Files in this item



application/pdf5436.pdf (1MB)
(no description provided)PDF


Title:Stacked Ensemble Learning For Range-separation Parameters
Author(s):Ju, Cheng-Wei
Contributor(s):Lin, Zhou; Kohn, Alexander W.; Geva, Nadav; French, Ethan
Subject(s):Theory and Computation
Abstract:\begin{wrapfigure}{l}{0pt} \includegraphics[width=2.2in]{Pic_ISMS.eps} \end{wrapfigure} High-throughput quantum chemical calculations, especially those based on Kohn-Sham density functional theory (KS-DFT), have achieved tremendous success in materials discovery. However, due to the notorious self-interaction error, many popular exchange-correlation (XC) functionals suffer from catastrophic failures for molecules with delocalized electronic densities. As an effective but expensive solution to this problem, the optimally tuned range-separation hybrid functional utilizes a system-specific range-separation parameter ($\omega$). [T. Stein, L. Kronik, and R. Baer, $J.$ $Am.$ $Chem.$ $Soc$. $\textbf{2009}$, 131, 2818.] To accelerate the search for optimal $\omega$ and makes it practical for the high-throughput materials screening, we developed a stacked ensemble learning approach based on composite molecular descriptors, and assessed its accuracy and efficiency using the LRC-$\omega$PBE functional and a diverse database of over 4,000 molecules that are equally divided into the training and test sets. Compared with the traditional optimal tuning method, our stacked ensemble learning algorithm reached a mean absolute error (MAE) of 0.005 $a_0^{-1}$ in the values of $\omega$ while reducing the time cost by four orders of magnitude. In addition, the predictive power in essential properties such as fundamental and optimal band gaps were not compromised. Given sufficient diversity in the training set, stacked ensemble learning provides a promising alternative scheme to determine range-separation hybrid functional and eventually any system-specific XC functionals.
Issue Date:2021-06-23
Publisher:International Symposium on Molecular Spectroscopy
Genre:Conference Paper / Presentation
Date Available in IDEALS:2021-09-24

This item appears in the following Collection(s)

Item Statistics