Analyzing drop coalescence in microfluidic devices with a deep learning generative model.

Area of Science:

• Chemical Engineering
• Machine Learning
• Data Science

Background:

• Predictive modeling in chemical engineering is vital for experimental design.
• Challenges include limited training data and significant label imbalance issues.
• Existing models struggle with accurate predictions due to these data limitations.

Purpose of the Study:

• To develop a deep learning generative model for creating synthetic data to overcome data scarcity and imbalance.
• To enhance the performance of predictive models in chemical engineering using generated data.
• To introduce a novel generative model, the double space conditional variational autoencoder (DSCVAE), for labeled tabular data.

Main Methods:

• Development of the double space conditional variational autoencoder (DSCVAE) for labeled tabular data generation.
• Incorporation of label constraints in both latent and original spaces within DSCVAE.
• Training and evaluation of random forest and gradient boosting classifiers using DSCVAE-generated synthetic data.

Main Results:

• The DSCVAE model generates consistent and realistic synthetic samples, outperforming the standard conditional variational autoencoder (CVAE).
• Predictive models enhanced with DSCVAE-generated synthetic data show considerable improvements in prediction accuracy.
• Evaluation on real experimental data confirms the effectiveness of the proposed synthetic data generation approach.

Conclusions:

• Deep learning generative models, particularly DSCVAE, offer a powerful solution for handling imbalanced data in chemical engineering classification problems.
• The use of synthetic data generated by DSCVAE significantly improves the accuracy of predictive models.
• This research provides valuable insights into data augmentation strategies for complex engineering applications.