challenges because of very thin changes between pore pressure and fracture gradient, frequently below

equivalent mud weight of 0.3 ppg. Traditional rule-based MPD tuning algorithms fail to hold optimal setpoints

This paper introduces a new reinforcement learning (RL) model of intelligent MPD set-point control, which is

more specifically backpressure and equivalent circulating density (ECD) optimization in narrow-margin

deepwater wells that characteristic of the Gulf of Guinea area, such as Ghana and Nigeria. A Deep Deterministic

Policy Gradient (DDPG) model was optimized using past MPD operational data in West African deepwater

campaigns which included a multi-objective reward function balancing between influx risk, loss risk, and ROP

optimization. The paradigm shows that the pressure control accuracy (23% improvement over rule-based

approaches) and mean absolute pressure deviation (42 psi to 32 psi) are lower in comparison to rule-based

The world chase of the hydrocarbon resource has moved drilling activities towards the deepwater environment

with intense geology such as the high-pressure, high-temperature (HPHT) oil reservoirs with small working

conditions. These issues are seen in the Gulf of Guinea, which contains the prolific hydrocarbon provinces

offshore Nigeria and Ghana, and the deepwater operations have added significant production capacity to the

region. The deepwater fields in Nigeria have a record of contributing more than 800,000 barrels per day to

national production with Ghana having the Jubilee, TEN and Sankofa fields that contribute around 160,000

situations can take place. Traditional drilling techniques are often inadequate in addressing such thin margins of

The making of managed pressure drilling (MPD) has become the supporting technology in overcoming such

hazardous conditions. In all its drilling activities, MPD makes use of dynamically controlled surface equipment

such as choke valves and auxiliary pumps to ensure total control of the bottomhole pressure (BHP). The method

allows the real-time management of the pressure, limiting the potential influx and loss circulation and permitting

the drilling works to be performed in the previously undrillable formations. The performance of MPD operations

is however critical on the accurateness and responsiveness of the set-point control system controlling the

Current MPD control systems are primarily based on proportional-integral-derivative controllers or model

The recent development in artificial intelligence, especially reinforcement learning (RL) has provided good

alternative solutions to the complex control problems in the drilling activities. Deep RL algorithms have shown

an impressive ability in optimization of drilling parameters such as weight on bit (WOB) and rotary speed

(RPM) to optimize ROP without causing downhole vibrations. The Deep Deterministic Policy Gradient (DDPG)

algorithm, which is algorithmically tailored to continuous action space, has been demonstrated to be of special

interest to drilling tasks in which the outputs of control systems need to be continuously tuned, as opposed to

being chosen between discrete values. The given research fills an important gap in the literature by creating and

testing a reinforcement learning model that is specifically tailored to MPD set-point control in narrow-margin

HPHT deepwater wells. The paper is dedicated to the Gulf of Guinea working situation, where historical data

second, this paper is supposed to show the quantifiable benefits of pressure control accuracy and drilling

performance that are provided by the framework in comparison with the rule-based approaches to MPD tuning;

and third, this paper should prove that the framework will be able to detect early kicks and respond during

Managed pressure drilling is a modern technology that was aimed at solving the difficulty of drilling in HPHT

reservoirs often encountered in deepwater conditions. The technology offers accurate annular pressure profile

controlling by manipulating surface backpressure using automated choke systems to give operators the

capability of maintaining constant bottomhole pressure (CBHP) within narrow operational windows. The basic

principle is to seal the annulus with a rotating control device (RCD) and supply flow via a choke manifold to

that drilling in less than 0.3 ppg EMW PP-FG window formations was possible at target depth, and that the

technology would allow it to obtain wells that could not have been reached by conventional techniques. On the

case studies can define MPD as a developed technology which can cope with extreme challenges in drilling but

The automation process has been developing over the last twenty years in the system of the MPD choke control

as it shifted to the period of the manual control systems to the modern system of the model-based control. Most

modern MPD control systems use hydraulic models to calculate real time downhole pressure and feedback

dependent on the outlook of the model, and discrepancies between the model and the real system results in poor

It has been suggested that nonlinear MPC (NMPC) strategies are required to deal with nonlinearities inherent to

MPD systems, especially when operating in two-phase flow conditions which appear when there is a kick event.

The Memorial University researchers have come up with elaborate solutions of simple PID controllers to

complex NMPC with fault management capabilities, which have been experimented both at the simulation and

lab level. Despite this progress, there is still much work to be done before robust control performance can be

attained throughout the complete range of conditions experienced in HPHT deepwater drilling, and especially

Machine learning applications have proved to be very successful in other drilling applications that include

prediction of ROP, estimation of ECD, and optimization of drilling parameters. Gamal et al. (2021) used

artificial neural networks (ANNs) and adaptive network-based fuzzy inference systems (ANFIS) to predict ECD

Random Forest models have demonstrated especially good performance in ECD prediction with the authors

showing R

of 0.9859 and RMSE of 0.0017 in testing datasets with input of parameters of surface drilling (Gao

et al., 2024). The benefit of surface measurements only removes the use of expensive downhole sensors yet still

has a high prediction accuracy. These methods can be immediately applied to MPD control systems and, in that

Reinforcement learning is a new paradigm to the automation of drilling, which promises the possibility of agents

learning how to act in the best control policy when interacting with their surroundings instead of being

programmed to do so. For instance, a virtual drilling agent developed by Huang et al. (2024) is based on DDPG

best solution to different drilling conditions including hard formation, embedded rock and unstable drilling

The DDPG algorithm belongs to the type of actor-critic algorithms for RL, which combines value-based and

policy-based approaches for dealing with actions that have infinite dimensions. The algorithm keeps two neural

networks, an actor network which computes the best action based on the current state, and a critic network

which approximates the value of state-action pairs. This is particularly suited to MPD control systems wherein

Keshavarz et al. (2024, 2025) suggested deep reinforcement learning representations of the real-time planning

of drilling operations with Markov Decision Process formulations and the application of the Gaussian process

algorithm to determine the safe operating window. Their work showed that in the process of wellbore cleaning,

Although, RL methods have been used to optimise the drilling parameters (WOB, RPM) their implementation

to MPD set-point control has not been investigated in the published literature. Moreover, the current RL drilling

programs are mainly devoted to single-objective optimization (usually, the ROP maximization), whereas the

MPD control should be considered in a multi-objective way and involve the influx risk, loss risk, and drilling

This paper fills these gaps by creating a DDPG-based framework that has been tailored in MPD set-point control

in narrow-margin HPHT wells. The new contribution consists of: (1) the MPD control problem is formulated as

a continuous-action RL problem; (2) a multi-objective rewarding function balancing the risk of influx/loss and

ROP maximization is developed; (3) the problem is trained and validated using data that is representative of Gulf

of Guinea deepwater operations; and (4) the problem is quantitatively compared to rulebased MPD tuning

flow rate (GPM), standpipe pressure (SPP), weight on hook (WOH), rotary speed (RPM), rate of penetration

(ROP), mud density (ppg), and choke position (%). The target variables in this work are bottomhole pressure as

well as PWD tools and ECD from downhole measurements. Extra derived features would be flow-in/flowout

Preprocessing of the data was done by the standard drilling data quality assurance. Values from null records,

sensor malfunction, or redundant data points were dropped, making valid observations 15,840. Outlier detection

and removal was done by the Interquartile Range (IQR) method, which removes about 4% of records that

contained anomalies in the pressure or flow measurements. All features were normalized to a z-score to

Stratified sampling was used to divide the dataset into the training (70%), validation (15%), and testing (15%)

subsets to maintain representative representation of well types and formation characteristics by the partitions.

Sensitivity analysis on the hyperparameters showed that the weights of the reward functions had a big impact