Analytical instrumentation
Continuous catalytic reforming is a conversion process in which many reactions occur with heat stages in between each reaction which is completed to convert the heavy hydrocarbons into converted into lighter products.
Real-time optimization (RTO) occurs in which numerous metrics are measured to predict optimal conditions for the maximum yield of various chemical products such as BTX or H2 to be produced as the qualifying measurements produced by the refinery may be inaccurate due to sampling delays, blending during transport, etc.
However, there are numerous machine learning algorithms which can be used to conduct RTO more efficiently than traditional nonlinear problem solutions.
The hybrid first-principles model, the NOMAD algorithm, and the TD3 algorithm are a few examples of ML being used to conduct RTO.
The results show that transfer learning algorithms relate to faster convergence and more model stability.
The paper also explores safe learning, in which constraints are upon the model to follow while multiple conflicting objective functions are fed.
A paper comparing a complete ML approach to RTO vs a nonlinear approach to RTO and a model predictive controller(MPC) layer controlled by H-DDPG algorithm show that a complete ML approach to RTO can lead to similar results as those with nonlinear equations, but in faster computational times.
This review also explores the challenges with machine learning such as data quality, model quality, surrogate model quality, security risks, and environmental risks.
The review then calls for more research on how wide application of these machine learning models can be made possible in the continuous catalytic reforming process and how to standardize model quality standards across research labs as the evidence shows how ML models are efficient and less computationally costly than nonlinear models.
There are three main processes involved with crude oil processing: separation, conversion and treatment [1].
Separations are made to crude oil through distillation of numerous hydrocarbon molecules within the oil [2].
The conversion process is completed through several processes including hydrocracking, catalytic cracking, etc.
Treatment occurs when various streams of chemicals, from the distillation process is combined appropriately.
Catalytic reforming (CR) is conversion process in which hydrocarbons are molecularly rearranged to become higher octane forms [3].
Continuous catalytic reforming (CCR) occurs with catalyst entering and exiting numerous reactors causing multiple simultaneous reactions occurring with heating stages in between each reaction before the product is separated using downstream fractionation [1].
One issue within the typical refinery is the constant change of composition of crude oil, and properties and how to change internal refinery settings in time to appropriately affect the oil composition.
There are various parameters which cover assay properties, compositional metrics and operational targets that oil needs to be tested on to ensure safety, efficiency and quality of output: cut points, distillation profile, API gravity, aromatic content, asphaltenes, Total Acid Number (TAN) and sulfur content [4].
However, there are many reasons as to why refineries have a difficult time getting these metrics to be accurate: sampling delays, blending during transport, and limited real-time compositional visibility [4].
These issues can be addressed with the help of real time optimization (RTO), with consistent continuous monitoring of crude to help detect anomalies or deviations in oil composition which could lead to mishaps.
Real-time optimization (RTO) can assist in managing these issues by being trained with previous refinery data and having numerous internal variables being placed in the objective function for the algorithm to optimize (A Survey of Optimization Methods from a Machine Learning Perspective).
However, the real-time optimization process’ performance depends on model accuracy, adequacy and online maintenance [5].
Continuous, feed-forward machine learning models of feed analyses can lead to at least 3-5% increase in operational efficiency [4].
The supporting technology of the Nuclear Magnetic Resonance (NMR) technology, and the On-line Process NMR (OP-NMR) in addition to the RTO ML model can lead to an increase in productivity [4].
The following technical review will cover the advantages of using optimization algorithms for real-time optimizers, comparisons to other models and the challenges involved in its implementation in real-time-optimization through perspective analysis.
There are many kinds of optimization approaches that can be used to implement RTO.
One of the kinds of optimization approaches involved in Dong et.al’s study was genetic algorithms in optimization models.
The main issue that Dong et.al was trying to resolve is that there are many deviations from optimal conditions in the crude oil upstream such as the composition and the flow rate of the naphtha feedstock [6].
This leads to less-than-ideal solutions created by the deterministic optimization models, which do not take these deviations into account [6].
Dong et.al used the minimax and Maximax criteria with an alpha = 5% to evaluate the objectives and based the model off one month of CCR data [6].
They then used Latin hypercube sampling method to sparse variability of feedstock within the multi-objective optimization model [6].
After that, they used a surrogate-based optimization method to solve objective operational optimization equations using genetic algorithms involving conflicting interests such as decreasing energy consumption and maximizing aromatic production (aromatic production is endothermic) [6].
In the dynamic optimization approach, the non-dominated point search was conducted after optimization and if convergence occurs, the solutions are finalized [6].
Else, the surrogate model is iterated and the cycle continues [6].
This entire algorithm (at least the dynamic approach) was robust as multiple statistical modeling techniques were involved in each iteration [6].
The optimal conditions found within the study were then evaluated through two objective optimization (yield of low mass aromatics maximization and minimization of energy consumption) and the four objective optimization problems (maximizing yield of low mass aromatics (Ya) and H while minimizing heavy aromatics and energy consumption).
This led to aromatic production increasing 1.05% and an energy consumption reduction of 8.8 GJ/h from original conditions for the two objective problems and 0.89% and 0.08% improvement in the four-objective-optimization problem [6].
The optimization model discussed in Dong et al’s paper was difficult to implement because it required sampling of the uncertainty which can lead to exponential computing cost increases.
Fortunately, this issue can be allayed with reinforcement learning which incorporates sampling.
Reinforcement learning involves an AI agent which will train and learn about the environment and does trial and error to achieve optimal results.
Guo et. al used real-time optimization methods involving reinforcement learning and transfer learning to optimize the continuous catalytic reforming process to increase quality of feedstocks which real time optimization methods may be hindered by due to model mismatches [7].
There are many studies focused on CCR, and process and parameter optimization, but RTO implementation has challenges.
Naphtha, which is the main part of the reaction in the CCR process, has 300 components which lead to differences in reaction rates and product amounts varying significantly depending on processes [7].
The relationship between operating conditions and products is very complex, and so making a correlation for the model is difficult.
The energy consumption of CCR is dependent on inlet temperature of the reformer and product distribution [7].
Guo et.al made a model of the CCR process and trained the RL model [7].
They then made the environment and agent, implemented the TD3 algorithm, in which the following variables were evaluated: inlet temperature of reactor #1,#2,#3,#4, feed flow rate, H/HC ratio, yield of H2 and benzene and toluene, carbon deposit of the used catalyst, and load of furnaces (four were involved for this process) [7].
After this process, the algorithm changes its policy based on the temporal difference error between the output of the state-action value function and the target value [7].
The TD3 algorithm uses the double Q-learning mechanism to overcome the overestimation problem present in DDPG algorithms (as TD3 is a variant of it) and is trained offline [7].
Unlike in Dong et.al, where uncertainty was sampled and statistical analysis was done, reinforcement learning was involved which reduces computing cost and streamlines optimization.
The algorithm was evaluated through comparison with other models (Fig 1) [7].
During the optimization process, the agent checks how much energy is used, and how much production is needed and adjusts accordingly.
In addition to regular machine learning, transfer learning is also added so that the retraining of the model, occurring when the feedstock properties change, is made to be easier [7].
Dropout layers are also added so that a fraction of connections within the ML model are shut off so overfitting is prevented and minimized [7].
The TL agent has similar performances, but training is much easier than shown on Fig 2, as convergence is achieved faster, which is also corroborated by Gao et. al (see Fig. 1).
The average reward with the TD3 agent was around 310, and the R^2 of each value exceeded 0.993 [7].
In addition, the table below (Table 1) shows that other machine learning programs/algorithms relate to similar data to the TD3 agent, which corroborates proper performance of the model [7].
While real-time-optimization is the traditional method of finding optimal refinery parameters, it has issues regarding naphtha and model mismatches, which can be reduced with the assistance of TL and RL, making training faster and efficient.
But Guo et.al are not the only ones studying this over the past few years.
In fact, Pasandide and Rahmani were the first to consider the deactivation of the catalyst, catalyst circulation rate, simulation and optimization of purification processes within the continuous catalytic reforming process [8].
Numerous variables such as the 1st catalytic bed inlet temperature, flash drum temperature, flash drum pressure were involved in the objective function(f) and constraints(g) and this equation was solved with the NOMAD algorithm, in which the vectors are within the constrained values and are sent to the optimizer until convergence is satisfied [8].
There were two kinds of studies being conducted.
One focused on the high-quality of the gasoline product which led to RON and furnace load being the main two variables being optimized for.
The second focused on more aromatic products being made.
Below are some of the optimized variables. (Table 1).
The results from the model show that coking reduced 24.5% and energy costs dropped 13.5% because of involving these variables in the objective function.
These studies corroborate similar results when conducting RTO which means that these machine learning algorithms are consistent with industry operations.
There are other factors which negatively affect, but that is what this table is suggesting.
Pasandide and Rahmani showed that variables can be placed in the objective function and appropriate constraints to calculate the optimal vectors for minimal coking and energy costs.
However, there were only two objectives in one objective function.
In Ma et.al’s paper, there are four objectives are present within two objective functions in the CCR process (maximizing research octane number, minimizing furnace load; maximizing aromatics yield, minimizing furnace load) [9].
There was a complex model present in the refinery named complex reactor mechanism model which was replaced by the 32 Kriging models based off of 32 components and 50 reactions (of which included dehydrogenation, Dehydrocyclization, Isomerization, Hydrocracking, Hydrodealkylation) which were used to predict the flowrate of the product [9].
The continuous catalytic reforming process is represented with a lumped reaction network which simplifies the reactions
The maximum error of the models was 3.59% with the average error being 0.66%, which shows that the cheaper 32 kriging models can replace the complex model, leading to more efficiency in the refinery due to the increase in aromatics and similar optimal conditions to Wang et.al and Pasandide and Rahmani and depicted in Table 1 [9].
While optimizers have been successful in determining optimal conditions for real-time-optimization of continuous catalytic reform, there are more complex forms of machine learning which may be more effective such as a hybrid first-principles model.
Xie et. al studied the effect of using first principles (FP) and hybridization of models with the Seq2Seq Neural network within the two-stage process of nitration of benzene.
There were seven reactors working in a series, with the organic phase going forward in the network while the acid phase was going backward [10].
The final product of dinitrobenzene is made after the products react in the 7th reactor and exit the network while the excess acid leaves the 1st reactor.
Reactor temperatures are the main determinants of process safety and can be measured using skilled operators [10].
Because of reactor temperature fluctuations, the study aims to make a prediction about the reaction temperature 5 minutes before the requested time, with the assumption that future operations in the network were constant, accounting for neglect from manual operators [10].
To make the first-principles model, they had seven main assumptions.
There are three subdivided heat types in each reactor: reaction heat, heat exchange with cooling water and sensible heat [10].
The mixture densities and mixture heat capacities of the material remain constant [10].
Heat exchange between reactor and cooling water is completed instantly.
The inlet and outlet molar flow is constant for each phase.
No side reactions are considered, and both stages are combined into one reaction and reactant concentration change from reactions are negligible.
The first-principles model is dynamically affected by the seven reactors’ temperatures at each timestep, the exogenous input including raw material stream flowrates, cooling water flowrates and inlet temperatures, and parameters which include heat capacity and effective reactor volume combined, and heat transfer coefficients of the cooling water and the heat exchange area.
The FP model is combined with the Unscented Kalman Filter to estimate model states and parameters based on real measurements and generated noise to conduct a co-estimation algorithm based on real-time measurements.
The latest parameters are calculated at each timestep and utilized to predict the temperature at multiple steps in the reaction.
The Seq2Seq network is a Transformer.
It utilizes previous temperatures and external factor data(u) as inputs for the encoder and the decoder uses the predictions from the UKF-FP model and combines the external factors to get the predicted temperature [10].
The network then uses the residuals and adds it to the predicted temperatures to get the final outputs.
The model was trained through a combination of the teacher forcing technique and corrections of the output at the end of the timestep.
The performance of the hybrid first-principles model was compared with the autoregressive exogenous input (ARX) and the Seq2Seq model alone. Unlike the hybrid-models, ARX and Seq2Seq models are purely data-driven [10].
This may have been an important factor in the unsatisfactory results of the ARX and Seq2Seq models.
The average median MSE of the ARX for the training and testing sets were 0.0035 and 0.004 respectively [10].
The median MSE of the Seq2Seq for the training and testing sets were 0.002 and 0.008 [10].
However, the median MSE of the UKF-FP model was 0.003 and 0.0025 respectively [10].
This data shows how the Seq2Seq model overfitted data as shown through the testing MSE magnitude being greater in training.
However, the FP model must be accurate for the hybrid model to be built.
The quantification of which needs to be researched further, investigating the minimum acceptable percentage of accuracy, minimum number of rewards, etc [10].
Also, the applications of this involved in continuous catalytic reforming process, specifically, also need to be researched thoroughly before any conclusions about results can be made as well.
While lump modeling has shown encouraging results like in Ma et.al’s study, the lump model has multiple limitations [11].
These models have difficulty reflecting variations in feedstock properties on compositional shifts within some lumps.
Typical optimization studies, similar to previously discussed, focus on operating conditions and may use statistical modeling to sparse variability (like Guo et.al).
However, while yield of aromatic products is affected by the reactor’s operational conditions, the main cause of the operator condition change is the reactions involved such as dehydrogenation and dehydrocyclization [11].
By adjusting the feedstock conditions and reactor parameters, a broader range of optimal conditions can be found [11].
This requires multi-objective optimization strategies which are balanced in safety while producing effective results [11].
In Shi et.al ‘s paper, the transformer model ML model with geometric Brownian motion was used.
The transformer is used to maximize entropy (TREM method) to improve accuracy of the naphtha composition [11].
The predicted composition of the molecular section of naphtha completed was just 0.084% from the actual composition, which accounts for feedstock variabilities which lumped kinetic models like in Dong et.al cannot cover [11].
At the same time, Shi et.al used the Online Synchronized Learning Multi-Objective Optimization (OSLMOO) method to account for market fluctuations to make an appropriate Pareto front which includes hydrogen yield, light aromatic profit, and heavy aromatic yield, which can lead to better profitability of CCR [11].
MOO does have some challenges, however.
If the objectives involved in the MOO are conflicting like in Shi et.al, then the improvement of an objective function could lead to the degradation of another objective [12].
Because of this, a non-dominated solution is found which may not be the optimal solution possible in the refinery, which could lead to missed opportunities for more profitability.
Also, if there are too many object functions, this could lead to all solutions being non-dominated, results which are difficult to interpret due to visualization being beyond 3D, or scaling issues involved with sample points as the required number of points needed to approximate the Pareto front exponentially increases with objective function increases [12].
While Guo et.al had incorporated RL and TL in their real-time-optimization algorithm developed for continuous catalytic reform, they had not explicitly compared the traditional real-time-optimization with nonlinear equations along with a model predictive control (MPC) layer to that of a machine learning algorithm being foundational behind the RTO and the MPC.
Ren et.al conducted a case study in the continuous stirred-tank reactor (CSTR) and into the performance of the RTO/ML algorithms and pure ML algorithms. [13] .
For the first method, Ren et.al used an RTO model with an nonlinear equation and an model predictive control (MPC) layer to solve the rolling optimization problem [13].
They set up the optimization problem for the RTO layer to be an objective function which includes the economic objective (J), the system state(x) and parameters(p) such as feed composition [13].
This nonlinear equation will be repeatedly updated when a steady state occurs.
The MPC layer involves similar variables in its optimization problem.
However, the MPC layer involves simplifications which may lead to inaccuracies in process dynamics simulation [13].
Because of the RTO model being steady-state and the MPC being dynamic, this may lead to conflicts between the two layers [13].
This is addressed through a dynamic learning approach by making the MPC calculate the control input based on the current states (which include reactor temperature and reactant concentration) and the setpoints from the RTO model [13].
The Markov Decision Processes are used to conduct the learning in which the agent reviews the current state from and receives a subgoal(g) from the upper layer [13].
The MPC layer then generates an action to approach g [13].
The environment is transitioned to the next state and reward is provided based on the reward function involving the state space and action space [13].
The upper RTO level would choose actions based on longer term goals while the MPC would choose actions to fulfill short term goals to eventually fulfill long term goals [13].
The Hierarchical DDPG learning method would allow these connected actions to occur by updating the RTO layer policy based on external rewards and the lower MPC policy using internal rewards.
The external reward would be the economic objective while the internal reward would validate the state based on tolerance [13].
For the second method, Ren et.al used the same H-DDPG algorithm to train both the RTO and the MPC layers [13].
The actor network is parameterized by the network weights and takes in the state value and goal values, and the critic network was parametrized by the trainable weights in the RTO layer [13].
In the MPC layer/controller, the same definitions are completed but the inputs are the meta state and the subgoal and the output is the action, which is affected by the policy, which is affected by the state and the subgoal.
The RTO layer makes decisions and adds noise to allow the agent to explore.
The current state after the decision is passed to the MPC controller which will select the action, calculate the internal and external rewards, replay some experiences(b), use the RTO Actor network to calculate target values.
It then updates the critic network in the MPC layer through action-value loss minimization, policy gradient, and network parameters.
It then updates the critic network, policy gradients, and network parameters of the RTO layers.
The RTO’s current meta state is measured and sent to the MPC layer and the cycle continues until the subgoal is achieved.
These two methods were compared by conducting both methods in a fifty-minute process.
The results show that the MPC controller with the nonlinear equations do not follow the given setpoints when operating conditions change while the MPC controller with the ML algorithm as the foundation of the former RTO model is following the given setpoints to a higher degree.
This shows positive results in full machine learning algorithmic methods and suggests that the RTO model may be able to be trained using machine learning versus nonlinear equations.
This study is corroborated by a previous study conducted by Quah et.al who investigated the artificial neural network using a particle swarm optimization method (ANN-PSO)’s performance in chemical processes compared to a first principles modeling non-linear programming (FP-NLP) model [14].
The case study which is used is also a CSTR process like Ren et.al did.
The parameter of interest was the average profit made per minute.
A PPO model is also compared as well [14].
The results show that FP-NLP had the highest average profit of $139.28/min [14].
However, the PPO model and the ANN-PSO models achieve over 99% of the profit of FP-NLP and PPO model has the fastest computation time (0.002 s) [14].
Quah et.al and Ren et al. both show similar results.
While the machine learning algorithm may have similar or greater rewards and optimal optimization conditions, it has consistently been shown to have a faster training time and higher efficiency.
However, there are some challenges in implementing ML in the refinery to appropriately optimize feedstocks.
For example, there are numerous simplifications and uniformity in some variables, which may not exactly portray the process conditions.
For example, in Shi et.al, they made 4 assumptions: “1. Each reactor in the cascade is modelled as a plug-flow reactor (PFR) with respect to the radial direction.2. Radial dispersion effects are considered negligible along the reactor’s length.
All reactions are treated as homogeneous phase reactions occurring in the gas phase.
Uniform catalyst activity is assumed throughout each reactor.” [11].
The catalyst is not uniform during the process as said in this textbook chapter [15].
Another example is in Dong et.al, in which a lumped kinetic model is used as the base of a surrogate mathematical model which is run as described above [2].
An issue present in the surrogate model is that the errors and biases of the foundational model may seep into the surrogate, leading to inaccurate results and unexpected composition anomalies [16].
Also, many high-quality data points must also be present and the compilation of them in an appropriate manner for the machine learning agent to follow can also pose challenges as implementation of multiple subsets of data may be extremely laborious and difficult to retrieve as they may be high dimensional, biased, constrained by physical laws, or biased [17].
In addition, if online learning occurs, the exploration conducted by the agent must be limited by safety constraints, else ideal values may be negatively affecting the refinery as the “ideal” condition may be impossible to achieve and negatively affect operations.
Offline learning is also difficult as realistic simulators which are confined to appropriate operating conditions are rarely available.
There are concerns regarding the right type of model and ML algorithm chosen because if these are incorrect for the application, it may lead to suboptimal performance.
There are also security risks as databases could be breached with weak data plaguing the model’s performance.
Significant amounts of energy are also used to power these ML systems which can lead to more usage, which can impact the environment [18]
In addition, the interpretation of the results is extremely difficult as the agent may make actions which the humans running the ML model may be confused at.
Artificial intelligence may be used to interpret these results and provide insight into potential opportunities.
In conclusion, optimization algorithms such as NOMAD and TD3 can optimize feedstocks in terms of composition (aromatics) and H2 yield, benzene, toluene yield, and inlet temperature of reactors from the continuous catalytic reform process.
They may also be used in computational chemistry in the refinery through machine learning molecular algorithms in which composition of naphtha can be predicted.
However, there are multiple challenges and downsides to using machine learning in the typical refinery.
The labelling and quality of training and testing data for the machine learning algorithms are paramount for model quality and performance as those are at the pillar of model training.
If the data is not clustered appropriately, this can lead to model mismatches if a surrogate model is trained based on the primary model.
In addition to data quality, mathematical simplifications and uniformity simplifications of certain parts of the process may lead to the agent inappropriately changing settings within the refinery which were unforeseen as the composition may have changed to something the agent may not have experienced before.
While reinforcement learning and transfer learning combined as machine learning techniques show a correlational relationship towards more rewards and best-fit as shown through Dong et.al, there needs to be guardrails, appropriate safety measures and caution is needed during the testing of machine learning models within a refinery for reasons discussed previously.
There needs to be more research in how machine learning can be widely integrated into the real-time optimization of naphtha feedstock through the continuous catalytic reforming process and how model quality can be standardized.
Dr. Raj Shah is a Director at Koehler Instrument Company in Holtsville, New York, where he has served for over three decades, contributing to the advancement of petroleum, fuels, lubricants, and analytical instrumentation technologies worldwide.
Over the course of his distinguished career in the energy and chemical engineering industries, he has become widely recognized for both his technical leadership and sustained service to global professional societies.
Dr. Shah is an elected Fellow by his peers at ASTM International, the Institute of Chemical Engineers (IChemE), the Chartered Management Institute (CMI), the Society of Tribologists and Lubrication Engineers (STLE), the American Institute of Chemists (AIC), the National Lubricating Grease Institute (NLGI), the Institute of Measurement and Control (InstMC), the American Oil Chemists’ Society (AOCS), the Institute of Physics (IOP), The Energy Institute (EI), and The Royal Society of Chemistry (RSC).
These fellowships reflect his multidisciplinary impact across chemical engineering, tribology, measurement science, energy technology, and applied chemistry.
He is also the recipient of the prestigious ASTM Eagle Award and ASTM’s highest honor, the Award of Merit (Fellow), recognizing more than 30 years of leadership and contribution to Committee D02 on Petroleum Products, Liquid Fuels, and Lubricants.
He recently co-edited the bestseller, Fuels and Lubricants Handbook: Technology, Performance, Properties, and Testing, a major reference work for the industry. Dr. Shah has now authored and co-authored over 750 technical publications, conference papers, and industry articles, and continues to be an active contributor to the scientific and engineering literature. Further information regarding his work and recognitions can be found at https://shorturl.at/I7000.
Dr. Shah earned his doctorate in Chemical Engineering from The Pennsylvania State University and is a Fellow of The Chartered Management Institute, London.
He is a Chartered Scientist (CSci) with the Science Council, a Chartered Chemist (CChem) with the Royal Society of Chemistry, a Chartered Engineer (CEng) with the Engineering Council, UK, and a Chartered Petroleum Engineer (CPEng) with the Energy Institute.
He was recently granted the honorific distinction of “Eminent Engineer” by Tau Beta Pi, the oldest and largest engineering honor society in the United States, an honor reserved for engineers demonstrating exceptional professional achievement and character.
Actively engaged in academia and mentorship, Dr. Shah serves on the Advisory Boards of Farmingdale State College (Mechanical Technology and Engineering Management), Auburn University (Tribology and Lubrication Science), and the State University of New York at Stony Brook (Chemical Engineering and Materials Science & Engineering).
He is also an Adjunct Professor in the Department of Materials Science and Chemical Engineering at Stony Brook University.
Throughout his career, he has remained deeply committed to advancing engineering education, standards development, and technical excellence within the global energy community.
More information on Dr. Shah can be found at https://shorturl.at/yYl85.
Mr. Soorya Shanmugam is a first-year chemical engineering student and Morrill Scholar at Ohio State University. In parallel, he serves as a petroleum research intern at Koehler Instrument Company in New York, researching petroleum and fuel related topics.
[1] Anonymous, “U.S. Energy Information Administration - EIA - independent statistics and analysis,” Refining crude oil - the refining process - U.S. Energy Information Administration (EIA), https://www.eia.gov/energyexplained/oil-and-petroleum-products/refining-crude-oil-the-refining-process.php (accessed Jan. 17, 2026).
[2] M. J. Kaiser, “A review of refinery complexity applications,” Petroleum Science, vol. 14, no. 1, pp. 167–194, Jan. 2017, doi: https://doi.org/10.1007/s12182-016-0137-y.
[3] J. G. Speight, “Catalytic Reforming - an Overview | ScienceDirect Topics,” www.sciencedirect.com, 2020. Available: https://www.sciencedirect.com/topics/engineering/catalytic-reforming
[4] P. J. Giammatteo, “Increasing efficiency, safety, and reliability in refineries with real-time crude feedstock analysis - Feb 28 2024 - Paul J. Giammatteo, Ph.D. - Petro Industry News Articles - Petro Online,” Petro Industry News, https://www.petro-online.com/article/safety/15/4ir-solutions/increasing-efficiency-safety-and-reliability-in-refineries-with-real-time-crude-feedstock-analysis/3406 (accessed Jan. 17, 2026).
[5] J. Matias and J. Jäschke, “Online model maintenance in real-time optimization methods,” Computers & Chemical Engineering, vol. 145, no. 107141, p. 107141, Feb. 2021, doi: https://doi.org/10.1016/j.compchemeng.2020.107141
[6] X.-J. Dong, J.-N. Shen, Z.-F. Ma, and Y.-J. He, “Robust optimal operation of continuous catalytic reforming process under feedstock uncertainty,” International Journal of Hydrogen Energy, vol. 47, no. 84, pp. 35641–35654, Oct. 2022. doi:10.1016/j.ijhydene.2022.08.161
[7] Y. Guo, H. Wang, L. Li, X. Sun, and S. Li, “Real-time optimization of continuous catalytic reforming using transfer learning and reinforcement learning for feedstock variability,” Journal of the Taiwan Institute of Chemical Engineers, vol. 179, Feb. 2026. doi:10.1016/j.jtice.2025.106419
[8] P. Pasandide and M. Rahmani, “Simulation and optimization of continuous catalytic reforming: Reducing Energy Cost and coke formation,” International Journal of Hydrogen Energy, vol. 46, no. 58, pp. 30005–30018, Aug. 2021. doi:10.1016/j.ijhydene.2021.06.151
[9] X. C. Ma, C. He, Q. L. Chen, and B. J. Zhang, “Modeling and optimization for the continuous catalytic reforming process based on the hybrid surrogate optimization model,” Computers & Chemical Engineering, vol. 191, p. 108841, Dec. 2024. doi:10.1016/j.compchemeng.2024.108841
[10] C. Xie, R. Yao, L. Zhu, H. Gong, H. Li, and X. Chen, “Hybrid dynamic modeling of an industrial reactor network with first-principles and data-driven approaches,” Chemical Engineering Science, vol. 289, no. 119852, p. 119852, Feb. 2024, doi: https://doi.org/10.1016/j.ces.2024.119852. Available: https://www.sciencedirect.com/science/article/pii/S0009250924001520#s0010
[11] Y. Shi, W. Zhong, X. Peng, K. He, and D. Xue, “Molecular management in continuous catalytic reforming operations by enhanced aromatics production through transformer-driven entropy maximization reconstruction,” Chemical Engineering Science, vol. 309, p. 121359, May 2025. doi:10.1016/j.ces.2025.121359
[12] J. S. Angelo, I. A. Guedes, H. J. C. Barbosa, and L. E. Dardenne, “Multi-and many-objective optimization: present and future in de novo drug design,” Frontiers in Chemistry, vol. 11, no. 1288626, Dec. 2023, doi: https://doi.org/10.3389/fchem.2023.1288626
[13] R. Ren and S. Li, “Enhancing hierarchical learning of real-time optimization and model predictive control for operational performance,” Journal of Process Control, vol. 155, no. 103559, p. 103559, Oct. 2025, doi: https://doi.org/10.1016/j.jprocont.2025.103559.
Available: https://www.sciencedirect.com/science/article/pii/S0959152425001878. [Accessed: Jan. 26, 2026]
[14] T. Quah, D. Machalek, and K. M. Powell, “Comparing Reinforcement Learning Methods for Real-Time Optimization of a Chemical Process,” Processes, vol. 8, no. 11, p. 1497, Nov. 2020, doi: https://doi.org/10.3390/pr8111497
[15] H. S. Fogler, “Catalyst Decay,” in Elements of Chemical Reaction Engineering, London, England: Pearson Education, 2006
[16] H. Jones and S. Chen, “Surrogate models: A comfortable middle ground?” Society of Actuaries, https://www.soa.org/digital-publishing-platform/emerging-topics/surrogate-models/#:~:text=First%2C%20when%20building%20a%20surrogate,Implementation (accessed Jan. 17, 2026).
[17] P. Daoutidis et al., “Machine learning in process systems engineering: Challenges and opportunities,” Computers & Chemical Engineering, vol. 181, no. 108523, p. 108523, Feb. 2024, doi: https://doi.org/10.1016/j.compchemeng.2023.108523.
Available: https://www.sciencedirect.com/science/article/pii/S0098135423003939?dgcid=coauthor. [Accessed: Jan. 26, 2026]
[18] F. Cuello, “What are the Risks of Machine Learning? | TrendSpider Learning Center,” Trendspider.com, Aug. 05, 2024. Available: https://trendspider.com/learning-center/what-are-the-risks-of-machine-learning. [Accessed: Jan. 26, 2026]
PIN 27.2 Apr/May 2026
.jpg)
Product directory
