The design of composite laminated panels is a combinatorial problem when the orientation of the fibers in each layer is restricted to a discrete pool of angles. Additionally, composite laminates often have many optimal and near-optimal designs, and the designer may benefit by knowing many of those designs. Genetic algorithms are well suited for laminate design because they can handle the combinatorial nature of the problem and they permit the designer to obtain many near-optimal designs. However, their computational cost is high for most structural optimization problems. This work describes several attempts to reduce the cost of optimizing composite laminates using genetic algorithms.

First, the use of a genetic algorithm to maximize the buckling load of a fixed thickness composite laminate is studied. Various genetic parameters, including population size, probability of mutation, and probability of crossover are optimized by numerical experiments. A new genetic operator - stack swap - is proposed and shown to be effective in reducing the cost of the optimization.

Second, the genetic algorithm is revised and improved for minimum thickness design of composite laminated plates. The influence on the genetic search of the penalty functions enforcing failure constraints is studied. Combining fixed and proportional penalty functions is found to be the most efficient strategy. Improved selection, mutation, and stack swap operators are proposed. The use of an operator called scaling mutation that projects designs towards the feasible domain is investigated. The improvements in the genetic algorithm are shown to reduce the average price of the search by more than 50%.

Next, the improved genetic algorithm for minimum thickness laminate design is applied to a more complex wing box-beam optimization problem. Tuning the genetic algorithm on this problem shows that, because the maximum length of a search is limited, the optimal population size does not grow with the size of the design space. If the probability of applying stack swap is reduced with the number of independent laminates in the wing box, stack swap enhances the performance of the genetic search on the wing box -beam problem.

Finally, the possibility of running many searches is investigated. It is empirically shown that several short searches can be more efficient than a long one, especially when high levels of reliability are required. An example is given where a genetic algorithm is specifically modified for better efficiency in the context of repeated short runs. A procedure is studied that enables predicting reliability at later stages of the search.