Modelling a reverse water-gas shift reactor with an artificial neural network

• Data-driven multi output neural network surrogate for a tubular rWGS fixed-bed reactor • SHAP-based explainable AI reveals temperature and H 2 :CO 2 as dominant levers, with pressure mainly shaping CO selectivity • ANN surrogate model is deployable for Power-to-Liquid python simulations For process intensified Power-to-Liquid (PtL) systems, the reverse water-gas shift (rWGS) reaction is a key step to convert CO 2 and H 2 into CO-rich syngas. Conventional modelling approaches either provide thermodynamic limits or require kinetic approaches which still struggle to represent catalyst- and reactor-specific effects. In this work, a multi-output artificial neural network (ANN) is developed as a data-driven surrogate model for a tubular Ni/γ-Al 2 O 3 fixed-bed rWGS reactor. The network is trained on 396 experimentally determined operating points covering 550-950°C, 1-8 bara, GHSV of 8,000-20,000 h -1 and H 2 :CO 2 ratios between 2 and 3. The ANN predicts both CO 2 conversion (X CO2 ) and CO selectivity (S CO ) with an R 2 of 0.97 and test-set errors of MAE 1.9% and RMSE 3.3%. Compared to Gibbs-equilibrium calculations, the surrogate accurately reproduces experimentally observed deviations from equilibrium in the low- and mid-temperature regime and captures the stronger pressure sensitivity of S CO relative to X CO2 . Explainable AI analysis using SHAP identifies temperature and H 2 :CO 2 as dominant drivers, with pressure mainly steering methanation-driven selectivity losses and GHSV becoming relevant at low temperatures. The proposed framework provides an experimentally validated, interpretable tool for fast rWGS reactor screening and PtL process optimisation.