Bayesian machine learning for inverse design of ultra-high-performance concrete

Area of Science:

• Materials Science
• Chemical Engineering
• Computational Science

Background:

• Designing ultra-high-performance concrete (UHPC) is complex due to diverse material feedstocks and strict performance demands.
• Machine learning offers potential for optimizing UHPC design, but model parameterization impacts accuracy and generalizability.

Purpose of the Study:

• To demonstrate a Bayesian inverse design method for UHPC using published data.
• To compare the accuracy and generalizability of machine learning models parameterized by compositional versus latent variables.
• To assess the models' ability to predict properties of untested materials and novel compositions.

Main Methods:

• Utilized a Bayesian framework for inverse design of UHPC.
• Employed hierarchical machine learning, including ensemble ridge regression and Gaussian process regression.
• Compared models parameterized by compositional variables against those parameterized by latent variables derived from empirical models.

Main Results:

• Models parameterized by latent variables showed improved generalizability, as indicated by reduced miscalibration area.
• Gaussian process regression unexpectedly yielded higher accuracy for compositional variables (test R² = 0.91) than latent variables (test R² = 0.77) in strength prediction.
• Latent variable models demonstrated superior prediction accuracy for untested fine aggregates and successfully identified novel high-strength UHPC compositions with reduced miscalibration error.

Conclusions:

• Latent variables within a Bayesian machine learning framework enhance generalizability across the variable space for UHPC design.
• Bayesian machine learning models can make robust predictions for untested feedstocks.
• This approach facilitates the prediction of new UHPC compositions with optimal properties, advancing material design.