Searcharxiv⌕ Search

arXiv subjects

Marianne Liu

Publications and source records attributed to Marianne Liu.

2 recordsLinked to original sources
arXiv2022

High-dimensional Bayesian Optimization of Hyperparameters for an Attention-based Network to Predict Materials Property: a Case Study on CrabNet using Ax and SAASBO

Expensive-to-train deep learning models can benefit from an optimization of the hyperparameters that determine the model architecture. We optimize 23 hyperparameters of a materials informatics model, Compositionally-Restricted Attention-Based Network (CrabNet), over 100 adaptive design iterations using two models within the Adaptive Experimentation (Ax) Platform. This includes a recently developed Bayesian optimization (BO) algorithm, sparse axis-aligned subspaces Bayesian optimization (SAASBO), which has shown exciting performance on high-dimensional optimization tasks. Using SAASBO to optimize CrabNet hyperparameters, we demonstrate a new state-of-the-art on the experimental band gap regression task within the materials informatics benchmarking platform, Matbench (~4.5% decrease in mean absolute error (MAE) relative to incumbent). Characteristics of the adaptive design scheme as well as feature importances are described for each of the Ax models. SAASBO has great potential to both improve existing surrogate models, as shown in this work, and in future work, to efficiently discover new, high-performing materials in high-dimensional materials science search spaces.

cond-mat.mtrl-sci↗
arXiv2022

Data-Driven Materials Discovery and Synthesis using Machine Learning Methods

Experimentally [1-38] and computationally [39-50] validated machine learning (ML) articles are sorted based on the size of the training data: 1-100, 101-10000, and 10000+ in a comprehensive set summarizing legacy and recent advances in the field. The review emphasizes the interrelated fields of synthesis, characterization, and prediction. Size range 1-100 consists mostly of Bayesian optimization (BO) articles, whereas 101-10000 consists mostly of support vector machine (SVM) articles. The articles often use combinations of ML, feature selection (FS), adaptive design (AD), high-throughput (HiTp) techniques, and domain knowledge to enhance predictive performance and/or model interpretability. Grouping cross-validation (G-CV) techniques curb overly optimistic extrapolative predictive performance. Smaller datasets relying on AD are typically able to identify new materials with desired properties but do so in a constrained design space. In larger datasets, the low-hanging fruit of materials optimization is typically already discovered, and the models are generally less successful at extrapolating to new materials, especially when the model training data favors a particular type of material. The large increase of ML materials science articles that perform experimental or computational validation on the predicted results demonstrates the interpenetration of materials informatics with the materials science discipline and an accelerating materials discovery for real-world applications.

cond-mat.mtrl-sci↗