AutoML based workflow for design of experiments (DOE) selection and benchmarking data acquisition strategies with...

Area of Science:

• Experimental Design
• Machine Learning

Background:

• Design of Experiments (DOE) is crucial for efficient parameter space exploration.
• Model-based Active Learning (AL) offers potential for optimizing data sampling strategies.

Purpose of the Study:

• Introduce a workflow for comparative DOE studies using automated machine learning.
• Examine the interplay between systematic data generation and model performance under uncertainty.

Main Methods:

• Developed a workflow integrating DOE and automated machine learning.
• Defined model complexity practically for machine learning contexts.
• Investigated uncertainties from sampling, data precision, and modeling.

Main Results:

• Not all AL strategies outperform conventional DOE; performance depends on data volume, complexity, and uncertainty.
• Replication-oriented strategies are advantageous for non-negligible noise and intermediate resources.
• Systematically analyzed trade-offs between data replication and broad sampling.

Conclusions:

• The proposed workflow simulates practical DOE testing and selection conditions.
• AL strategy effectiveness is context-dependent, not universally superior to DOE.
• Balancing data replication and exploration is key for resource allocation in ML-driven DOE.