|
Heterogeneous catalyst design for CO2 conversion using machine learning |
|---|---|
| รหัสดีโอไอ | |
| Title | Heterogeneous catalyst design for CO2 conversion using machine learning |
| Creator | Sukanlaya Kornnum |
| Contributor | Pawin Iamprasertkun, Advisor |
| Publisher | Thammasat University |
| Publication Year | 2568 |
| Keyword | CO2 hydrogenation, Methanol, Copper-based catalyst, Machine learning, Space-time yield |
| Abstract | Carbon dioxide (CO₂) hydrogenation to methanol has attracted considerable attention as a carbon utilization strategy because methanol serves as an important chemical feedstock and sustainable fuel. Copper-based catalysts are among the most widely studied systems for this process; however, optimizing catalyst composition and operating conditions remains challenging due to their complex interactions. Furthermore, machine learning studies in catalysis often rely on random train-test splitting, which may overestimate predictive performance and provide limited insight into model generalizability toward newly developed catalysts. In this work, a literature-derived dataset comprising 694 experimental data points collected from 78 publications published between 1998 and 2022 was assembled to investigate Cu-based catalysts for CO₂ hydrogenation to methanol. Thirteen machine learning models were systematically evaluated under a three-scenario evaluation framework combining two data splitting strategies random split and temporal split with two feature conditions: GHSV included and GHSV excluded. GHSV was identified as a target leakage variable due to its direct mathematical presence in the STY formula, contributing ΔR² = 0.493 of artificial performance inflation. Under random split with GHSV included (Scenario A), XGBoost achieved the best performance with R² = 0.932. Under temporal split with GHSV included (Scenario B), LightGBM generalized best at R² = 0.305. Under the most rigorous condition temporal split with GHSV excluded (Scenario C) LightGBM achieved R² = -0.188, with total methodological inflation of ΔR² = 1.100 between Scenario A and C. Analysis revealed a fundamental composition shift between the training (1998-2018) and test (2019-2022) periods, with ZnO-based catalysts declining from 69.1% to 25.5% while In-promoted systems absent from training comprised 28.8% of the test set, explaining the generalization failure as a dataset evolution rather than a modelling deficiency. Without GHSV interference, interpretability analyses using SHAP, Feature Permutation Importance, and Partial Dependence Plots identified SBET and metal loading as dominant catalyst design parameters, with pressure as the primary operational lever, and provided practical design guidelines: Cu loading 20-35 wt.%, SBET > 150 m2 g-1, pressure 2-10 MPa, and temperature 490-560 K. |
