A new method has been developed to predict the undrained shear strength from geophysical logs and core measurements using neural networks. A limited number of undrained shear strength measurements are performed on cores, resulting in a limited, discontinuous data set. A series of neural networks were developed to predict the undrained shear strength of the sediments where measurements were not available. The prediction is based on a learned relationship between the suite of geophysical logs and the measurements of undrained shear strength.
The use of neural networks to predict the undrained shear strength in cores requires that a statistically significant number of shear strength measurements are available to train and test the neural network. Geophysical logs are necessary to develop a relationship with the shear strength measurements. Shear strengths were measured on core samples in intervals ranging from 15 cm to greater than 200 cm. The neural network was able to resolve the shear strength measurements to a depth resolution of approximately 8 cm. The predicted undrained shear strength values, when tested against actual shear strength measurements, show good correlation.
The expense involved in drilling and coring marine sediment in the world's oceans for scientific research is high. Testing and sampling of each core section must be prioritized and shared within the scientific community. Non-destructive testing of sediment core complements sediment sampling, accessing more information from each core section than from sampling alone. The multi-sensor core logger (MSCL) is an instrument that non-destructively measures the physical properties of sediments, such as bulk density, resistivity, compressional wave velocity and magnetic susceptibility at very high depth resolution. Each core section from the Integrated Ocean Drilling Program (IODP) and the Ocean Drilling Program (ODP) is run through the MSCL. In addition to the MSCL, in-situ non-destructive measurements are made through downhole wireline logging. Wireline logs give a continuous record of several geophysical properties.
Traditionally, however, measurements of the undrained shear strength (Su) of a sediment require that the test disturbs the sediment. To maximize the amount of undrained shear strength data available from a borehole, a neural network computer program was used to estimate Su values based on a learned relationship between Su measurements and the MSCL and downhole wireline log data available. If only a limited number of Su measurements are possible, the neural network program is used to fill in the gaps of Su data where measurements were not possible.
A neural network is a pattern recognition or function approximation computer program that emulates the learning and predictive behavior of the human brain, in which future responses or actions are dictated by the outcome of previous experiences (Haykin, 1994).
A typical neural network consists of a series of input data, one or more hidden layers of neurons that give a weight to the data inputs, gauging their importance, and an output layer.
