Augmented adaptive learning-based prediction and mixed-integer nonlinear optimization for catalyst discovery in CO2-to-methanol process

The CO2-to-methanol (CTM) process faces critical challenges primarily due to the lack of viable methods for accurately predicting catalyst performance and identifying optimal catalyst candidates based on limited experimental data while adhering to chemical reaction constraints. To address these challenges, this study presents a novel and high-fidelity framework that integrates augmented adaptive learning-based prediction and mixed-integer nonlinear optimization for identifying effective catalyst compositions, synthesis conditions, and operating conditions for the CTM process. 1272 experimental samples are initially collected as raw data. Outliers are detected and removed using the fast Fourier transform method, followed by feature selection based on Pearson correlation analysis and domain knowledge, resulting in 368 refined samples. To overcome the challenges of experimental data sparsity and incorporate chemical reaction constraints, an augmented adaptive deep neural network (AA-DNN) model is developed, including two components: a variational autoencoder for data augmentation and an adaptive DNN for performance prediction. Using the refined dataset, the predictive performance of the AA-DNN model is compared against conventional DNN and augmented DNN models, demonstrating a 6.58% to 10.34% improvement in prediction accuracy. The developed AA-DNN model is subsequently used to formulate two mixed-integer nonlinear optimization problems to demonstrate its effectiveness in catalyst discovery. At the optimal catalyst components, compositions, synthesis conditions, and operating conditions, the Cu/ZnO/In2O3 (CuZnIn) catalyst demonstrates the best performance, improving CO2 conversion by 4.73% and methanol yield by 2.9% compared to the experimental dataset. Notably, the optimization process requires only 7.03 s, representing a significant reduction in screening time compared to conventional experimental approaches. These findings provide valuable insights into data-driven catalyst development for the CTM process. Additionally, due to the shared characteristics of CO2 conversion pathways, the proposed framework offers a promising and extensible tool for accelerating catalyst discovery across a broad range of CO2 conversion processes.