(681f) Deductive and Inductive Modeling of Electronic Properties in the Organic Molecular Design Space

Socio-economically important researches such as designing single-molecular electronic devices, depend on reliable virtual screening of electronic properties in the chemical space of organic molecules. Clearly, for this task, the only viable solution is to employ statistical property models trained on quantum chemistry datasets. For a given modeling framework, e.g. kernel-ridge-regression with a specified kernel function, the most efficient strategy is to "delta-learn", i.e., to train the model on the prediction errors of fast, but qualitatively accurate, legacy quantum chemistry methods -- rather than directly on outcomes from accurate many-body methods. I will present proof-of-concepts, using non-trivial "Big Datasets", for accurate "delta-learning" of several molecular properties, with hierarchially improved quantum chemistry methods. Latest results on novel electronic properties of isolated molecules and their composite forms will also be discussed.

References

1. Ramakrishnan et al., "Quantum Chemistry Structures and Properties of 134 Kilo Molecules",  Scientific Data 1 (2014).
2. Ramakrishnan et al., "Big Data Meets Quantum Chemistry Approximations: The Delta-Machine Learning Approach", Journal of Chemical Theory and Computation 11 (2015) 2087-2096.
3. Ramakrishnan et al., "Many Molecular Properties from One Kernel in Chemical Space",  Chimia International Journal for Chemistry 69 (2015) 182-186.
4. Ramakrishnan et al., "Electronic Spectra from TDDFT and Machine Learning in Chemical Space", The Journal of Chemical Physics 143 (2015) 084111.
5. Ramakrishnan et al., "Machine Learning, Quantum Mechanics, and Chemical Compound Space", arxiv:1510.07512.