Nonlinear dimensionality reduction and Bayesian optimization for accelerating design of materials.

Area of Science:

• Materials Science
• Chemical Engineering
• Data Science

Background:

• Optimizing biobased foam formulations is complex due to costly experiments and high-dimensional property spaces.
• Bayesian optimization (BO) with Gaussian process regression (GPR) offers data-efficient experimental guidance but is sensitive to input dimensionality.

Purpose of the Study:

• To evaluate nonlinear dimensionality reduction (DR) methods (t-SNE, UMAP) against PCA for optimizing biobased foam formulations.
• To assess the effectiveness of DR-assisted BO in identifying high-yield-stress regions and reconstructing optimal formulations.

Main Methods:

• Utilized a dataset of 26 methylcellulose-cellulose fiber foam formulations with rheological and mechanical data.
• Trained Gaussian processes (GPs) on low-dimensional representations from PCA, t-SNE, and UMAP.
• Applied Bayesian optimization (BO) to latent representations and mapped compositions to reduced coordinates for formulation reconstruction.

Main Results:

• All DR methods, when integrated with BO, identified high-performing foam formulations with yield stress comparable to the experimental optimum.
• Principal Component Analysis (PCA) served as a baseline, while nonlinear methods (t-SNE, UMAP) showed comparable performance when tuned.
• Nonlinear DR-assisted BO proved effective in navigating the complex formulation space.

Conclusions:

• Nonlinear dimensionality reduction combined with Bayesian optimization provides a data-efficient framework for optimizing soft-matter materials.
• This approach facilitates the exploration of complex formulation spaces, guiding the discovery of materials with desired rheological properties.