Developing high-dimensional machine learning models to improve generalization ability and overcome data insufficiency...

Area of Science:

• Biotechnology
• Machine Learning
• Biorefinery Processes

Background:

• Accurate machine learning models are crucial for advancing biorefinery efficiency.
• Simulating mixed sugar fermentation presents challenges due to data limitations and model generalization.

Purpose of the Study:

• To develop a strategy for enhancing machine learning model generalization and overcoming data insufficiency in mixed sugar fermentation simulation.
• To improve the accuracy and applicability of artificial neural network (ANN) models in biorefinery contexts.

Main Methods:

• Utilized multiple inputs single output (MISO) and multiple inputs multiple outputs (MIMO) models for fermentation simulation.
• Developed a consensus yeast (CY) model by consolidating data from multiple yeast strains.
• Adjusted pretrained CY models to assess performance with reduced datasets.

Main Results:

• MISO models with combined initial glucose, xylose, and time inputs showed superior generalization compared to time-only inputs.
• MIMO models achieved high accuracy with an average R² of 0.99.
• The CY model achieved an R² of 0.90, and adjusted models retained high accuracy (R² 0.95 for yeast, 0.93 for bacterial) with over 50% less data.

Conclusions:

• The proposed strategy significantly enhances model generalization and addresses data scarcity in fermentation simulations.
• This approach expands the applicability of ANN models in biorefineries and reduces data curation costs.