5:15 PM - 6:30 PM
[MIS18-P03] Machine Learning Approaches for Predicting the Lithology and Core Recovery Rate
Keywords:Scientific drilling, machine learning, Surface drilling data
As primary aim of scientific drilling is to recover core samples from sediment layers under the seabed, improving core recovery rate is very important because it can enhance the operation efficiency as well as scientific examinations using recovered samples. Obtaining the information of expected core recovery rate is helpful for operating the drilling equipment to improve the recovery rate. In addition, obtaining the information of lithology of drilling layer is also helpful for both scientific and operational aspects. Even an approximation could potentially provide valuable information for conducting coring operations. However, there is no direct information regarding the core recovery rate and lithology. The recovery rate and lithology can be known after retrieving a coring tool.
This study focuses on the discussion of possibility to know the sediment property, lithology, and core recovery rate from the surface drilling data, that is only the information monitored during the operations, from analytical and machine learning approaches.
A previous study proposes an analytical method to provide the shear stress of drilling layers as an important property. In this study, we first summarize the method applied in the previous paper and attempt for NanTroSEIZE operation in 2018-2019. The result shows that the estimated shear stress is in very good agreement with the shear stress measured by LWD operations.
A main topic of the paper is to predict the core recovery rate and lithology by applying the state-of-the-art deep convolutional neural networks. Learning data was created from surface drilling data and information of lithology and core recovery rate from recovered core samples. Machine learning was then applied and prediction model was created. During creating the model, we investigated the effect of input length window and network hyper parameters such as convolution kernel widths, number of filters, and network depth for the validation performance. The results, that are obtained from the best model, shows high classification performance between 75 to 95% for each lithology.
The model also provides an adequate performance for predicting the core recovery rate. On the other hand, core recovery rate seems to be highly influenced by the drill bit motions. However, since there is no direct information, we established the analytical models of drill bit motions such as torsional vibration and vertical motions, and prepare the learning data by adding the analytical data to the surface drilling data. From the results, it can be observed that adding analytical data may improve the prediction performance.
Although high prediction performance was achieved by applying the machine learning, it can be argued that the network structures for both lithology classification and core recovery prediction are very complicated, and the prediction performance is sensitive to the hyper parameters. This reveals that the prediction allows for difficult classification. It was also noted that the number of data sets of learning data seems not sufficient for making firm conclusions. We will continue further investigation.
An important topic in the future is the real-time prediction during the operations of Chikyu. Toward it, we are developing the drilling data acquisition system and real time data analysis system. We incorporated the prediction model into the system, and we attempted the real-time prediction of lithology and core recovery rate during the NanTroSEIZE in 2019. Unfortunately, due to the operational restrictions from the viewpoints of the schedule and hole conditions, a coring operation was not conducted during the real-time prediction test. Through this attempt, it is confirmed that the system with the prediction models seems to provide adequate range of predictions, and operates sufficiently well without any trouble throughout the operation.
This study focuses on the discussion of possibility to know the sediment property, lithology, and core recovery rate from the surface drilling data, that is only the information monitored during the operations, from analytical and machine learning approaches.
A previous study proposes an analytical method to provide the shear stress of drilling layers as an important property. In this study, we first summarize the method applied in the previous paper and attempt for NanTroSEIZE operation in 2018-2019. The result shows that the estimated shear stress is in very good agreement with the shear stress measured by LWD operations.
A main topic of the paper is to predict the core recovery rate and lithology by applying the state-of-the-art deep convolutional neural networks. Learning data was created from surface drilling data and information of lithology and core recovery rate from recovered core samples. Machine learning was then applied and prediction model was created. During creating the model, we investigated the effect of input length window and network hyper parameters such as convolution kernel widths, number of filters, and network depth for the validation performance. The results, that are obtained from the best model, shows high classification performance between 75 to 95% for each lithology.
The model also provides an adequate performance for predicting the core recovery rate. On the other hand, core recovery rate seems to be highly influenced by the drill bit motions. However, since there is no direct information, we established the analytical models of drill bit motions such as torsional vibration and vertical motions, and prepare the learning data by adding the analytical data to the surface drilling data. From the results, it can be observed that adding analytical data may improve the prediction performance.
Although high prediction performance was achieved by applying the machine learning, it can be argued that the network structures for both lithology classification and core recovery prediction are very complicated, and the prediction performance is sensitive to the hyper parameters. This reveals that the prediction allows for difficult classification. It was also noted that the number of data sets of learning data seems not sufficient for making firm conclusions. We will continue further investigation.
An important topic in the future is the real-time prediction during the operations of Chikyu. Toward it, we are developing the drilling data acquisition system and real time data analysis system. We incorporated the prediction model into the system, and we attempted the real-time prediction of lithology and core recovery rate during the NanTroSEIZE in 2019. Unfortunately, due to the operational restrictions from the viewpoints of the schedule and hole conditions, a coring operation was not conducted during the real-time prediction test. Through this attempt, it is confirmed that the system with the prediction models seems to provide adequate range of predictions, and operates sufficiently well without any trouble throughout the operation.