Multi-Objective Evolutionary Algorithm Archive for Design Optimization

The patent performs preliminary actions by evaluating multiple design alternatives in parallel using high-performance computing resources before the optimization algorithm begins. This pre-evaluation creates an initial population with known performance characteristics, reducing the number of expensive simulations needed during the optimization process to achieve convergence to the Pareto optimal front.

The patent uses approximation models or surrogate models that create simplified copies of the complex simulation models. These surrogate models can rapidly evaluate design alternatives without requiring full finite element analyses, thereby maintaining solution diversity while dramatically reducing computational cost. The surrogate models are trained on a subset of expensive simulations and then used to guide the evolutionary search.

The patent implements adaptive feedback mechanisms that continuously monitor the convergence behavior of the evolutionary algorithm. Based on this feedback, the maximum number of generations is dynamically adjusted - extending the optimization when convergence is slow and maintaining high precision requirements, while reducing generations when convergence is rapid or when computational resources are constrained. This feedback loop balances productivity and precision based on actual optimization progress.

The patent transforms the static maximum generation limit into a dynamic parameter that adapts during the optimization process. The generation limit is adjusted based on convergence criteria, resource availability, and the complexity of the design space being explored. This dynamic approach allows the system to achieve sufficient precision with fewer generations in simple problems while allocating more generations to complex problems requiring higher precision.

The patent segments the optimization process into multiple phases with different simulation requirements. Early generations use coarser models or simplified evaluations that require less time, while later generations progressively use more detailed models as convergence approaches. This segmentation allows the algorithm to explore the design space efficiently in early stages and refine solutions in later stages, achieving reliable convergence without uniformly high computational cost across all generations.

The patent implements periodic evaluation intervals where full expensive simulations are performed only at selected generations rather than every generation. Between these periodic full evaluations, the algorithm uses surrogate models or simplified assessments to guide the search. This periodic action maintains reliability by periodically verifying convergence with accurate simulations while reducing overall time loss by avoiding expensive simulations at every single generation.