UAE-optimizer

Machine learning–guided optimization of ultrasound-assisted extraction parameters for maximizing bioactive compound yield.

Workflow

1. Use Latin Hypercube Sampling (LHS) to design experiments, ensure representative data coverage. The generation is fully data driven and parameterized through config.json.
   1. List each variable with its [min, max] range in config.json.
   2. A number of total_samples Latin Hypercube samples were generated to cover the parameter space.
   3. Generate all $2^n$ boundary combinations.
   4. Construct a Latin Hypercube Sampling (LHS) design in n dimensions.
   5. A diverse subset with a number of pretest_size points was selected using a MaxMin (greedy space-filling heuristic) algorithm for initial experiments (Set = 1), with the remaining reserved for follow-up (Set = 2).
      1. MaxMin: at each step, selecting the point that maximizes the minimum distance to the already chosen points.
      2. Boundary combinations are always included in Set 1.
   6. Outcome visualized with seaborn.pairplot.
2. Use k-nearest neighbors algorithm (k-NN) regression to model and optimize the parameter space, eg. enzyme concentration, ultrasound temperature, time, and power.(work in progress)

Dependencies

• numpy
• pandas
• pyDOE
• scikit-learn
• matplotlib
• seaborn

pip install -r requirements.txt

Acknowledgements

This project was supported by guidance from Prof. Yinsheng Zhang at Zhejiang Gongshang University, whose feedback greatly improved the analysis. I also gratefully acknowledge Prof. Lina Chen's lab at NJMU for sharing the raw chemical data.