Journal
COMPUTATIONAL MATERIALS SCIENCE
Volume 172, Issue -, Pages -Publisher
ELSEVIER
DOI: 10.1016/j.commatsci.2019.109286
Keywords
Density functional theory; Machine learning; Multi-fidelity learning; Polymers; Bandgap
Categories
Funding
- Toyota Research Institute through the Accelerated Materials Design and Discovery program
Ask authors/readers for more resources
The fidelity of data is of paramount importance in the construction of reliable and accurate machine learning (ML) models. Low-fidelity data, although noisy, can usually be obtained for a large number of material samples. High-fidelity data, on the other hand, is time-consuming and oftentimes, only available for a limited number of target samples. While the former can provide useful information to help generalize the ML models over large materials space, the latter is useful to build more accurate surrogate models. Information fusion schemes that utilize the data available at multiple levels of fidelity can outperform traditional single fidelity based ML methods, such as Gaussian process regression. In this work, a variant of the multi-fidelity information fusion scheme, namely multi-fidelity co-kriging, is used to build powerful prediction models of polymer bandgaps. To benchmark this strategy, we utilize a bandgap dataset of 382 polymers, obtained at two levels of fidelity: using the Perdew-Burke-Ernzerhof (PBE) exchange-correlational functional (low-fidelity) and the Heyd-Scuseria-Ernzerhof (HSE06) functional(high-fidelity) of density functional theory. The multi-fidelity model, trained on both PBE and HSE06 data, outperforms a single-fidelity Gaussian process regression model trained on just HSE06 band-gaps in a number of scenarios and is also able to generalize better to a more diverse chemical space.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available