| dc.description.abstract | Bond Dissociation Energy (BDE) is one of the most imperative properties of molecular bonds
and the prerequisite for understanding the kinetic mechanism and drug analysis. The transition
from the experimental approach to the computational-derived theory, and Machine Learning
(ML) approach has demonstrated great potential to predict reaction properties and accelerate
chemical space analysis in today's world. The available solutions made by computationalderived or graph neural networks gain promising results and insights, but most of them are
resource-intensive and demand a large amount of training data. In this thesis, we apply a new
Machine Learning solution called AIP-BDET to solve this Chemistry problem with high speed,
accuracy, reliability, and scalability on low-end GPU hardware.
Our proposed model focused on the interpretation of reliability metric whereas consistent
prediction is made under the comparison with DFT-BDE and Exp-BDE and delivered
incredible speed with only 20 minutes of model training using the NVIDIA Quadro P1000, and
single-digit millisecond latency. The AIP-BDET model achieved comparable performance
against advanced graph networks with a minor performance loss at around 0.1
MAE on DFT-BDE reference overall, but some top-3 single models outperformed them on
several metrics. The code to reproduce can be found on GitHub: | en_US |
