Application of a feedforward neural network in the search for kuroko deposits in the hokuroku district, Japan

摘要: A feedforward neural network with one hidden layer and five neurons was trained to recognize the distance kuroko mineral deposits. Average amounts per hole of pyrite, sericite, gypsum plus anhydrite as measured by X-rays in 69 drillholes were used train net. Drillholes near between Fukazawa, Furutobe, Shakanai mines used. The training data selected carefully represent well-explored areas where some confidence ore assured. logarithmic transform applied remove skewness each variable scaled centered subtracting median dividing interquartile range. learning algorithm annealing conjugate gradients minimize mean squared error ore. then all 152 that had gypsum, pyrite. contour plot net predicted shows fairly wide 1 km or less ore; known deposit groups is within contour. high low distances on margins contoured are part result boundary effects contouring algorithm. For example, short west (Hanaoka) deposits basement. However, northeast Furotobe, just off figure, coincide location Nurukawa Omaki deposit, south Shakanai-Hanaoka deposits, seems be an extension contour, but beyond 3 limit from drillholes. Also interest only a few kilometers Fukazawa estimated many ore, apparently reflecting network's recognition extreme local variability geology