More than mimicry? The challenges of teaching chemistry to deep models
Brett Savoie, Purdue University
Deep generative chemical models refer to a family of machine learning architectures that can digest chemical property data and suggest new molecules or materials with targeted properties. These models have generated interest for their potential to predict chemistries based on structure-function design rules learned directly from data. Nevertheless, these models are extremely data hungry and have several common failure mechanisms. In this talk I will summarize contemporary strategies for training these models, discuss where progress has been made, and provide some opinions on where work is still needed.
a) Iovanac, N. C.; Savoie, B. M. "Simpler Is Better: How Linear Prediction Tasks Improve Transfer Learning in Chemical Autoencoders." The Journal of Physical Chemistry A 2020, 124 (18), 3679–3685. https://doi.org/10.1021/acs.jpca.0c00042 [doi.org].
b) Iovanac, N. C.; Savoie, B. M. "Improving the Generative Performance of Chemical Autoencoders through Transfer Learning." Mach. Learn.: Sci. Technol. 2020, 1 (4), 045010. https://doi.org/10.1088/2632-2153/abae75 [doi.org].
c) Iovanac, N. C.; MacKnight, R.; Savoie, B. M. "Actively Searching: Inverse Design of Novel Molecules with Simultaneously Optimized Properties” ChemRxiv, https://doi.org/10.26434/chemrxiv.14643360.v1 [doi.org]