the buckling pressure from 112 Mpa to 113 MPa. This one MPa difference is a significant improvement given how highly engineered cans are. In this case, four parameters needed to be tuned to find the best lid design.

Can lid

SigOpt Visualization of simulation results tracking the best value metric

Visualization of SigOpt modified lid design with new buckling pressure results

Case Study #4 – Calphad Aging Model

In this example, Novelis is optimizing the strength of aluminum created during manufacturing operations. Specifically, in this use case Novelis is optimizing the aging process to yield the highest strength of the material. This computation is used by a code called ThermoCalc where Novelis provides chemical composition to yield a time and temperature profile. The goal is to optimize these parameters to simulate the yield strength as a function of temperature and time. This model has several parameters to balance to find the best results.

Time and Temperature Graph

Comparisons of optimizers with SigOpt

The GA optimizer was able to find a similar result to SigOpt but it took that optimizer 10,000 ThermoCalc simulations whereas SigOpt only required ~250 simulations. This is a substantial time reduction since 10,000 ThermoCalc simulations require about a day to complete.

Novelis also compared SigOpt to the other platforms based on usability, number of maximum parameters, and other modeling variables as shown in the below chart. SigOpt delivers the best out of these other optimizers with the #1 reason being the ease of use. The Enterprise level GUI is very important to make the experiment design much easier.

PropertySigOptAxDragonfly
Parameters100No limits in functionality. 20 advisedGitHug Repository not updated in more than a year
Metrics25
P-constraintsNoneNo limits in functionality
M-Constraints4No limits in functionality
CostSubscriptionOpen Source
UsabilityEnterprise Level GUIProgramming expertise required. No GUI

Table comparing different properties of optimizers with SigOpt

Batch vs Sequential runs

Novelis also performed a study on the impact of batched vs sequential runs. When running simulations that require massive compute systems organizations often have to balance who has access to these systems by scheduling blocks of time each team can use the supercomputer. The Novelis R&D team is only allotted certain time slots to conduct their experiments. Thus they need to run their simulations in batches due to the scheduling processes of the supercomputer.

Batched vs Sequential runs results. Lower is better

The sequential observations came to a better result faster than batching the observations. So being able to run the simulations sequentially would be more valuable if the scheduling allows it.

Takeaways

Use CaseDesign Improvement Using SigOptSimulations Reduced
#1 - Beverage Can Bottom DesignMax Buckling Pressure:
7.5 bar → 8.6 bar (15%)
Max Column Load:
1.5kn → 3.0kn (100%)
15,000 → 99
(99% Reduction)
#2 - Automotive Axial Crush AnalysisLoad Factor:
0.75 → 0.9 (20%)
250 → 26
(90% Reduction)
#3 - Can Lid Design OptimizationsMax Buckling Pressure:
112 MPa → 113 MPa
Not Measured
#4 - Aging Model OptimizationsLog Error:
Lowest error of all optimizers
~10,000 → ~250
(98% Reduction)

Results of all use cases

Key Learnings

  • Modeling Case studies are useful as benchmarks and one-to-one comparisons
  • Optimization is exactly the same for experimental work (e.g.lab and plant trials) too
  • Bayesian optimization will likely give better results than traditional DEO for limited observation budgets

Take Action

To try SigOpt today, you can access it for free at sigopt.com/signup

PXL_20211015_211818478.PORTRAIT
Steven Stein Product Marketing Lead

Want more content from SigOpt? Sign up now.