Data Science for Engineering Design Optimization
Data Science

Data Science for Engineering Design Optimization

Computer aided engineering (CAE) has been a boon to automotive industry for decades now. Over the period, several computer applications have been developed for every stage in Automotive design. Numerous tools are available for CAD/modeling, pre-processing for solvers to conduct simulations using Finite element analysis or computational fluid dynamics and then post-processors to visualize and understand the design outputs. These tools require special skills and it takes several hours for each simulation cycle. The automotive design involves thousands of simulations across various disciplines in design and each automobile from concept to final design takes about 18 months or more.

In most engineering simulations, the objective is to arrive at optimal parameter values to achieve the desired results – meaning, for example, arriving at minimum thickness (weight), material or shape of each component in the system under consideration. Unlike most research papers that focus on limited dimensions (design variables), the reality is that the variables involved are in large numbers. Each simulation takes several hours and is expensive.

While design process is streamlined over the years, the market conditions have changed enormously. Faster go-to-market on one hand and need for innovative design on the other are placing stress on the design cycles. Product development is becoming increasingly expensive because of regulatory compliances, testing and validation. These factors call for the review of current practices and innovations in design.

The objectives of improvements in design process to meet the new demands should reduce cycle time and provide platform for automation. The current design systems for pre-processing, solvers and post-processing are all work in silos whereas most disciplines are interconnected. The data preparation and exchange of the intermediate results to next in line system is cumbersome. There are tools to automate some part of this process but still cycle times are longer than desired.

Surrogate Models and Design Optimization

Response Surface Models (RSM) and optimization tools have been gaining importance. The elastic processing power of the cloud and advancements in data sciences have provided the opportunity to augment the design process through surrogate models for responses and then optimizing the designs using these models. Wherever large number of simulations are involved, using the initial few sample design combinations and their responses, the surrogate models can be built, and these models can then take the role of solvers reducing significant number of simulations.

The overall process looks like the following:

Optimized designs are of good quality when the models from machine learning tools are built with least error.  Larger the size of DoE, lesser is the model error but each simulation in solver is expensive and time consuming. Hence the challenge is to get good design space coverage with least number of samples to produce models with highest accuracy.

Design of Experiment – DoE is a selection of multiple design combinations wherein the design variable values take different values within the allowed values. These sample designs are then passed through traditional solvers along with their geometric parameters to get the response values for each design combination. Each simulation in solvers is computationally intensive and expensive. Advanced algorithms are used to achieve minimum samples requirement.

Modeling – Several advanced algorithms and machine learning tools are now available to build fairly accurate surrogate models for linear or non-linear behavior of responses with respect to design variables. Dimensionally reduced models some time provide best results by reducing the noise. Metamodel concepts eliminates the designer bias and model parameter tuning requires adept skills in data science.

Optimization – It is a CPU intensive quick checks of combinations of design variable values that provide nearest designs meeting the objectives and constraints setup for design. Choice of algorithms is key to ensure best designs are uncovered by traversing through all possible zones in design space in a manner that converges faster to meet the objectives. Reducing the design variables to only those that are sensitive to set objectives is a big plus.

Unlike the output of solvers in traditional computer aided engineering (CAE), which is large data, this data science-based approach has an opportunity to store the information of all the current and past analysis. Statistical data analysis on this data in addition to insights from previous models will provide good inputs to designers. Augmented analytics now allows to merge human intelligence with machine learning to get most of the data produced in the process.

At Intuceo, we have built a metamodel based optimization application using a good balance of all influencing factors, i.e., minimal DoE, fine-tuned surrogate modeling engine and advanced optimization logic.