Small datasets, big predictions: learning methods for uncertainty-aware modelling of multi-fidelity material properties
(2024)
Files
DeBreuck_Thesis_v240123_print.pdf
Open Access - Adobe PDF
- 14.94 MB
Details
- Authors
- Supervisors
- Rignanese, Gian-Marco
- Abstract
- Functional materials play a critical role in various technological applic- ations, and the availability of open databases has paved new paths for materials design. In this thesis, I address the challenge of supervised materials design with limited datasets by investigating various learning methods and proposing an all-encompassing framework called MODNet. Traditional approaches often rely on large datasets, which may not be readily available in practice. MODNet overcomes this limitation by combining a feedforward neural network with physically meaningful feature selection and joint learning. This results in faster training and superior performance on small datasets compared to existing graph-network models. MODNet showcases excellent performance on the Matbench test suite and achieves a mean absolute test error of 0.009 meV/K/atom on the vibrational entropy of crystals at 305 K, significantly outperforming previous studies. The thesis also addresses the critical task of uncertainty assessment in material science, utilizing an ensemble MODNet model to build confidence intervals and quantify uncertainty in individual predictions. Furthermore, the potential of active learning is explored by using Bayesian Optimization to iteratively explore metals within the materials space. The significance of considering imbalance and bias in the training set for successful real-world applications of machine learning in materials science is emphasized. Additionally, the thesis investigates techniques that leverage multiple quality sources for the same property, with a focus on the electronic band gap. By learning from differences between high- and low-quality values as a correction, substantial improvements in results are achieved compared to relying solely on high-quality experimental data. This research significantly contributes to advancing machine learning in material property predictions and offers valuable insights into uncertainty assessment and data quality aspects in the field. By providing a comprehensive framework and effectively addressing critical challenges, this research opens new opportunities for accelerating materials design and discovery.
- Affiliations
Citations
De Breuck, P.-P. (2024). Small datasets, big predictions: learning methods for uncertainty-aware modelling of multi-fidelity material properties. https://hdl.handle.net/2078.5/31555