The
efficient and effective design of chemical processes and products
heavily relies on the accurate prediction of essential properties.
In this work, a deep-learning architecture integrating a bidirectional
long short-term memory (Bi-LSTM) network, an attention mechanism,
and a back-propagation neural network (BPNN) is developed for the
prediction of energy conversion efficiency of organic solar cells.
Inspired by the success of artificial intelligence in natural language
processing, we first developed a novel strategy for molecular signature
encoding and information embedding in order to depict the compositional
structures of molecules. Then, an advanced recurrent neural network
named Bi-LSTM is employed to process the molecular information, while
the BPNN is applied to correlate energy conversion efficiency. During
this procedure, the attention mechanism is used to identify molecular
constituents that are important to the property of interest. To evaluate
the performance of the proposed approach, the energy conversion efficiencies
of more than 20,000 organic photovoltaics are used to train and test
the model. Result comparisons with several other modeling approaches
indicate that the proposed method is competitive in prediction accuracy
and possesses good transferability to small data sets. Additionally,
the proposed method is capable of identifying decisive molecular constituents,
providing instructive information for the reverse design of organic
solar cells.
The mechanism models based on structure-oriented lumping (SOL) deliver a satisfactory prediction on the properties and yield distribution of the products from fluid catalytic cracking (FCC). However, with high complexity and low computing efficiency, such a model is increasingly unable to meet the needs of refineries to produce lighter and greener products using heavier and poorer feedstocks. Therefore, in this paper, a modeling approach hybridizing molecular mechanism and data models was proposed to describe the maximizing iso-paraffins (MIP) technology of the FCC process. This proposed model showed assured prediction accuracy with shortened computing time and thus was appropriate for online application. In this work, model simplification was carried out: less molecules and reactions (3078 and 5216, respectively) were adopted, along with a simplified reactor model, which largely reduced the computation load. CatBoost algorithm was also adopted for constructing a data model, to compensate for the accuracy loss resulting from the simplified SOL mechanism model. Combining with the mechanism model, it ensured the accuracy of prediction while greatly shortened the computing time. Furthermore, to overcome the strong coupling between the process variables to be solved, this work adopted the method of case-based reasoning (CBR) to optimize the process and expanded the case base with the prediction results of the hybrid model, which ensured the feasibility of the solution parameters and shortened the computing time. The hybrid model and the corresponding process optimization strategy proposed were then applied to an industrial FCC MIP process for verification. The results show that the hybrid model could assure the prediction accuracy (comparable with the conventional mechanism model) while the computing time was reduced from more than 20 h to less than 1 min. In the process optimization validation test, the total liquid yield increased by 1.19% on average for 43 out of 50 sets of operating configurations and the coke yield decreased by 1.05% on average. This work provides a good solution for the online process optimization of FCC.
Four kinds of π-complexation adsorbents are synthesized via ion exchange method or incipient wetness impregnation method with Amberlyst 35, SBA-15, TUD-1, and KIT-6 as supports, and AgNO 3 as active ingredient. The samples are characterized by N 2 adsorption/desorption. Fourier transformed infrared spectroscopies, transmission electron microscopy, X-ray diffraction spectrum analysis, and coupled plasma optical emission spectrometry are also used as adsorbents for ethylene/ethane adsorptive separation. The results show that the specific surface area and the dispersion of silver ions affect the separation performance of the adsorbent. Ag−Amberlyst 35 has the highest ethylene/ethane selectivity among these adorbents while the adsorbed amounts of ethylene of the three mesoporous silica complexation adsorbents are higher. Adsorption thermodynamics analysis suggests that the interaction of ethylene with adsorbent is a mild chemical adsorption. An adsorption kinetics study indicates that the adsorption of ethylene on silver-supported mesoporous materials is not a simple diffusion-control process. The adsorption behavior of ethylene on the π-complexation adsorbent has an energy barrier in the range of 24−33 kJ/mol. Among the adsorbents in this work, the KIT-6-based adsorbent has the best mass kinetic performance due to its three-dimensional regular interconnected mesoporous structure.
This
study proposes a hybrid approach for the modeling of the fluid
catalytic cracking (FCC) process, with the aim to establish an adaptive
and accurate product yield prediction model. Because of the uncertainties
in crude oil quality and the complexity of the FCC process, which,
for example, has highly coupled process variables with high dimensionality
and strong interference, it is difficult for existing first-principles-based
methodologies to deliver accurate results. To tackle this, this study
proposes a machine-learning-based modeling approach that integrates
an intelligent feature selection strategy with random forest for the
process modeling. First, the adaptive immune genetic algorithm (AIGA)
is applied to screen for the most relevant process indicators from
the collected process data, including the operation parameters for
the relevant process devices and the property data of the feed stream.
Second, random forest (RF) is employed to establish the FCC process
models based on the selected process indicators in the first step.
The approach is illustrated by its application in a real FCC production
process, for which 10 months of historical production data were used
to train and test the proposed AIGA-RF model to determine the product
yield predictions for four products. Comparisons between the proposed
method and other methods were also conducted. The result indicates
that the proposed method is able to remove the disturbance variables
and is found to be adaptable to different product yield prediction
scenarios. It could be a good reference for online process optimization
and control of FCC processes.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.