Meta optimization based on real-time benchmarking of multiple surrogate models for autonomous flow synthesis

Area of Science:

• Organic Synthesis
• Chemical Engineering
• Machine Learning

Background:

• Optimizing microflow-based organic synthesis is complex.
• Bayesian optimization (BO) guides flow reactor conditions but model selection is inefficient.

Purpose of the Study:

• To introduce meta optimization (MO) for real-time surrogate model benchmarking.
• To enable efficient optimization of reaction conditions in microflow synthesis.

Main Methods:

• MO benchmarks multiple surrogate models in real-time without pre-work.
• Evaluates expected values from regressors to build surrogate models.

Main Results:

• MO consistently outperformed various Bayesian optimization (BO) methods across four flow synthesis datasets.
• Conventional BOs showed variable performance depending on the surrogate model (e.g., Gaussian process, random forest).

Conclusions:

• Meta optimization (MO) provides a superior and efficient method for optimizing microflow reaction conditions.
• Real-time benchmarking is crucial for selecting optimal surrogate models in BO.