A machine learning approach to model electrodialysis fouling

Electrodialysis removes charged components from liquids

Electrodialysis removes charged components from liquids

Electrodialysis removes charged components from liquids

Some charged components are annoying...

Some charged components are annoying...

Some charged components are annoying...

Some charged components are annoying...

There are many factors that influence fouling

Foulant props.

Salt concentration

Crossflow velocity

Current

pH

Membrane props.

Modelling particle interactions is expensive

$ F_{p,i} = \sum{F_{p,j}} + F_{mem} + F_d \\ + F_{am} + F_h + F_l + F_e + ... $

→ need for an efficient description!

Different models have different purposes

Neural Differential Equations

R. Chen et al. 2018

C. Rackauckas et al. 2019

Neural Differential Equations

R. Chen et al. 2018

C. Rackauckas et al. 2019

Neural Differential Equations

R. Chen et al. 2018

C. Rackauckas et al. 2019

Neural Differential Equations

R. Chen et al. 2018

C. Rackauckas et al. 2019

Neural Differential Equations

R. Chen et al. 2018

C. Rackauckas et al. 2019

It is a black-box differential equation

It is a black-box differential equation

Let us test this on a simple example...

Let's perform some experiments

Foulant props.

Salt concentration

Crossflow velocity

Current

pH

Membrane props.

Let's perform some experiments

Foulant props.

Salt concentration

Crossflow velocity

Current

pH

Membrane props.

22 timeseries at different process conditions

Some examples of experiments

Model training results

Model test results

A continuous representation of the fouling rate

Extrapolation in time works well

Summary

1. Electrodialysis fouling is a complex process
2. Fouling dynamics can be modelled with neural ODEs
3. Neural ODEs can be predictive even for a limited dataset

Perspectives

• Integration of mechanistic knowledge (mixed neural ODEs)

Summary

1. Electrodialysis fouling is a complex process
2. Fouling dynamics can be modelled with neural ODEs
3. Neural ODEs can be predictive even for a limited dataset

Perspectives

• Integration of mechanistic knowledge (mixed neural ODEs)

Summary

1. Electrodialysis fouling is a complex process
2. Fouling dynamics can be modelled with neural ODEs
3. Neural ODEs can be predictive even for a limited dataset

Perspectives

• Integration of mechanistic knowledge (mixed neural ODEs)

Summary

1. Electrodialysis fouling is a complex process
2. Fouling dynamics can be modelled with neural ODEs
3. Neural ODEs can be predictive even for a limited dataset

Perspectives

• Integration of mechanistic knowledge (mixed neural ODEs)

Summary

1. Electrodialysis fouling is a complex process
2. Fouling dynamics can be modelled with neural ODEs
3. Neural ODEs can be predictive even for a limited dataset

Perspectives

• Integration of mechanistic knowledge (mixed neural ODEs)