Manufacturing polypropylene (PP) composites to meet customers’ needs is difficult, time-consuming, and costly, owing to the ever-increasing diversity and complexity of the corresponding specifications and the trial-and-error method currently used to satisfy the required physical properties. To address this issue, we developed three models for predicting the physical properties of PP composites using three machine learning (ML) methods: multiple linear regression (MLR), deep neural network (DNN), and random forest (RF). Further, the industrial data of 811 recipes were acquired to verify the developed models. Data categorization was performed to account for the differences between data and the fact that different recipes require different materials. The three models were then deployed to predict the flexural strength (FS), melting index (MI), and tensile strength (TS) of the PP composites in nine case studies. The predictive performance results differed according to the physical properties of the composites. The FS and MI prediction models with MLR exhibited the highest R² values of 0.9291 and 0.9406. The TS model with DNN exhibited the highest R² value of 0.9587. The proposed models and study findings are useful for predicting the physical properties of PP composites for recipes and the development of new recipes with specific physical properties.

Modern chemical processes comprise multiple interconnected unit operations with complex dependencies, in which even minor faults can lead to product loss, quality degradation, energy waste, resource loss, and environmental pollution. These issues can be mitigated through anomaly-detection models, which enable early detection of abnormal process conditions and timely corrective actions. However, conventional statistical and data-driven fault-detection methods often assume that the process is either static or governed by simple dynamics, thus failing to adequately capture the nonlinear and high-dimensional characteristics of real-world industrial systems. To address these limitations, this paper proposes a novel anomaly-detection framework called the LSTM–transformer autoencoder (LTAE). The model employs a long short-term memory (LSTM) encoder to effectively capture dynamic temporal changes in the process data, and a transformer-based decoder with multi-head attention learns long-range dependencies and critical information in parallel to achieve precise reconstruction. When applied to the Tennessee Eastman (TE) benchmark process under a single-mode setting, the proposed model maintained a low false alarm rate (FAR) of 0.00, similar to the baseline models, while enhancing the fault detection rate (FDR) from 0.83 to 0.86. Furthermore, in a multimode setting using a real-world industrial polyethylene (PE) production process, the model significantly reduced the FAR from 0.85 (dual-attention LSTM, transformer) to 0.05, while sustaining an FDR of 1.00. These results demonstrate that the proposed model delivers robust anomaly detection performance in both single- and multimode environments, serving as a practical solution to minimize resource waste, reduce unnecessary energy consumption, and mitigate environmental risks.

Machine learning (ML) is a promising method for finding optimal materials for the synthesis of effective catalysts. However, the application of ML to the exploration of effective heterogeneous thermal catalysts, such as those for dry reforming of methane (DRM), is challenging because the performance of such catalysts is based on various complex factors. To address this issue, we propose an interpretable machine learning (IML) framework to predict the performance of a DRM catalyst. CH 4 conversion, a performance indicator of DRM catalysts, is a function of design variables (such as composition and components of active metal, support, and promoter), pretreatment conditions (such as calcination and reduction), and operating conditions. The framework predicts catalyst performance while considering the contribution of reaction temperature, the most dominant variable in DRM reaction, by using an ML model. First, the shapley additive explanation (SHAP) and partial dependence plot (PDP) are systematically used as an IML tool in a framework for developing reliable and transparent ML models. The prediction accuracy of the model improved (R 2 : 0.91 to 0.96) after data set modification using SHAP. Second, PDP was used to obtain optimal operational and pretreatment variables. A dataset was generated via PDP to reflect operation and pretreatment variables while reducing the contribution of the reaction temperature to CH 4 conversion predictions. The contribution of each variable to the conversion was described plausibly, and promising DRM catalyst candidates at each reaction temperature (600, 700, and 800 °C) were recommended using the developed framework.