PhD opportunities

Drill Bit behavior detection with Machine Learning: Application to event detection and performance optimization

Thesis proposal

Area of expertiseGeoscience and Geoengineering
Doctoral SchoolGRNE - Gťosciences, Ressources Naturelles et Environnement
SupervisorM. Thomas ROMARY
Co-supervisorM. Laurent Marc GERBAUD
Research unitGeosciences
ContactGERBAUD Laurent
Starting dateOctober 1st 2021
KeywordsPhysic informed Machine Learning, Drilling, Mechanic
AbstractThe profitability of geothermal energy requires a drastic reduction in its costs. Drilling costs represent 30 to 50% of the total cost of the project and remain the main obstacle to its development. These costs should therefore be reduced as much as possible by drilling quickly and over a long period of time in basement, metamorphic rocks where drilling speeds (ROP) are very low, in the order of 1 to 2 m/h, with the risk of material breakage due to extreme conditions (temperature, vibration, pressure, etc.). Drill bit behavior is one of the most challenging topics for increasing ROP and its management is often more important than the bit itself. It is therefore fundamental to know in real time how the bit behaves (dysfunction detection, safe drilling) and what the optimum parameters are.
The geosciences department has developed and patented a process that allows monitoring drilling dysfunctions at the bit in real time in order to deal with them and thus increase drilling speed. Recent internal works using Clustering techniques and Convolutional Neural Networks show that drilling dysfunction such as whirl or Stick Slip vibrations can be accurately detected. Finally, the geosciences department has gathered a broad expertise in drill bit behavior through the development of a 3D bit rock interaction software (DIG3D), models of cutter-rock interaction and a huge amount of data from more than 20 years of drilling experience in its experimental drilling platform.
As part of this thesis, we wish to develop a new approach to monitor and optimize drilling in real time based on machine learning. This work will consist of two main components:
- The first step will be to develop a supervised Machine Learning model to predict the drill bit behavior in a controlled environment (laboratory drilling). The model will be trained with data from 20 years of experimental laboratory drilling tests as well as theoretical results from 3D simulations. One of the solutions explored will be the Physics Informed Neural Networks. This approach uses a modified loss function to train the neural network, to ensure that the model predicts the observations while honoring the laws of physics. The auto-differentiation (back-propagation of the errors) within the neural networks allows to estimate the parameters. This approach is very attractive and will be extended and modified to be applicable in the context of geothermal drilling, using experimental and numerical simulations data. Once the model trained, it will be used both to predict the bit behavior and/or to predict the optimal parameters by inversion.
- Disrupted by many untapped phenomena, it has been shown that drill bit behavior cannot be estimated with surface field data and that it is difficult to detect all phenomena affecting the penetration speed. The second step will be to use transfer learning techniques to develop a new model for the field where the model developed from the lab data will be used as a starting point.
The expected results of the thesis are the development of a procedure composed of learning models that can detect major drill bit events during drilling and provide the optimal drilling parameters for fast, long and safe drilling. This procedure should allow real-time monitoring in the first instance and should pave the way for drilling automation in the second instance.
ProfileMathematics, Mechanics or Civil Engineering, IA.
FundingConvention CIFRE