Digital Chemical Engineering最新文献

Autoencoder-based reduced-order machine learning models have been developed for modeling and predictive control of nonlinear chemical processes with high dimensionality such as discretization of reaction–diffusion processes. However, in the presence of data noise, autoencoders may over-fit the training data and subsequently learn an inaccurate low-dimensional representation of the process variables. This leads to an inaccurate prediction model when the models are integrated with model predictive control (MPC). To address this issue, this work develops a novel machine-learning-based reduced-order modeling method by integrating SpectralDense layers into autoencoders and incorporating them with recurrent neural networks. We demonstrate that the new architecture of autoencoders using SpectralDense layers is more robust against over-fitting than conventional autoencoders in the presence of data noise, which improves the prediction accuracy in MPC. A diffusion–reaction process simulation example is used to demonstrate that the robust autoencoders outperform those using conventional layers for reduced-order modeling in predictive control.

Cold start is a critical operating scenario for the proton exchange membrane fuel cell (PEMFC), particularly in the field of transportation. Under sub-freezing temperatures, the water inside the cell will freeze and obstruct gas flow paths as well as cover catalyst reaction sites, resulting in a failed startup. This study proposes an optimization method for the -30°C cold start of PEMFC based on a data-driven surrogate model to improve cold start performance and reduce irreversible damage to the cell. A validated PEMFC cold start mechanism model is utilized as the basis for developing an extreme learning machine (ELM) based data-driven surrogate model, which is trained using data collected from the mechanism model and has higher computational efficiency compared with the original model. In addition, the NSGA-II multi-objective optimization algorithm is employed to optimize the current loading strategies and operating parameters using the surrogate model as fitness function. The objectives are to enhance the minimum voltage and reduce startup duration time. Moreover, experimental validation confirms the effectiveness of the proposed method. The test results demonstrate that a cold start from -30°C is achieved within 97 s, with the minimum voltage reaching 0.44 V. Notably, there is a reduction in startup time by 26 s and an increase in the minimum voltage by 0.06 V compared to the base case. This study establishes a foundation for researchers to adjust operating settings during cold start based on diverse applications and requirements.

Cryogenic processes with mixed refrigerants are prevalent in energy-intensive chemical industries, enhancing energy efficiency while reducing costs and unit size. However, the curse of dimensionality and process design constraints pose significant hurdles for effective screening and optimization. To tackle this, we developed a neural network model for natural gas liquefaction prediction. Trained on an extensive Aspen HYSYS database, our ML model accurately simulates LNG processes, with an impressive $R^2$ test value of 99.63, operating almost ten million times faster than HYSYS. It effectively addresses vital process design constraints, including liquid slugging and temperature cross, crucial for optimization. By integrating the ML model with genetic and Nelder–Mead algorithms, we achieve an 8.9% reduction in total exergy, outperforming Aspen HYSYS within the same time frame. Our study underscores ML’s significance in modeling energy-intensive chemical processes, providing insights into the exergy profile and enabling feature importance analysis.

The measurement of batch quality indicators in real time operation is plagued with many challenges, hence soft sensing has become a promising solution within industrial research. However, small data has traditionally been a severe problem, hindering the ability to create accurate, reliable soft sensors, especially within industrial research and development for new product formulations. Nevertheless, it is often the case that modelling knowledge is available for a related system. In order to exploit this, we have developed a generalisable transfer learning methodology which takes advantage of previous modelling efforts to accelerate and improve the construction of models for new systems. Specifically, we adapted a recently developed advanced data-driven soft sensing methodology made for an existing process formulation and integrated a feature-based transfer learning approach to facilitate the modelling of two new industrial process systems, each of which containing notable differences to the original. The performance of the transfer soft sensors was tested rigorously and compared to a benchmark approach under different data availability conditions. It was shown that, the proposed transfer mechanism yielded high accuracy, and is robust to small data scenarios, indicating its potential for use in soft sensing of novel systems.

Chromatography is one of the most valuable techniques chemists possess at their disposal, conducive to everything from developing vaccines, food, beverage, and drug testing to catching criminals. The diverse applications allow it to be used for analytical and preparative purposes. On the other hand, droplet microfluidics has significantly evolved from simple droplet generators to complex and integrated tasks through specially designed networks. Microfluidics finds itself at the center of various Lab-on-Chip studies, enabling single-cell analysis, biochemical synthesis, etc. We demonstrate a microfluidic chromatograph machine that can produce an ordered droplet arrangement for a large number of drops. The droplets are sent into the device using a novel methodology where the conventional droplet train is made into smaller batches. The study describes the use of droplet batch methodology and compares it with the traditional droplet train approach. Using this platform, different droplet sequences are sent through the chromatograph, which preferably allows some droplets to exit first while others take a longer time to flow across the chromatograph based on the droplet properties and device design. The droplet sequences contain various drops; however, the type of drops in these sequences is limited to 2. The chromatograph can handle any number of drops in a single machine is enough for handling diverse droplet sequences. The stability of the microfluidic chromatography is also studied by varying the droplet properties and the droplet batch size.

Area-selective atomic layer deposition (AS-ALD) is a beneficial procedure that facilitates self-alignment for transistor stacking by concentrating oxide growth on targeted areas of a substrate. However, AS-ALD is difficult to incorporate into semiconductor manufacturing industries due to difficulties such as minimal process data and a lack of insight into reactor design. To enable the industrial scale-up of AS-ALD, in silico modeling is necessary to characterize the process. Thus, this work proposes a multiscale computational fluid dynamics modeling framework that simultaneously describes the surface chemistry and ambient fluid behavior for an Al₂O₃/SiO₂ substrate. The multiscale model first involves ab initio molecular dynamics simulations to optimize molecular structures involved in the AS-ALD reactions. Next, a kinetic Monte Carlo simulation is performed to describe the stochastic surface chemistry behavior to determine the surface coverage, and deposition and byproduct rates. Lastly, computational fluid dynamics is performed to study the spatiotemporal behavior of the flow. The surface and flow field simulations are carried out in an integrated fashion. Various AS-ALD discrete feed reactor configurations with differing injection plate geometries were developed to investigate their impact on the processing time to achieve full surface coverage and film uniformity. Results indicate that the multi-inlet reactor model achieves minimal processing time while producing a high-quality film with the AS-ALD process.

Digital Twins (DTs) are high-fidelity virtual models that behave-like, look-like and connect-to a physical system. In this work, the physical systems are operations and processes from energy-intensive industrial plants and their local communities. The creation of DTs demands expertise not just in engineering, but also in computer science, data science, and artificial intelligence. Here, we introduce the Adaptive Digital Twins (ADT) concept, anchored in five attributes inspired by the self-adaptive systems field from software engineering. These attributes are self-learning, self-optimizing, self-evolving, self-monitoring, and self-protection. This new approach merges cutting-edge computing with pragmatic engineering needs. ADTs can enhance decision-making in both the design phase and real-time operation of industrial facilities and allow for versatile 'what-if' scenario simulations. Seven applications within the energy-intensive industries are described where ADTs could be transformative.

Steam methane reforming (SMR) is one of the most widely used hydrogen (H₂) production processes. In addition to its extensive utilization in industrial sectors, hydrogen is expanding it share as a clean energy carrier, and more sustainable and efficient H₂ production methods are continuously being explored and developed. One method replaces conventional fossil fuel-based heating with electrical heating through the flow of electrons across the reformer. At UCLA, an experimental setup was built of an electrically heated steam methane reforming process. This paper describes the system components, explains the digitalization of the experimental setup and introduces methods for building a first-principles-based dynamic process model using parameters estimated via data-driven methods from process experimental data. The modeling approach uses a lumped parameter approximation and employs algebraic equations to solve for gas-phase variables. The reaction parameters are calculated from steady-state experimental data, and the temperature change is modeled with respect to change in electric current using a first-order dynamic model. The overall dynamic process model is then used in a computational model predictive control (MPC) scheme to drive the process to a new H₂ production set-point under unperturbed and steam flowrate disturbance cases. The performance and robustness of the proposed MPC scheme are compared to the ones of a classical proportional–integral (PI) controller and are demonstrated to be superior in terms of closed-loop response, robustness, and constraint handling.

The production of therapeutic proteins in mammalian cell culture is an essential unit operation in biopharmaceutical manufacture that can benefit from the predictive insights of effective process models, leading to accelerated process development and improved process control. This review outlines and evaluates current approaches to predictive model development for mammalian cell culture and protein production. Classical mechanistic and data driven approaches are analysed, together with potential challenges in model development and application, including the experimental requirements for parameter estimation. Hybrid models, which may offer greater robustness, are then explored along with hybrid model architecture and the steps involved in model development. Successful examples from other cell fermentation processes are also considered, for application to the development, monitoring and control of mammalian processes.