Interpretable mechanism-guided and data-driven model for completion design

• Introduces an interpretable mechanism-guided completion optimization model. • Weighted fuzzy model identifies geological sweet spots with high correlation (ρ = 0.51) • Reveals high-yield patterns boosting productivity by 40% - 75% under geo-constraints. • Optimized completion design increases pump rate by 14% and proppant intensity by 22%. • Overcomes black-box opacity to ensure transparent completion design. Effective completion design is crucial for unconventional reservoirs, yet existing workflows and black-box AI models struggle to handle geological heterogeneity and interpretability. To address these challenges, this study proposes an interpretable, mechanism-guided, and data-driven completion optimization model. First, a weighted fuzzy C-means model is developed to identify geological sweet spots, where petrophysical parameters are assigned weights based on their correlation with oil production to enhance physical consistency. Subsequently, under geological constraints, engineering parameters are analyzed using production-weighted clustering to isolate high-yield patterns. A dynamic fracturing-stage design method integrating sweet-spot probability and Shannon-entropy-based boundary fuzziness is then introduced to adaptively delineate geologically consistent stages. Finally, a two-step completion optimization strategy, historical-knowledge initialization followed by CART-based interpretable refinement, maps engineering parameters to production outcomes with transparent, rule-based logic. Validation in the Ordos Basin showed the weighted fuzzy model significantly outperformed conventional methods, with sweet spot identifiers strongly correlating with oil production (correlation coefficient of 0.51). The geology-constrained clustering revealed that high-intensity parameter patterns yield 40%–75% higher productivity. When applied to a representative test well, the optimized design increased pump rate by 14% and proppant intensity by 22%, successfully converting all low-yield stages into high-yield ones. These results confirm that the proposed model overcomes black-box limitations, offering a transparent and robust pathway for intelligent completion design.