Design of fuel additives using neural networks and evolutionary algorithms

It is difficult and challenging to design high-performance fuel additives i n an industrial-design setting where data are sparse and noisy, and fundame ntal knowledge is often limited. An automated framework is presented for th e design of such fuel-additive molecules that minimize the intake-valve dep osit in the automobile. A hybrid model that combined functional descriptors from a first-principles degradation model with a statistical/neural-networ k model was developed to predict additive performance, given the additive s tructure. The results of the predictive model are discussed for differentia l industrial case studies. An evolutionary method using specialized represe ntation and constrained operators to enforce formulation constraints was us ed to generate optimal additive molecules that meet desired performance cri teria. The evolutionary design strategy in combination with the hybrid pred iction model was successful in identifying novel additive molecules that al so possess good synthesis potential.