Prediction and optimization of epoxy adhesive strength from a small dataset through active learning

Area of Science:

• Materials Science
• Polymer Science
• Chemical Engineering

Background:

• Experimental datasets in polymer science are often limited in size (< 100 samples) and expensive to generate.
• Small datasets hinder the application of machine learning (ML) algorithms for extracting crucial chemical insights.
• Developing high-performance functional materials requires efficient methods to overcome data scarcity.

Purpose of the Study:

• To predict and optimize adhesive materials using a data-driven approach despite limited experimental data.
• To demonstrate the efficacy of active learning and Bayesian optimization in accelerating materials discovery.
• To overcome the challenges posed by small datasets in polymer science research.

Main Methods:

• Combined laboratory experimental design with an active learning pipeline and Bayesian optimization.
• Utilized an initial dataset of 32 adhesive samples prepared from bisphenol A-based epoxy resins and polyetheramine curing agents.
• Employed a Gradient Boosting machine learning model for successive prediction of adhesive joint strength within the active learning framework.

Main Results:

• Achieved an optimal adhesive material preparation with a high adhesive joint strength of 35.8 ± 1.1 MPa after three active learning cycles.
• The Gradient Boosting model demonstrated good predictive accuracy, with R-squared, RMSE, and MAE of 0.85, 4.0 MPa, and 3.0 MPa, respectively.
• Successfully identified high-performance adhesive formulations from a very small initial dataset.

Conclusions:

• Active learning significantly accelerates the design and development of tailored functional materials.
• This data-driven methodology is effective for optimizing materials even with highly limited experimental data.
• The study highlights the potential of integrating ML with experimental design for efficient materials discovery in polymer science.