AI-accelerated discovery of defect-engineered heteroatom-doped carbon electrocatalysts for electrochemical CO₂ reduction to C₂⁺ products

Selective electrochemical CO₂ reduction to multi-carbon products remains a central bottleneck for carbon-neutral chemical manufacturing, with most catalysts achieving C₂⁺ Faradaic efficiencies below 75%. This investigation integrates high-throughput density functional theory with ensemble machine learning to screen 8000 heteroatom-doped carbon configurations. Ensemble models combining Random Forest, Gradient Boosting, and XGBoost achieve cross-validated R 2 = 0.8760 ± 0.0274 with Gaussian residuals, enabling prediction of C₂⁺ Faradaic efficiency from electronic, geometric, and adsorption descriptors. Computational optimization identifies ternary N–S–P–doped architectures with Stone–Wales defects (7.3 at%, 4.6% defect density) predicted to exhibit C–C coupling barriers of 0.61 eV within the thermodynamically favorable 0.5–0.75 eV window. DFT-optimized atomic configurations reveal three synergistic mechanistic pathways; nitrogen substitution lowers the local work function from 4.5 to 4.2 eV, reducing the onset overpotential by 0.18 V; complementary S–P electronegativity contrast establishes a Bader charge asymmetry of Δ q = 0.31 at adjacent carbon sites, preferentially stabilizing the *COCO transition state; and Stone–Wales rearrangement elongates C–C bonds to 1.44 Å, collectively reducing the coupling barrier by Δ Δ G ‡ = 0.30 eV relative to pristine graphene, confirmed by CI-NEB calculations on 12 representative ternary configurations. *COCO adsorption energy as the dominant descriptor (importance = 0.1471). Multi-objective Pareto optimization yields 301 configurations, with the highest-performing candidate CAT06928 predicted to achieve 89.5 ± 1.7% C₂⁺ Faradaic efficiency, 301 ± 9 mA cm⁻² current density, and C₂⁺/CO selectivity ratio of 15.71. Techno-economic modeling projects costs below $0.50 per kg and 51.4% life-cycle carbon footprint reduction relative to fossil-fueled thermal CO₂ conversion, providing quantitative design principles for metal-free electrocatalysts requiring experimental validation. • Screened 8000 N,S,P-doped defect-rich carbon catalysts via DFT. • Ensemble ML achieves R² = 0.8760 ± 0.0274 for FE prediction. • *COCO adsorption energy emerges as key C–C coupling descriptor. • Optimal design: 7.3 at% heteroatoms, 4.6% defects, 0.61 eV barrier. • Predicted C₂⁺ FE 89.5 ± 1.7% with 51.4% lower CO₂ footprint than thermal.