(697e) Monitoring Catalysts Synthesis Using Real-Time Emission Spectroscopy and Advanced Machine Learning Models

Flame spray pyrolysis (FSP) is ideal for designing desired catalysts since it allows for control over material properties. Combining in-situ laser-induced breakdown spectroscopy (LIBS) with FSP can monitor nanoparticles’ chemical properties through their emission spectra. The difficulty in LIBS spectra analysis can be overcome using advanced machine learning (ML) models.

In this work, a series of Pd/Ce_xZr_yMn_z (where x, y, z = 0,1,2) catalysts were synthesized via FSP. Their LIBS spectra were analyzed using the linear-supervised vector classifier (LSVC) and polynomial-supervised vector regressor (PSVR) to predict phase presence, lattice constant, and oxygen vacancy percentage (OV%). The crystalline structure, lattice constant, and OV% were determined by X-ray Diffraction (XRD) and Raman spectroscopy. Samples can present five different crystal structures, including cubic CeO₂, tetragonal ZrO₂, tetragonal Mn₃O₄, tetragonal α-MnO₂, and tetragonal β-MnO₂. An average accuracy can reach above 85% over the LSVC algorithm. For lattice constant prediction, the cross-validated R² and RMSE over the PSVR algorithm are calculated as 0.72 and 0.0462, respectively. For OV% prediction, the PSVR algorithm can lead to the cross-validated R² and RMSE of 0.71 and 6.924%, respectively. Considering the complexity of LIBS spectra and the small dataset size of 76 samples, both LSVC and PSVR models can reasonably predict phases, lattice contact, and OV% from LIBS patterns. The models can successfully predict the chemical properties of same-type compounds without requiring further post-characterization, which can speed up the discovery of novel functional catalysts.