(Special session) MMIJ long-term thematic project - Stat up of second period and Follow up of first period

[3507-19-12] Current approach and future perspective towards understanding the copper matte-slag formation using high-speed imaging and deep learning

○Shungo Natsui¹, Yuko Goto^1,2, Jun-ichi Takahashi², Hiroshi Nogami¹ (1. Tohoku University, 2. Sumitomo Metal Mining Co., LTD.)

Chairperson：小池 克明（京都大学）

Keywords:copper smelting, suspended combustion test, high-speed microscopic videography, convolutional neural network (CNN)

An analysis based on a convolutional neural network (CNN) was carried out to classify the different combustion patterns of Cu concentrate-SiO₂ mixtures tablets under oxidation gas to estimate their combustion behavior and phase changes. A suspended-combustion-test method involving high-speed digital microscopy and thermal measurements was employed to characterize the combustion behavior of each sample. The time series images-based pattern recognition method enabled the calculation of the chemical composition of the blended concentrates by transforming the network output into a probability distribution. The combustion of the blended-concentrate tablet was different from that of each single-concentrate tablet in terms of the combustion pattern like shapes of the molten part, and the temperature profiles. It can be interpreted that the change in the free surface shape of a tablet plays an important role in combustion pattern recognition, therefore when blended samples were used as training data as well as single samples, a good correlation could be obtained between the measured and predicted values of its chemical compositions. Predicting the combustion patterns of Cu concentrates by constructing a CNN database comprising further experiments with various Cu concentrates and blending conditions should be possible.