Year: 2026
Type: dissertation
Abstract: Accurate prediction of amine pKa is central to carbon-capture solvent design, where it governs the trade-off between fast CO2 absorption and low-energy solvent regeneration. This thesis develops machine-learning pKa predictors in a three-part progression. First, five standard architectures are benchmarked across five molecular representations on curated ChEMBL and IUPAC amine datasets; the best, XGBoost on Benson-group features, attains Test R2 0.92 and MAE 0.60. Second, a three-stage transfer-learning network trained on the purpose-built pKahub corpus over FairChem uma-s-1p2 embeddings attains MAE 0.27 on an 86-amine carbon-capture (CCUS) hold-out under a strict leak-free protocol, while remaining a domain predictor rather than a universal one (Novartis MAE 1.54, R2 ≈ 0.1). Finally, a physics-informed head coupling a linear free-energy anchor, a van’t Hoff temperature law, and multiprotic ordering, refined by evolutionary architecture search, attains MAE 0.25 on the same hold-out and 0.88 out-of-domain.
Author: Aslam, Kamal
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
SDG: Climate action
Open Access status: green
OpenAlex ID: W7213478551