A neural network model with multi-layer NALU algorithm for the prediction of microstructures of aluminum alloys was carried out. The principle and model of NALU algorithm were explained in detail

and the approaches of modeling and training for the prediction with local grain characteristics were introduced. A satisfying model can be trained with the combination of both optimization algorithm RMSProp and Adam

and appropriate adjustment of learning rate. Taking aluminum alloy casting as instance

26 input factors including alloying elements

processing conditions as well as equipment parameters were concerned

and microstructure parameters containing 18 local grain characteristics were targeted. Hence a neural network model containing 4 layers of NALU algorithm was established with（26

32）、（32

28）、（28

23）、（23

18） as the number of input/output neurons. The mean square error for the trained network model reaches 0.000 6. Prediction error of average grain size of 8 groups of aluminum alloys under different conditions is less than 9%

suitable for intelligent prediction of microstructures and design of new aluminum alloys. The results indicate that it is feasible to construct an intelligent prediction model for the microstructure and properties of alloys by multiple-layer NALU. Multi-layer NALU neural network model can realize the intelligent prediction for exquisite expression of aluminum alloy microstructure by local grain characteristics.