Computational Fluid Dynamics-Informed Machine Learning-Enabled Digital Twin for Vapor Maldistribution Prediction in Structured Packed Distillation Columns
- 1School of Chemical Engineering
- 2Oklahoma State University
- 3Fractionation Research Institute
Abstract
Vapor maldistribution in structured-packed distillation columns reduces separation performance. While experimental measurements are essential for characterizing column hydrodynamics, they often lack the resolution to capture velocity distributions or to systematically quantify how column geometry and operating conditions affect vapor maldistribution throughout the packed bed. High-fidelity, first-principles computational fluid dynamics (CFD) simulations can provide detailed spatial flow information, but their high computational cost limits the rapid evaluation of alternative column designs and operating conditions. To address this limitation, a CFD-informed machine-learning (ML) reduced-order model (ROM) framework was developed to rapidly predict vapor flow uniformity in structured-packed distillation columns. 3D CFD simulations were performed for 13 column geometries by varying column diameter, inlet diameter, nozzle-to-packing clearance, and vapor mass flow rate, yielding a dataset of 65 simulation cases for parametric analysis and ML model training. Vapor maldistribution was quantified using the area-weighted coefficient of variation (CoV) of streamwise velocity across 13 packing layers. CFD results indicated that vapor maldistribution was most pronounced near the packed-bed entrance and progressively decreased through the packing. Higher vapor flow rates and smaller inlet diameters generally increased maldistribution in the inletdominated region, while the effect of nozzle-to-packing clearance depended on column geometry and scale. For instance, increasing the inlet diameter reduced the Layer-1 CoV from 51.01% to 30.73% in the 47.75-in. column and from 48.82% to 22.32% in the 120in. column. A Gaussian process regression (GPR)-based ROM was subsequently trained using the CFD dataset and evaluated through leave-one-geometry-out cross-validation. Across 65 held-out cases and 845 packed-layer predictions, the ROM achieved a mean absolute error of 0.851, a root mean squared error of 1.524, and a coefficient of determination of 0.980. Prediction errors were primarily concentrated in the lower packing layers, where the flow was strongly influenced by the inlet geometry, and decreased substantially farther downstream. The resulting CFD-informed ML framework offers a computationally efficient digital-twin development workflow for predicting vapor maldistribution in previously untested column geometries and, within the investigated design space, enables rapid geometry screening and design optimization of structuredpacked distillation columns.
Related
- Computational Evaluation of Carbazole–Chalcone Urea and Thiourea Derivatives Against EGFR: Molecular Docking, Pharmacokinetic Profiling, Target Prediction, and Molecular Dynamics Simulation
- Computational Prediction of Ligand-Binding Sites of Proteins by Combining Hydrophobic Area Search and Residue-Based Scoring
- AGAPE (Computational G-Quadruplex Stabilization Prediction): The First Machine Learning Workflow for G-Quadruplex Stabilization Prediction
- Multi-omics integration reveals the role of PFOS in sarcopenia via ESRRG and RBP4: from computational prediction to experimental validation
- A computational dynamics and damping–induced transitions in beam–foundation systems with viscoelastic constraints
BibTeX
@misc{nino2026computational,
title = {{Computational Fluid Dynamics-Informed Machine Learning-Enabled Digital Twin for Vapor Maldistribution Prediction in Structured Packed Distillation Columns}},
author = {Alek Nino and Hang Yi and Tony Cai and Izak Nieuwoudt and Yu Feng},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009199/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009199/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Computational Fluid Dynamics-Informed Machine Learning-Enabled Digital Twin for Vapor Maldistribution Prediction in Structured Packed Distillation Columns
AU - Alek Nino
AU - Hang Yi
AU - Tony Cai
AU - Izak Nieuwoudt
AU - Yu Feng
PY - 2026
DA - 2026/09/21
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009199/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009199/v1
AB - Vapor maldistribution in structured-packed distillation columns reduces separation performance. While experimental measurements are essential for characterizing column hydrodynamics, they often lack the resolution to capture velocity distributions or to systematically quantify how column geometry and operating conditions affect vapor maldistribution throughout the packed bed. High-fidelity, first-principles computational fluid dynamics (CFD) simulations can provide detailed spatial flow information, but their high computational cost limits the rapid evaluation of alternative column designs and operating conditions. To address this limitation, a CFD-informed machine-learning (ML) reduced-order model (ROM) framework was developed to rapidly predict vapor flow uniformity in structured-packed distillation columns. 3D CFD simulations were performed for 13 column geometries by varying column diameter, inlet diameter, nozzle-to-packing clearance, and vapor mass flow rate, yielding a dataset of 65 simulation cases for parametric analysis and ML model training. Vapor maldistribution was quantified using the area-weighted coefficient of variation (CoV) of streamwise velocity across 13 packing layers. CFD results indicated that vapor maldistribution was most pronounced near the packed-bed entrance and progressively decreased through the packing. Higher vapor flow rates and smaller inlet diameters generally increased maldistribution in the inletdominated region, while the effect of nozzle-to-packing clearance depended on column geometry and scale. For instance, increasing the inlet diameter reduced the Layer-1 CoV from 51.01% to 30.73% in the 47.75-in. column and from 48.82% to 22.32% in the 120in. column. A Gaussian process regression (GPR)-based ROM was subsequently trained using the CFD dataset and evaluated through leave-one-geometry-out cross-validation. Across 65 held-out cases and 845 packed-layer predictions, the ROM achieved a mean absolute error of 0.851, a root mean squared error of 1.524, and a coefficient of determination of 0.980. Prediction errors were primarily concentrated in the lower packing layers, where the flow was strongly influenced by the inlet geometry, and decreased substantially farther downstream. The resulting CFD-informed ML framework offers a computationally efficient digital-twin development workflow for predicting vapor maldistribution in previously untested column geometries and, within the investigated design space, enables rapid geometry screening and design optimization of structuredpacked distillation columns.
ER -