Surrogate Models for CO ₂ Utilization Pathways: Accelerating Industrial Park Design and Enterprise-Wide Optimization

This study develops and validates surrogate models in a standardized environment for three key CO₂ utilization processes: methanol, ammonia, and urea production. Linear and nonlinear models were constructed using Latin Hypercube Sampling data from Aspen Plus simulations and evaluated across multiple data set sizes (40-320 points). Performance was assessed using a comprehensive set of statistical metrics (<i>R</i> ², adjusted <i>R</i> ², predicted <i>R</i> ², 5-fold cross-validation <i>R</i> ², RMSE, and MAE), supported by confidence interval plots to capture both predictive accuracy and generalization. Results show that quadratic models more effectively capture nonlinear trends in the methanol and ammonia processes, particularly for larger data set sizes, while linear models maintain strong generalization and computational efficiency, especially for urea. For small sample sizes, inflated <i>R</i> ² values masked overfitting in quadratic models, an issue revealed only through predicted <i>R</i> ², cross validation, and error metrics. Convergence of all performance measures at higher data sizes confirmed the importance of sufficient sampling to ensure robust surrogate behavior. Computational performance benchmarking of selected models illustrated significant savings in prediction times compared to those of simulation. Beyond individual process insights, the validated surrogate models presented here serve as a benchmark for different CO₂ utilization processes in a standardized testing environment. They provide a consolidated and reusable resource for CO₂ utilization research and are intended as modular building blocks that can be readily embedded into enterprise-wide optimization frameworks of industrial eco-park network models. A case study is also provided to demonstrate error propagation in interconnected process units replaced by surrogates resulting in low deviations from simulation. By reducing computational burden while maintaining predictive reliability, these models support both academic investigations and practical decarbonization initiatives.