AI-DRIVEN PREDICTION AND ANALYSIS OF EXPERIMENTAL METHANE CONVERSION IN TRI-REFORMING CATALYSTS

Abstract

The discovery of catalysts using machine learning requires experimental datasets of high fidelity without literature bias. This research provides a Random Forest regression model that has been trained on fresh triplicate measurements (550-750 ^o C) of Ni-catalyzed tri-reforming (TRM), and it gives a state-of-the-art test result with R²=0.953, RMSE=1.43%. The model completely removes multi-dimensional confounding in heterogeneous literature models by isolating temperature alone, proving that 100 percent of the variation is due to temperature by permutation importance analysis and partial dependence analysis. Visualization of response surfaces in three dimensions shows the identification of three kinetic regimes: activation-limited (550-600 ^oC), exponential growth (600-680 ^oC) and equilibrium-limited (>680 ^oC) with an operating window of 680-720 ^oC (>30% conversion, uncertainty of 1.2 percentage points). Hold-one-replicate-out cross-validation shows that it is time-robust (mean RMSE=1.40% across Conv1/2/3), and extrapolation surfaces predict physics at 827 °C with no spurious artifacts. The autonomous reactor control is edge deployable with computational efficiency (O (n log n), 0.1ms inference). The major innovations comprise data-efficient convergence of 25 points, removing the sample hunger of deep learning, and scalable uncertainty quantification of ±1.962 experimental standard that is safety critical. The experimental-only method with dataset purification (RMSE= 4x better than literature RF ensembles) reduces the error by 4x. The generalization of the temperature-collapse paradigm is applied in the context of thermal catalysis, and this makes TRM prediction a reproducible computing benchmark between the lab-to-plant translation. It is a groundbreaking work in hybrid computing-chemical research, as it shows that single-feature RF is more effective than the high-dimensional complexity on sparse industrial data. Model Predictive Control with quantified confidence is used with deployable digital twins to make autonomous hydrogen production economies based on the precision of the ML.