The accurate representation of complex reaction kinetics and fundamental industrial phenomena constitutes a critical and persistent challenge in petroleum refining process modeling. The complexity of species and reactions in refining models limits comprehensive understanding of kinetic characteristics, thereby posing significant challenges for model validation. To address this, this study proposes a post-hoc interpretability analysis approach for existing refining kinetic models, enabling systematic investigation of reaction contributions, species interactions, and core reaction networks. First, individual reaction molar flow analysis characterizes how a single reaction influences specific species throughout the entire process. Subsequently, a Directed Species Interaction Analysis (DSIA) algorithm traces influence propagation through routes of varying lengths, generating quantitative metrics for directed interaction strength between any species pair. For key industrial species, these metrics enable extraction of core reaction networks that reveal the critical routes governing their formation and consumption. When applied to naphtha catalytic reforming models, the approach demonstrates a clear transformation pathway from paraffins to naphthenes and progressively to aromatics, elucidating the reaction routes underlying the formation and accumulation of major aromatic products. The resulting visualizations identify distinct kinetic behaviors across alternative kinetic models, offering a transparent and mechanistic basis for model comparison and validation. This approach enhances the understanding of reaction kinetics within refining process models, ultimately improving the fidelity of process simulations for industrial applications.
High-throughput analysis of microfluidic droplets and bubbles is essential for chemical engineering but remains challenging due to the inherent loss of high-frequency details in standard deep learning models. This study proposes a novel Hand-crafted Feature Fusion framework that explicitly integrates physical priors, specifically Local Binary Patterns and Discrete Wavelet Transform, into a two-stage instance segmentation network. We design an adaptive attention-based fusion module embedded within both the Feature Pyramid Network and Region Proposal Network to synergize explicit texture cues with implicit semantic features. Validated on a large-scale dataset comprising over 64000 instances, our method achieves a test mAP of 0.808, significantly outperforming state-of-the-art architectures. Crucially, the framework effectively resolves the detection bottleneck for minute targets and elevates the small-object accuracy to 0.764, representing an improvement of nearly 20% over the baseline. This work demonstrates that incorporating physical priors offers a superior strategy for precise scientific image analysis compared to generic data-driven models.
Simulation and modeling technologies play an increasingly critical role in enhancing feedstock utilization, improving product quality, and enhancing enterprise competitiveness in the refining industry. Molecular-level kinetic models for refining processes typically involve highly detailed and complex data, significantly increasing computational costs. Additionally, the data granularity within these models fails to align with the practical requirements of industrial production. To address these, this study proposes the Reaction Network Coarse-Graining Integration (RNCGI) algorithm to achieve flexible adjustments to model complexity and data granularity. Species with similar molecular structures and kinetic behaviors within the reaction network are identified and integrated. Subsequently, a kinetic parameter mapping approach is developed to preserve the critical reaction behaviors. Finally, the reaction coarse-graining process significantly reduces the number of reactions within the network. The RNCGI algorithm was validated using the naphtha catalytic reforming process. The original complete network with 206 species nodes was integrated into two versions with different data granularities: a medium-scale network with 42 nodes and a small-scale network with 5 nodes. Both coarse-grained networks calculated the product composition accurately and provided clearer insights into core species transformation mechanisms compared to the original network. Additionally, the RNCGI algorithm ensures the completeness of reaction routes. By flexibly handling detailed data in molecular-level models, this work preserves the necessary complexity of models tailored to industrial needs, supporting the development of refinery-wide production planning and scheduling strategies.
The petrochemical industry faces mounting challenges from environmental constraints, resource scarcity, market volatility, and rising production costs, driving the urgent need for precision-oriented process optimization. The inherent complexity of petrochemical reaction networks (PCRNs) hinders their practical deployment, bringing practical bottlenecks including excessive problem scale, heavy computation burden and poor mechanistic interpretability. To address this, we propose a Knowledge-based Overlapping Reaction Community Detection (KORCD) framework that integrates domain knowledge, graph theory, and artificial intelligence for systematic analysis and simplification of PCRNs. KORCD employs a quasi-crystallization algorithm to detect overlapping reaction communities centered on key products, preserving intact reaction pathways while revealing critical intermediates and cross-community interactions. Subsequent network flow analysis further distills each community into a skeletal mechanism by retaining high-importance reactions. The method is demonstrated on a steam cracking process using an industrial naphtha feedstock. Community detection identifies seven chemically meaningful modules governing distinct product formation routes. Simulation validation shows that KORCD-based simplification achieves up to 60% reduction in reaction count while maintaining mean relative errors below 3% for major products. Moreover, computational time decreases sharply at low-to-moderate pruning ratios, offering an optimal trade-off between accuracy and efficiency at 40–60% pruning. This work establishes KORCD as a robust, interpretable, and industrially viable approach for knowledge extraction and model reduction. From a practical engineering perspective, the chemically constrained KORCD method delivers skeletal mechanisms to support digital-twin deployment, low-cost real-time simulation, closed-loop process optimization and molecular-level intelligent control for petrochemical plants.
The scheduling of multi-line polyolefin production is a complex decision-making process characterized by sequence-dependent changeovers, strict physicochemical constraints, and dynamic market environments. Traditional optimization methods often suffer from high computational costs and a lack of flexibility in online adjustments. To address these challenges, this paper proposes a Deep Reinforcement Learning (DRL) framework for dynamic scheduling tasks. We first construct a high-fidelity simulation environment that meticulously models realistic industrial constraints, including transition materials, shutdowns, and inventory limits. A Soft Actor-Critic (SAC) agent with a tuple-based action space is employed to mitigate the combinatorial explosion associated with multi-line decisions. Furthermore, a dynamic action masking mechanism embedded with domain knowledge is introduced to strictly enforce hard constraints and significantly improve sample efficiency. Case studies based on real-world industrial data demonstrate that the proposed method can autonomously generate valid schedules that satisfy complex production requirements. Comparative experiments further reveal that the action masking mechanism accelerates training convergence, and the DRL agent exhibits superior adaptability to dynamic price fluctuations.
Crude oil scheduling is a critical but highly complex sequential decision-making problem in refinery operations. Traditional mathematical programming methods suffer from exponential computational complexity with increasing scale, while traditional reinforcement learning approaches struggle to guarantee the satisfaction of numerous process and product quality constraints. This highlights a critical issue in the current field of scheduling: the lack of an optimization methodology that can simultaneously achieve high computational efficiency and robust constraint satisfaction. To bridge this gap, we propose a novel hybrid framework based on constraint stratification. The framework embeds critical hard constraints, such as entity connectivity limit, directly into the scheduling environment for intrinsic satisfaction through action masking and shaping. Concurrently, it employs safe reinforcement learning algorithms to manage soft constraints, such as tank inventory levels and product quality specifications, by optimizing a primary objective while keeping constraint violations below a predefined threshold. Through comparative experiments on scheduling cases, the Constrained Policy Optimization algorithm was identified as the most effective safe reinforcement learning method. The results demonstrate that our proposed framework significantly outperforms traditional methods. It achieves the high computational efficiency and scalability of reinforcement learning while providing a much stronger safety guarantee than penalty-based approaches, offering a robust and practical solution for complex industrial scheduling problems.
The coalescence behavior of microdroplets affects the flow stability, transport and reaction performances, as well as phase-separation efficiency of the microdroplet-based reactors and separators. However, the coalescence dynamics of buoyancy-driven rising monodisperse microdroplets with rear-end collision in unconfined space remain insufficiently understood, despite its significance for designing microchemical systems based on microdroplet swarms. In this work, a string of monodisperse microdroplets generated using the micro-capillary jetting technique is taken as the research object. A self-developed image recognition system is employed to obtain high-resolution dynamic parameters. The results reveal that the film-thinning rate increases approximately linearly with terminal velocity. The influence of physical properties is manifested in their effects on the velocity distribution within the liquid film. Droplet size will increase the film drainage time. A critical capillary number is proposed to characterize the coalescence state of rising monodisperse microdroplets with rear-end collision, providing a basis for optimizing microchemical systems by either promoting or suppressing coalescence.
Improving the gas–liquid mass transfer rate in microdevices is essential for enhancing chemical reaction performance, but it has traditionally required high energy input or complex device fabrication. This study reports superior gas–liquid mass transfer performance in a newly designed T‐junction microchannel with a simple structure. Compared with the mass transfer contribution of approximately 30% in a conventional T‐junction microchannel, the contribution of the bubble generation stage in the modified device ranges from 50%–80%. The parameters of bubble generation frequency and liquid slug length are studied to identify the mechanism underlying the enhanced performance. Importantly, through a self‐developed image recognition system with high temporal and spatial resolution, this study reveals that the liquid‐side mass transfer coefficient not only depends on operation parameters but also relies on bubble residence time. Finally, considering channel length and mass transfer time, a new semi‐empirical model is developed.
In the context of low-carbon strategies, China has joined the global leadership series of prefabricated underground structures (PUS). However, PUS faces significant challenges in construction, maintenance, and restoration subjected to multi-hazard impacts, particularly at the assembled joints. Addressing this research gap, this study focuses on the PUS life-cycle resilience based on joint restorability. The full-scale experimental studies were conducted on typical assembled joint of PUS and its restored specimen. Integrating experimental results with intelligent monitoring techniques, a resilience assessment framework for the PUS was proposed. The effectiveness of this framework was validated through a case study. Key findings include: (1) Even with minor restoration level undetected rotations, the assembled joint can rapidly reach major restoration thresholds after cumulative operational impacts. (2) Intelligent monitoring techniques enable rapid and accurate responses to restoration needs. It reduces recovery time and cost by more than 10 times, with a 52 % resilience enhancement compared to traditional techniques. (3) The proposed joint restoration technology enhances the yield and ultimate load-bearing capacities by 11.8 % and 7.9 %, respectively, primarily owing to the enhancement of interface bond strength. (4) Restored joints exhibit higher resilience under subsequent daily operations and extreme conditions. Compared to original assembled joints, restored joints show 29 % improvements in rotation tolerance at minor restoration thresholds. Along with 22 % resilience enhancement, the restored joint significantly reduces more than 90 % restoration costs. This study not only provides key restoration technologies and intelligent resilience management methods for the life-cycle resilience of PUS, but also has significant practical application value and resilience enhancement implications.
The increasing urban building density has driven the utilisation of underground space for constructing urban municipal infrastructure and transportation systems, alleviating land scarcity. However, the construction of underground projects within densely populated cities poses challenges such as traffic congestion, secondary excavations, and settlement of adjacent building foundations. Accordingly, the simultaneous construction of subways and utility tunnels (SCSUT) has emerged as a practical solution for developing underground spaces in high-density areas. By integrating the constructions of urban rail transit and municipal infrastructure from a long-term perspective, SCSUT could reduce road occupancy and minimize impacts on existing structures. Despite its advantages, SCSUT implementation remains challenging, with few practical cases available. To gain insight into the SCSUT, this study first examines its necessity and global implementation status. Meanwhile, this research identifies the existing implementation strategies of the SCSUT in terms of construction management, pipeline layout, and typical construction schemes. Furthermore, an in-depth analysis of the largest existing SCSUT project is conducted. Findings indicate that the SCSUT, currently in its early development stage, faces several critical challenges, including the lack of technical standards, complex approval processes and management models, inadequate geological surveys and preliminary planning, and frequent design changes. To realize the significant potential benefits of the SCSUT, future efforts should focus on developing technical standards based on relevant engineering cases, enhancing early-stage planning, and leveraging digital technologies to support construction. This study is anticipated to be a valuable reference for policy improvement and other SCSUT projects similar to the case study.
This study tackles the challenge of accurate yield prediction in fluid catalytic cracking (FCC) units by comparing conventional supervised regression with time series forecasting methods using industrial data collected from the distributed control system (DCS) of an FCC plant. We introduce a shifted forecast paradigm that preserves temporal relationships between predictors and targets. Our preprocessing pipeline, which employs trimmed mean smoothing, addresses common industrial data challenges. Results demonstrate that the forecasting approach significantly outperforms supervised regression, achieving a mean absolute percentage error (MAPE) of 1.56% for 3-hour shifted predictions compared to 6.20% for supervised regression. The model maintains robust performance even with extended shifts during predictions, showing an MAPE of 3.55% for 14-day forecasts. This research provides valuable insights for implementing predictive analytics in industrial FCC operations, demonstrating the superiority of forecasting methods over traditional supervised regression approaches for process yield prediction.
In the context of carbon neutrality and carbon peaking, molecular management has become a focus of the petrochemical industry. The key to achieving molecular management is molecular reconstruction, which relies on rapid and accurate calculation of oil properties. Focusing on naphtha, we proposed a novel property prediction model construction procedure (MDs-NP) employing molecular dynamics simulations for property collections and gamma distribution from real analytical data for calculating mole fractions of simulation mixtures. We calculated 348 sets of mixture properties data in the range of 273 K-300 K by molecular dynamics simulations. Molecular feature extraction was based on molecular descriptors. In addition to descriptors based on open-source toolkits (RDKit and Mordred), we designed 12 naphtha knowledge (NK) descriptors with a focus on naphtha. Three machine learning algorithms (support vector regression, extreme gradient boosting and artificial neural network) were applied and compared to establish models for the prediction of the density and viscosity of naphtha. Mordred and NK descriptors + support vector regression algorithm achieved the best performance for density. The selected RDKFp and NK descriptors + artificial neural network algorithm achieved the best performance for viscosity. Using ablation studies, T, P_w and CC(C)C are three effective descriptors in NK that can improve the performance of the property prediction models. MDs-NP has the potential to be extended to more properties as well as more-complex petroleum systems. The models from MDs-NP can be used for rapid molecular reconstruction to facilitate construction of data-driven models and intelligent transformation of petrochemical processes.
Ethylene is one of the most important chemicals, and scheduling optimization is crucial for the profitability of ethylene cracking furnace systems. With the diversification of feedstocks and the high variability in prices, supply chain fluctuations pose significant challenges to the scheduling decisions. Dynamically responding to these fluctuations has become crucial. Traditional mixed integer nonlinear programming (MINLP) models lack the capability of supply chain response, while receding horizon optimization (RHO) models require parameter prediction and repeated optimization solving. To address this challenge, we propose a deep reinforcement learning-based framework that includes an ethylene dynamic scheduling environment and a decision agent based on deep Q-network. Across three test cases, compared to the MINLP and RHO models, this framework significantly minimizes losses caused by supply chain fluctuations, thereby increasing daily average net profits by 9%-27%, demonstrating its significant potential for application in responsive scheduling in the presence of supply chain fluctuations.
Machine learning-assisted retrosynthesis planning aims to utilize machine learning (ML) algorithms to find synthetic pathways for target compounds. In recent years, with the development of artificial intelligence (AI), especially ML, researchers’ interest in ML-assisted retrosynthesis planning has rapidly increased, bringing development and opportunities to the field. In this review, we aim to provide a comprehensive understanding of ML-assisted retrosynthesis planning. We first discuss the formal definition and the objective of retrosynthesis planning, and organize a modular framework which includes four modules: data preparation, data preprocessing, pathway generation and evaluation, and pathway verification. Then, we sequentially review the current status of the first three modules (except pathway verification) in the ML-assisted retrosynthesis planning framework, including ideas, methods, and latest progress. Following that, we specifically discuss large language models in retrosynthesis planning. Finally, we summarize the extant challenges that are faced by current ML-assisted retrosynthesis planning research and offer a perspective on future research directions and development.
Optimal control of the fed-batch biopharmaceutical process remains an open research and industrial challenge. The fed-batch process is characterized by non-steady operation, partially observable states, and batch uncertainties. Recently, deep reinforcement learning (DRL) has emerged as a powerful tool for chemical process systems. However, directly applying DRL to the simulation-based biopharmaceutical process has been proven a failure due to the mismatch between reinforcement learning algorithms and simulation-based environments. To fill this research gap, we proposed a novel DRL-based control framework that innovatively incorporates human knowledge. A case study of open-source simulator (IndPenSim), is used to verify the effectiveness of the proposed framework. The DRL controller demonstrates advantages in batch yields, computational loads, and measurement requirements compared to the existing control methods. An improvement of 14 % average penicillin yield is achieved with only 0.01 similar to 0.03 s online computation time per step, showing its potential in the control applications for fed-batch biopharmaceutical process.
Natural gas hydrates represents a huge source of energy. At the same time substantial leakages of natural gas from hydrates contributes significantly to climate changes. One of the most important reasons for these natural gas fluxes is leakage of seawater in to the hydrates from seafloor, through fracture systems. Hydrate dissociates if surrounding seawater is less than hydrate stability limit. Another interesting aspect of natural gas hydrates is the potential for safe CO2 storage. These different aspects of hydrates in natural sediments put demands on thermodynamic models. In addition to accurate description of pressure temperature hydrate stability there also a need to describe hydrate dissociation in concentration gradients towards surrounding water or surrounding gas as two examples. In this work we present new experimental data and an extensive thermodynamic model for hydrate. In contrast to conventional thermodynamic models for hydrate the model is consistent since all thermodynamic properties are derived from the Gibbs free energy. In this work we examine mixtures of CH4, C2H6, N2, CO2 from the China Sea and some synthetic mixtures, using this model. Maximum CO2 content in these mixtures are 60 mol% and the rest is dominated by CH4. Agreement between experimental data and model calculations are generally good and average deviations are below 5.5% for all the systems and conditions examined. Another aspect of the model is the ability for incorporation of effects of mineral surfaces. Specifically it is illustrated that adsorption of water on rust dominates liquid water drop out from gas as compared to water dew-point. Production of natural gas with such high CO2 content requires a strategy for CO2 separation and storage. It is proposed that the CH4 is separated from the C2H6, CO2 and N2 and cracked to H2 and CO2 using steam. Thermodynamic analysis indicates a significant potential for safe CO2 storage in natural gas hydrate and H2 as the only export product.
Fluid Catalytic Cracking (FCC) is one of the most important conversion processes in oil refineries, widely used to convert high-boiling, high-molecular-weight hydrocarbon components from crude oil into more valuable products like gasoline and diesel. Advanced simulation and optimization technologies are critical for improving the operational efficiency and economic performance of the FCC process. First-principles-based simulators rely on parameter estimation and are computationally intensive, making them unsuitable for online optimization. In recent years, with the development of deep learning, data-driven models have made significant progress in FCC modeling. However, due to their black-box nature and difficulty with extrapolation, they are rarely used for optimization. To bridge this gap, we propose an integrated framework that combines hybrid modeling and surrogate model-based optimization. This approach combines plant and simulation data to train a multi-task learning prediction model, which then serves as a surrogate for operational optimization. Validated on a large-scale FCC unit in southern China, the model predicts product yields with an error margin of under 4.84% for all products. Following optimization, yields of LNG, gasoline, and diesel rose by an average of 0.10 wt%, 1.58 wt%, and 1.05 wt%, respectively, resulting in a 3.67% increase in product revenues. This highlights the substantial potential of this framework for industrial applications.