the greyness of machine learning based hybrid models for bioprocess predictive
modelling and uncertainty estimation
April at 3pm (UK time)
The interaction between
the complex underlying cell metabolism and multiphase environment, leads biochemical
process to often exhibit high batch-to-batch variation. This poses a challenge
to the development of mechanistic kinetic models that provide both accurate
state predictions and reliable uncertainty estimates. To tackle this, a hybrid
modelling methodology was adopted in this work, whereby a probabilistic Gaussian
Process (GP) machine learning model was embedded into the mechanistic kinetic
model structure to simulate time-dependent bioprocess dynamics.
Two case studies
were conducted to explore the practical challenges associated with building a
hybrid bioprocess model. First, the hybrid modelling methodology was applied to
model the fermentation of the yeast X. dendrorhous, where the
impact of different hyperparameters on the trade-off between predictive accuracy
and uncertainty was investigated. Then the methodology was applied to model the
fermentation of the fungus Cunninghamella echinulate, where the effect
of inductive bias by incorporating different amounts of physical knowledge (i.e.,
hybrid model “greyness”) on overfitting and overparameterisation was
systematically explored. Compared to benchmark kinetic models, both hybrid
models exhibited good predictive accuracy and superior predictive confidence,
demonstrating the suitability of hybrid modelling for robust process
optimisation, control and online monitoring.
This webinar will last no
longer than one hour.
The webinar will be
presented by Alexander Rogers, University of Manchester.
To register please contact