Optimizing asphalt mix design process using artificial neural network and genetic algorithm

Selection of aggregate gradation and binder content for asphalt mix design, which comply with specification requirements, is a lengthy trial and error procedure. Success in performing mix design rely largely on experience of the designer. This paper presents development of an automatic mix design process with the ability to both predict and optimize asphalt mix constituents to obtain desired mix properties. A successful automatic process requires the use of local past experience translated into a design aid tool, which then predicts properties of asphalt mix without actually testing the mix in laboratory. In the proposed approach, simple multilayer perceptron structure Artificial Neural Network (ANN) models were developed using 444 Marshall mix design data. The ANN models were able to predict both air voids and theoretical maximum specific gravity of asphalt mix to within ±0.5% and ±0.025, respectively, for 99.6% of the time. After that, the ANN models were called by a non-linear constrained genetic algorithm to optimize asphalt mix, while satisfying the Marshall requirements defined in the formulation as constraints. Durability of the optimized mix is ensured by introducing a constraint on adequacy of asphalt film thickness. The developed mix design aid tool is compiled into a computer software called Asphalt Mix Optimization (AMO) that can be used by road agencies as a mix design tool. A case study is presented to demonstrate the ability of the model to optimize aggregate gradation and minimize binder content in asphalt mix. The computed ANN outputs and the optimized gradation were found to compare well with laboratory measured values. Although, Marshall compacted mixes were used in demonstrating the approach, this method is general and can be applied to any mix design procedure.