Design of engineering materials using hybrid neural-networks and evolutionary algorithms

The design of engineering materials satisfying different performance criteria is an important problem spanning such varied areas as solvents, polymers, additives, and pharmaceuticals. The problem comprises of a forward phase of performance prediction of the engineering material and an inverse problem of construction of the product from desired performance requirements. Real-life industrial problems are characterized by uncertainty in knowledge underlying product performance and the lack of accurate measurements. This precludes the use of a totally fundamental or a completely statistical approach, to performance prediction. In this work, a hybrid neural-network approach was proposed to address the performance prediction problem. The specific domain of application was the intake-valve deposit (IVD) performance of fuel-additives. Deposits on the intake-valve of the automobile affect driveability, cold-start, and exhaust emissions. The primary function of fuel-additives is minimization of deposits by removal of deposit precursors. Under the hybrid approach, a phenomenological model was used to determine functional descriptors directly from the additive structure and fuel characteristics. These were then used as inputs to an artificial neural-network (ANN) to predict the IVD performance of the additive. The hybrid model proved to be more accurate than traditional models, while using fewer descriptors. Evolutionary algorithms based on Darwin\u27s theory of evolution have proved successful in property-based polymer design. They were adapted to the fuel-additive domain by use of specialized representation and constrained operators. In conjunction with the hybrid ANN model, the evolutionary search was successful in identifying novel and optimal additive structures, with reasonable synthesis potential. The sensitivity of the evolutionary search to its internal parameters and the search-space were examined through parametric sensitivity and search-space characterization studies. Constrained genetic operators and adaptive parameter tuning schemes were implemented to add efficiency and robustness to the stochastic genetic search. The evolutionary search was brought under an interactive framework, to alleviate problems due to its stochastic nature. Under this framework called GENESYS, the designer was allowed to interact with the design system. Structural characteristics of initial and evolving solutions, weights on different objectives, and the objectives themselves could be modified. This allowed for an implicit inclusion of experiential design knowledge, towards improving the convergence problems of evolutionary algorithms