VIBES framework slashes predictive uncertainty by 80% in complex chemical systems
Two-stage Bayesian UQ reduces computational burden by combining Sobol screening with variational inference.
Bayesian uncertainty quantification in complex chemical processes often becomes computationally intractable due to high-dimensional parameter spaces. The newly proposed VIBES (Variational Inference-based Bayesian Estimation with Sobol screening) framework tackles this head-on with a two-stage approach. First, it uses Sobol global sensitivity analysis to screen and identify the few dominant variables and parameters that drive uncertainty in process outputs. Then, variational inference is performed only on this reduced-dimensional subspace, dramatically cutting computational cost while preserving accuracy.
The authors demonstrate VIBES on a real-world biomass valorization process: bioadhesive production via base-catalyzed depolymerization of kraft lignin and crosslinking with isolated soy protein. A custom Python-Aspen interface automates simulation and parameter estimation, enabling gradient-based Bayesian calibration with automatic differentiation. Results show that even when only the reduced-space variables are optimized, predictive uncertainty bounds shrink by more than 80% across all model outputs. The methodology is generic and readily generalizable to other biomass conversion pathways and high-dimensional chemical engineering systems, offering a scalable path to uncertainty-aware decision-making.
- Two-stage framework: Sobol GSA for dimensionality reduction, then variational inference for Bayesian UQ on reduced space.
- Applied to lignin-based bioadhesive production; Python-Aspen interface enables automated simulation and gradient-based calibration.
- Reduces predictive uncertainty bounds by over 80% compared to full-space Bayesian estimation.
Why It Matters
Enables scalable, uncertainty-aware optimization of complex chemical processes, cutting computational costs for industry and research.