Application of Machine Learning for in-Situ Analysis of Catalyst Quality

Flame spray pyrolysis (FSP) provides a fast and continuous route for the preparation of nano-sized oxide-supported transition metal catalysts that play a vital role in oxidizing unburned methane at low temperature in natural gas-fired combustion engines. Determining the performance of these catalysts requires post-synthesis characterization, which hinders the ability to continuously produce high-quality catalysts. Our work incorporates machine learning and laser-induced breakdown spectroscopy (LIBS) catalyst synthesis, allowing one to tune the experimental parameters to yield the desired product selectivity. LIBS is an ideal in-situ analysis technique to monitor particle information regardless of sample phase but difficult to analyze without machine learning due to various spectral emissions mixed with fuel, precursors, and by-products.

Herein, Pd supported on Ce-Zr-Mn oxide solid solution catalysts synthesized by FSP are employed for low temperature methane oxidation. X-ray diffraction (XRD) and Raman spectroscopy were carried out to determine the chemical structure of the synthesized catalysts. LIBS, XRD, and Raman spectroscopy data were then used to obtain the support vector machine (SVM) algorithms and predict catalyst quality. Support vector classifier was used to classify the LIBS data to predict crystalline phases and oxygen vacancy percentages, whereas support vector regressor (SVR) was chosen to predict the crystalline lattice constant from the LIBS data. To evaluate the performance of support vector classifier (SVC) and SVR; LIBS, XRD, and Raman data were collected from various FSP experiments of unique state. These parameters were tuned by different precursor/gas flow rate, precursor concentration, and the precursor concentration ratios of the dopant metals. The performance of the classifiers and regression models were calculated by comparing the predicted and experimental data. Our results recorded an average accuracy of 85.86% to detect the presence of crystalline phases and 93.08% to predict low, medium, or high oxygen vacancy ranges. The SVR model was able to predict lattice constants with an average error of 0.73%. The high performances of these classifiers/model allow one to gain a perspective of the quality of the synthesizing catalysts before post-synthesis characterization, ultimately accelerating the process of producing high-quality catalysts.