Application of Machine Learning to Fischer–Tropsch Synthesis for Cobalt Catalysts

Machine Learning was used to make a prediction model for six property parameters (COconv (%); CH4 (%); CO2 (%); C2–C4 (%); C5+ (%); and COconv*C5+ (%)) from 16 feature parameters of 169 Fischer–Tropsch synthesis experiments for cobalt catalyst. The Random Forest method was chosen as the "black-box" prediction tool, and cross-validation tests revealed small mean average errors: 10.9 ± 1.8; 3.7 ± 0.7, 1.5 ± 0.6, 3.6 ± 0.8, 7.4 ± 1.4 and 9.0 ± 1.3, respectively. The most important feature parameters were Brunauer–Emmett–Teller (BET) surface area, pressure (P), and temperature (T). The Decision Tree was chosen to build an explainable model, which revealed that BET in the range of [199.5, 529] leads to small COconv (%); C5+(%); and COcon*C5+(%) values. Outside this range, the samples tend to have large conversion and selectivity values. Most importantly, the BET effect is not linear or monotonic; it can be extracted using machine learning methods, and current work has shown how to use it. The resulting rules can be used for further experiments to maximize the efficiency and develop a new catalyst system. Particularly, porous Co3O4 with a small admixture of Al2O3 particles, Co2O3 by classical calcination of quartz sand impregnated with a cobalt nitrate solution, and a catalyst using a hydroxyapatite support were identified as good candidates for further deep screening of the best catalysts.