Editorial: Machine Learning Optimization of Chemical Processes

The integration of machine learning into chemical process optimization has undergone a remarkable evolution in recent years, moving rapidly from proof-of-concept demonstrations in controlled laboratory settings to deployment in real industrial environments. While early applications of data-driven methods in chemical engineering were largely confined to regression-based soft sensors and statistical process control, the contemporary landscape encompasses deep learning architectures, Bayesian inference, reinforcement learning, and physics-informed neural networks operating across the full complexity of industrial chemical systems. This transition has been accompanied by a parallel maturation in the methodological toolkit available to practitioners: reaction optimization is now understood not merely as a parameter-tuning exercise but as a structured decision problem that can be addressed through systematic experimental design, surrogate modeling, and closed-loop automation [1]. At the same time, the broader promise of predictive chemistry, that is, the use of machine learning to guide reaction deployment, reaction development, and even reaction discovery, has moved from aspiration to demonstrable practice across academic and industrial laboratories alike [2]. This breadth and maturity of application define the current Special Issue on Machine Learning Optimization of Chemical Processes, published in the journal Processes.

Machine learning in chemical engineering is no longer a peripheral curiosity. As industries face mounting pressures to improve efficiency, ensure process safety, and meet increasingly stringent sustainability targets, the ability to extract actionable intelligence from process data has become a central engineering imperative. The nine contributions assembled in this Special Issue, comprising seven research articles and two reviews, collectively reflect the expanding frontier of this field, spanning reaction engineering, semiconductor manufacturing, battery safety, aquaculture, minerals processing, and process control. Together, they demonstrate that machine learning is not a singular methodology but a versatile toolkit whose value is realized through thoughtful integration with domain knowledge.

The two review papers assembled in this issue provide essential conceptual anchoring. One (Shen, Luo, and Su, Contribution 1) offers a comprehensive treatment of Bayesian optimization (BO) for chemical synthesis in the era of artificial intelligence, situating Gaussian process surrogate models and acquisition functions within the practical constraints of chemical experimentation. Bayesian optimization is examined for its capacity to address the inherent data scarcity and experimental cost challenges in chemical synthesis, and this review situates it as perhaps the most principled approach available for sequential experimental design in chemical settings, where each experiment carries real material and time costs. This framing resonates with a broader trend in the field: the emergence of large language models purpose-built for chemistry, capable of jointly reasoning over reaction prediction, retrosynthesis, and property estimation within a unified generative framework, further compressing the iterative loop between hypothesis generation and experimental validation [3]. The second review (Alghamdi and Haraz, Contribution 2) turns to the aquaculture domain and examines the integration of artificial intelligence and Internet of Things (IoT) technologies in smart biofloc systems for sustainable aquaculture. The review presents a comparative analysis of AI models, including LSTM, Random Forest, and SVM, for water quality prediction and process management, addressing not only performance metrics but also model interpretability and failure mode analysis. This contribution broadens the scope of the issue and underscores the universality of machine learning optimization principles across process industries well beyond traditional chemical manufacturing.

A recurring and increasingly urgent theme running through this issue, and through the field at large, is the reliable quantification of predictive uncertainty. As machine learning models are deployed in settings where erroneous predictions carry direct safety or economic consequences, whether thermal runaway in batteries, inverse process design in semiconductor manufacturing, or reinforcement-learning-driven control of refinery units, the ability to distinguish confident from unreliable predictions becomes a prerequisite for responsible deployment rather than an optional refinement. Recent surveys of uncertainty quantification methods for engineering systems have cataloged the growing arsenal of approaches available to practitioners, ranging from Bayesian neural networks and Gaussian process regression to ensemble-based and conformal prediction techniques, each offering different trade-offs between computational cost, calibration quality, and interpretability [4]. The contributions in this issue that adopt Bayesian frameworks, whether in the reinforcement learning agent for catalytic cracking control or in the surrogate modeling of semiconductor device behavior, can be read as concrete instantiations of this broader methodological imperative.

Closely related to the question of uncertainty is that of interpretability. As machine learning models are entrusted with process-critical decisions, domain experts increasingly demand not merely accurate predictions but explanations that align with physical and chemical intuition. The growing literature on explainable artificial intelligence in process engineering has begun to formalize this requirement, cataloging techniques such as SHAP, LIME, and attention-based visualization as tools for bridging the gap between black-box performance and engineering trust [5]. Several contributions in this issue engage directly with this imperative: the aquaculture review explicitly incorporates SHAP and LIME analyses into its comparative evaluation of predictive models, while the retrosynthesis model for radical reactions employs attention weight analysis to confirm that the network has learned chemically meaningful reaction patterns, including cascade cyclizations and photocatalytic steps, rather than merely fitting statistical regularities in the training data.

The research articles in this issue address a diverse set of optimization and modeling challenges, each illustrating a distinct facet of the machine learning toolkit applied to chemical processes. One contribution (Xu, Dai, Zang, and Zhu, Contribution 3) reports the development of a continuous-flow microreactor system for the enzymatic synthesis of polydatin, in which a glycosyltransferase-sucrose synthase cascade enables in situ UDP-Glc regeneration to reduce process costs. A kinetic study elucidates the apparent reaction orders and a sequential substrate-binding mechanism, demonstrating how precise residence-time control and enhanced mixing in microflow translate directly into systematic process optimization. This work sits within a broader movement toward inline-monitored continuous flow chemistry, in which real-time spectroscopic feedback, such as inline FTIR analysis, is used to close the loop between reaction monitoring and optimization in heterogeneous catalytic transformations, enabling autonomous adjustment of operating conditions without interrupting production [6]. Another contribution (Zhao, Xing, Jiang, Shu, and Sun, Contribution 4) investigates thermal runaway behavior in 18,650 lithium-ion batteries across five states of charge, combining infrared thermography, real-time temperature monitoring, mass loss analysis, and gas composition detection with quantitative explosion risk modeling using flammability limit analysis. The multi-parameter, data-rich experimental approach and the resulting mechanistic insights into state-of-charge-dependent thermal runaway dynamics are of direct relevance to battery safety engineering and provide a well-characterized experimental foundation for future machine learning-based predictive modeling of thermal runaway risk.

Another contribution (Xu, Dong, Du, Liu, Peng, and Yu, Contribution 5) introduces the first deep learning-based retrosynthesis model specifically designed for radical reactions. Built upon the Chemformer architecture, pretrained on ZINC-15 and USPTO datasets, and fine-tuned on a newly curated database of 21,600 radical reactions, the model achieves a Top-1 retrosynthesis accuracy of 69.3%, surpassing established models by substantial margins. This work represents a significant advance in computer-aided synthesis planning for a reaction class that has been historically underserved by AI-driven tools, and it exemplifies a broader effort within the field to bridge chemical domain knowledge with machine learning architectures so that predictive performance on organic synthesis tasks is grounded in mechanistically meaningful representations rather than purely statistical pattern matching [7]. Yet another contribution (Geng, Guo, Sun, Gao, and Ni, Contribution 6) addresses process-device co-optimization in trench-gate MOSFET manufacturing by developing an LSTM-based proxy model for TCAD simulation, coupled with Bayesian optimization for inverse process parameter design. The proxy model achieves prediction deviations below 3.5% relative to full physical simulations, while the BO-driven inverse optimization navigates trade-offs between competing electrical performance objectives, with recipe predictions deviating by no more than 8.3% from experimental data. Such surrogate-model-driven approaches to process design echo the growing role of digital twins in chemical and semiconductor engineering, where physics-informed surrogate models are increasingly used to enable rapid, low-cost exploration of process parameter spaces that would otherwise require prohibitively expensive full-scale simulation or experimentation [8].

A further contribution (Qin, Ye, Zheng, and Jin, Contribution 7) presents a Bayesian deep reinforcement learning framework for the optimal operation of a fluid catalytic cracking unit. By embedding Bayesian neural networks within the reinforcement learning agent, replacing conventional deterministic network weights, and integrating a primal-dual constraint-handling mechanism, the proposed method achieves more stable control performance and higher economic profit under process parameter fluctuations and external disturbances compared to standard RL approaches. The application to fluid catalytic cracking, a cornerstone unit operation in petroleum refining, underscores the industrial relevance and economic stakes of intelligent process optimization, and it anticipates a future in which reinforcement-learning-based control loops are embedded within increasingly autonomous, self-optimizing plants, a vision conceptually related to the emerging paradigm of self-driving laboratories, in which closed-loop experimentation, robotic execution, and machine-learning-guided decision making are integrated to accelerate materials and process discovery with minimal human intervention [9]. Finally, one contribution (Hu, Chen, Cen, Yin, and Deng, Contribution 8) proposes a novel video-based method for superheat degree identification in aluminum electrolytic cells, combining the VideoMamba state space model with a channel attention mechanism and a learnable nonlinear Fourier transform block. The approach achieves 85.7% identification accuracy on fire hole video data while maintaining lower computational complexity than transformer-based alternatives, illustrating how specialized neural architectures, designed with domain-specific physical priors in mind, can outperform generic deep learning baselines in demanding industrial monitoring tasks.

Rounding out the collection, one contribution (Arce Munoz and Hedengren, Contribution 9) addresses the practical challenge of controlling industrial thickeners, unit operations widely used in mineral processing and wastewater treatment whose underlying dynamics are notoriously difficult to capture with first-principles models due to nonlinear settling behavior and time-varying feed characteristics. The authors propose a transfer learning strategy in which a control-oriented neural network model, pretrained on data from a well-characterized source thickener, is efficiently adapted to a new target unit using only a limited amount of site-specific data, substantially reducing the data collection burden typically associated with deploying data-driven control on a new piece of equipment. This work speaks directly to one of the most persistent obstacles in the industrial adoption of machine learning-based control: the cost and time required to accumulate sufficient plant-specific data before a model can be trusted for closed-loop operation. By demonstrating that knowledge learned on one thickener can be meaningfully transferred to another with different geometry and feed conditions, the contribution offers a template for scaling data-driven control strategies across fleets of nominally similar industrial units, a problem of considerable practical significance well beyond thickener control itself, and one that resonates with the broader push toward reusable, generalizable machine learning models in process engineering.

Surveying this collection as a whole, several themes emerge that define the current state and future trajectory of machine learning in chemical process optimization. First, the integration of uncertainty quantification, whether through Bayesian neural networks, Gaussian processes, or ensemble methods, is increasingly recognized as essential for deploying machine learning in high-stakes process environments where overconfident predictions carry real safety and economic consequences. Second, the challenge of data scarcity remains central: transfer learning, multi-task learning, physics-informed modeling, and active learning strategies, of which Bayesian optimization is a prime example, are all responses to the fundamental reality that labeled process data is expensive to generate, a challenge directly confronted by the thickener control contribution described above. Third, interpretability is no longer an afterthought; contributions in this issue explicitly address how models can be understood and trusted by domain experts, a prerequisite for industrial adoption. Fourth, the scope of chemical processes continues to broaden, from reaction engineering and pharmaceutical synthesis to semiconductor fabrication, battery systems, aquaculture, and minerals processing, reflecting the growing recognition that optimization principles developed in chemical engineering have value across a wide range of process industries.

Looking beyond the immediate contributions of this issue, the field appears to be converging toward an increasingly unified conception of the chemical reaction space as a mathematical object that can be systematically represented, searched, and optimized using machine learning [10]. Within this conception, the individual advances reported here, including improved surrogate models, better-calibrated uncertainty estimates, more interpretable predictions, more transferable control strategies, and more autonomous experimental platforms, are not isolated technical achievements but interlocking pieces of a larger infrastructure for data-driven chemical engineering. As this infrastructure matures, the boundary between computational prediction and physical experimentation is likely to become progressively more porous, with machine learning models serving simultaneously as hypothesis generators, experimental planners, and real-time process controllers.

This Special Issue does not attempt to provide a definitive account of machine learning optimization in chemical processes, as the field evolves too rapidly for any single volume to do so. What it does offer is a rigorous, multifaceted snapshot of a discipline that is maturing with remarkable speed, moving steadily from algorithmic novelty toward engineering practice. The guest editor thanks all contributing authors for the quality and diversity of their submissions, the reviewers for their diligent and constructive assessments, and the editorial team of Processes for their support throughout the publication process. It is hoped that the contributions gathered here will serve as both a useful reference and a stimulus for further research at the productive intersection of machine learning and chemical process engineering.