Saturday, October 9, 2010

SDM demystifies multidisciplinary system design optimization - SDM Pulse Fall 2010

By Genevieve Flanagan, SDM ’10

Genevieve Flanagan
SDM ’10
The concept of “model-based design” has generated a lot of buzz in recent years, primarily because virtual simulations and analysis can be used in conjunction with our existing product development processes to release a better, less expensive product in less time. (Figure 1 shows an idealized view of how this might be applied in a product development cycle.)
In practice, however, the application of model-based simulation and design has typically been narrowly scoped and focused on a single objective. While this works well during the early design phase of a component or subsystem, it is less helpful during early system-level specification development or system integration phases.

SDM’s class in multidisciplinary system design optimization provides a fresh design approach to developing complex engineering systems, one that combines both technical and qualitative objectives. Taught by Associate Professors Olivier de Weck and Karen Willcox, the course begins by defining the architecture of a model of a multi-disciplinary system. Design variables and objective functions are identified and categorized into subsystem disciplines. Dependencies are described in a N2 matrix format, and then optimized to reduce the coupling.

Figure 1. In this idealized view of model-based design,
development in the virtual domain parallels physical
development, speeding the product to market.
Published with permission from John Deere
With the model defined, the next part of the course is to examine single-objective optimization. Numerical techniques (i.e. Newton’s Method and Steepest Descent) and heuristic methods (i.e. Genetic Algorithms and Particle Swarm Optimization) are introduced. These are used to find the optimal design vector to satisfy a single system objective.
Once the basics of these optimization techniques are understood, the course moves onto multi-objective optimization. Using many of the same algorithms learned earlier, the trade-offs among the multiple objectives of a system are realized by displaying the relationship between objectives as their common design variables change within a design space. Simulation analysis methods are presented to help students understand the success of the simulation and design variable sensitivity.
Students complete five assignments during the semester to reinforce their understanding of course concepts. However, the class primarily revolves around a semester-long group project, which is generally taken from thesis research or from industry.
Figure 2. This graphic shows that very different
results come from optimizing on two separate
objectives independently—cost to manufacture
and total cost of ownership.
My group for the project included Justin Kraft (SDM ’09), a senior engineer for John Deere. We chose to work on optimizing the design of a battery powered autonomous vehicle for John Deere. A model of this product existed, but it was narrowly scoped. We supplemented the model with additional disciplines to add to its system-level functionality. Our new subsystems then used information from the initial model to input new cost, reliability, and other functions, resulting in no loss in technical fidelity while providing new quantitative and qualitative features to the model.
When we ran the single-objective optimization routines, the flaws in that type of analysis were clear. Optimizing on two separate objectives independently—cost to manufacture and total cost of ownership—led to two very different design vectors for the vehicle (see Figure 2). Reduced manufacturing costs resulted in a low-powered small vehicle while total cost of ownership focused on increasing efficiencies and battery life.
In the end we chose a multi-objective optimization using a heuristic genetic algorithm to visualize the trade-off between the two cost goals resulting in a Pareto front (essentially the balance point at which you cannot make progress toward one goal without detracting from the other). This chart (see Figure 3) showed the optimal necessary system design options and allowed us to evaluate the most appropriate design choices. Further analysis revealed design vector sensitivity so that we could see what design variables and model parameters would most affect our result—and where additional concessions could be made with minimal impact to the objectives.
Figure 3. Multi-objective optimization reveals the Pareto
front (or balance point between goals), making it easier
to visualize the trade-offs between cost goals.
An analysis like this one, which reveals design parameters and how the trade-offs necessary to reach multi-disciplinary objectives are achieved, is at the heart of our idealized view of model-based design. This is precisely the sort of information that so often is not available early in the process when its impact would be most useful.
Although this course required a significant investment in time and thought, I have gained the tools necessary to take complex problems out of the purely heuristic realm in order to address very real situations. Overall, I think multidisciplinary system design optimization is valuable for those interested in embracing model-based design to improve product quality, delivery, and production time.

No comments:

Post a Comment