Post-Combustion CO₂ Capture — ML Surrogate Model

A physics-grounded machine-learning surrogate for an MEA-based absorption column.
Predicts capture efficiency, reboiler duty (the key CCS economic lever), and
solvent degradation as functions of operating conditions — then finds the operating
envelope that minimises energy penalty at ≥90 % capture.

Background

Post-combustion capture using 30 wt% monoethanolamine (MEA) is the benchmark CCS
technology. The absorption column physics follow:

• Mass transfer: two-film theory with chemical enhancement
  K_OG = 1 / (1/k_G + 1/(k_L · E))
  where the enhancement factor E ≈ √(k₂·C_MEA / k_L²) (Hatta regime)
• NTU-HTU model for column sizing
• VLE: simplified Austgen correlation
  P*_CO₂ = 10⁴ · α² · exp(4000·(1/333 − 1/T_lean))
• Reboiler duty decomposed into sensible heat + heat of absorption + stripping steam
• Degradation: thermal + oxidative + carbamate precipitation pathways

Surrogate ML replaces expensive rigorous simulations (Aspen Plus / gPROMS) for
real-time optimisation.

Targets

Target	Unit	Typical range
capture_efficiency	fraction	0.20 – 0.995
reboiler_duty_GJ_t	GJ / t CO₂	2.5 – 5.0
solvent_degradation_pct_per_1000h	% MEA / 1000 h	0.3 – 7.5

Features (inputs)

Feature	Description	Range
y_CO2	CO₂ mole fraction in flue gas	0.04 – 0.16
y_O2	O₂ mole fraction	0.02 – 0.08
y_N2	N₂ mole fraction (balance)	—
T_flue_K	Flue-gas temperature [K]	313 – 363
LG_ratio	Liquid/gas mass ratio [kg/kg]	2.5 – 8.0
lean_loading	Lean solvent loading [mol CO₂/mol MEA]	0.15 – 0.40
MEA_conc_mol_L	MEA concentration [mol/L]	3.0 – 7.0
P_total_bar	Column pressure [bar]	1.0 – 1.5
T_lean_K	Lean solvent temperature [K]	308 – 328
col_height_m	Packing height [m]	5 – 20

Quick Start

pip install -r requirements.txt

# 1. Generate synthetic dataset
python src/data_generation.py

# 2. Train surrogate models
python src/train.py

# 3. Find optimal operating envelope
python src/optimize.py --target_capture 0.90 --max_degradation 4.0

# 4. Evaluate and plot
python src/evaluate.py

Models

Three model families are compared:

• Gradient Boosted Trees (XGBoost) — best overall accuracy
• Random Forest — interpretable feature importance
• Neural Network (MLPRegressor) — captures high-order interactions

Multi-output wrapper trains one model per target. Best model is saved to models/.

Results (example)

Model: XGBoost
─────────────────────────────────────────────────────
Target                        R²      MAE
capture_efficiency           0.984   0.009
reboiler_duty_GJ_t           0.971   0.052 GJ/t
solvent_degradation          0.966   0.083%/1000h
─────────────────────────────────────────────────────

Optimal operating point (η_cap ≥ 0.90, Q_reb minimised):
  LG_ratio        = 5.8
  lean_loading    = 0.32
  MEA_conc_mol_L  = 5.5
  T_lean_K        = 315
  → Q_reb = 2.91 GJ/t CO₂   |   η_cap = 0.927   |   deg = 1.2%/1000h

Results Visualisations

Project Layout

src/data_generation.py   Physics-based LHS data generation
src/features.py          Feature engineering (interaction terms, scaling)
src/train.py             Train + CV + save best model
src/optimize.py          Constrained optimisation over surrogate
src/evaluate.py          Metrics, parity plots, SHAP feature importance
notebooks/               Exploratory analysis
tests/                   Unit tests for data + model pipeline
docs/                    Physics derivation notes

References

• Austgen et al. (1989) — MEA VLE correlation
• Kvamsdal & Rochelle (2008) — Temperature bulge & NTU modelling
• Nuchitprasittichai & Cremaschi (2011) — Surrogate-based CCS optimisation
• Liao et al. (2022) — ML surrogate for amine scrubbing