On the base of samples taken from ash-sludge collector of Pavlodar Aluminum Plant we have created neural network for making forecasts of concentration distributions of different elements compounding production waste of the plant. For every analyzed element separate neural network was created. Levenberg-Marquardt algorithm was chosen for training. Architecture of neural network includes 5 layers, where one layer is input, one – output and three between them are hidden layers. Neural network demonstrates high accuracy on all of three samples of data obtained by means of partitioning of samples taken from different locations of the lake. Much higher concentration in every location is observed for Silicon (Si), Calcium (Ca), Cuprum (Cu) and Ferrum (Fe). The less concentrations were obtained for Manganese (Mn), Vanadium (V), Titanium (Ti), Scandium (Sc), Gallium (Ga). Accuracy of neural network calculations depends on setting parameters such as number of layers, training algorithm.