CFD · Machine Learning · Membrane Optimization · KU Leuven

Vanadium Redox Flow Battery
Simulation & ML Optimization

2D multiphysics CFD modeling, multi-ion membrane transport, and ML-driven membrane design optimization for grid-scale energy storage.

01

Overview

Vanadium redox flow batteries (VRFBs) are a leading technology for long-duration grid storage — but their performance hinges critically on membrane properties. This project builds a complete physics-to-optimization pipeline: a 2D COMSOL CFD model with multi-ion membrane transport, validated against experimental charge–discharge data, then coupled to machine learning surrogates for membrane design optimization.

A database of 70 membranes across three types (CEM, AEM, AIEM) was analyzed using Mutual Information, Random Forest, and Gradient Boosting Regressor. GBR was selected as the NSGA-II surrogate based on LOO-CV evidence, targeting simultaneous maximization of Coulombic, Voltage, and Energy efficiency while constraining vanadium crossover.

70Membranes
2DCFD Model
GBRSurrogate
NSGA-IIOptimizer
VRFB Architecture
Anolyte V²⁺/V³⁺ Catholyte VO²⁺/VO₂⁺ Anode electrode Membrane Cathode electrode Stack ⟳ ⟳ e⁻ flow (external circuit) Pump Pump
Anolyte Catholyte Membrane
02

Modeling & Optimization Workflow

01
COMSOL · CFD · Multi-ion

2D Multiphysics CFD Model

Physics-based 2D model couples Navier–Stokes flow, convection–diffusion species transport, Nernst–Planck multi-ion membrane transport, and Butler–Volmer electrochemistry — validated against experimental charge–discharge curves.

  • Model reduction: 3D short-stack → 2D domain — reduces cost while preserving ion-exchange membrane (IEM) physics impossible in 3D
  • Isothermal, incompressible electrolyte; 60% pump efficiency; full boundary condition set at ribs, membrane, and outlet
  • Performance metrics: CE, VE, EE, and net power accounting for pump overhead
02
Random Forest · GBR · LOO-CV

ML Surrogate Selection

A database of 70 membranes (CEM, AEM, AIEM) was analyzed. Mutual Information identified IEC as the dominant driver of CE, VE, and EE. Random Forest and Gradient Boosting Regressors were compared via Leave-One-Out Cross-Validation on the small dataset.

  • Non-linear signals detected: Spearman vs Pearson divergence ruled out linear models
  • Both RF and GBR split first on IEC — validating the physics-driven feature signal
  • GBR selected: tighter LOO-CV residuals, higher explained variance, cleaner error distribution — optimal for embedding in NSGA-II
03
NSGA-II · Pareto · LHS

Multi-Objective Membrane Optimization

GBR surrogate feeds NSGA-II multi-objective optimizer targeting the CE–VE–EE Pareto front with vanadium permeability (Dr) as a hard constraint. Latin Hypercube Sampling (LHS) seeds the design space for COMSOL parametric studies.

  • Three design levers: IEC (primary, 34–37% importance), thickness (resistance control, 43–47%), water uptake (bounded window 20–50%)
  • Benchmark target: SPEEK/ZC-GO-2 profile — EE 91.4%, CE 98.5%, VE 92.3%, Dr 0.189
  • Design rule: maximize EE as objective, use Dr as constraint — not CE-only screening
03

Key Results

CFD model validated against experimental charge–discharge data 2D COMSOL model with multi-ion Nernst–Planck membrane transport reproduces experimental CE, VE, and EE profiles across flow rates and current densities.
IEC identified as dominant membrane performance driver Mutual Information analysis across 70 membranes ranked IEC highest for CE, VE, and EE. Both RF and GBR tree splits confirmed the physics — IEC anchors the design signal.
GBR selected as NSGA-II surrogate via LOO-CV evidence Six-metric LOO-CV comparison favoured GBR: tighter absolute-error distribution, higher explained variance, unbiased residuals — providing the strongest out-of-sample generalisation in the small-data setting.
Design playbook derived from 70-membrane dataset Water uptake 20–50%, IEC tuned not maximised, filler-driven tortuosity (GO/ZrO₂), and EE as objective with Dr as constraint — four gates that convert correlation evidence into actionable membrane choices.
LHS-seeded COMSOL parametric pipeline automated in Python Latin Hypercube Sampling generates the CFD design space; Python drives COMSOL via API for automated simulation, result extraction, and surrogate training dataset construction.
04

Skills Demonstrated

Multiphysics CFDCOMSOL · Nernst–Planck · Butler–Volmer
ML SurrogateGBR · Random Forest · LOO-CV
Multi-objective Opt.NSGA-II · Pareto · LHS
Feature AnalysisMutual Information · Pearson/Spearman
Python AutomationJupyter · COMSOL API · Scikit-Learn
ElectrochemistryIon transport · Crossover · IEC
Materials Database70 membranes · CEM / AEM / AIEM
Model Reduction3D → 2D · Boundary conditions
05

Collaboration

Open to collaboration in flow battery simulation, ML-driven materials optimization, and electrochemical systems modeling.

Get in Touch