Towards improving chemical synthesis planning assistants
- Andronov, Mikhail
- Schmidhuber, Jürgen (Degree supervisor)
- Wand, Michael (Degree committee member)
- 2026
PhD: Università della Svizzera italiana
English
In this thesis, we explore a set of techniques to improve existing deep learning–based automated chemical synthesis planning systems. First, we develop and analyze methods to reduce the latency of transformer-based synthesis planning systems. We accelerate transformer-based single-step retrosynthesis and reaction prediction models using speculative decoding and a novel speculative beam search algorithm, achieving substantial speedups without loss of accuracy and demonstrating their impact on multi-step synthesis planning. Secondly, we study reagent prediction as an integral but underexplored component of synthesis planning. We formulate it as a sequence-to-sequence learning problem and show that accurate reagent modeling can both enhance reaction prediction and mitigate missing information in reaction datasets. We then propose a self-supervised approach to improve the quality of reagent information in reaction data by grouping reagents by functional role, implemented in an interactive web application for data preparation. Together, our contributions advance the efficiency, completeness, and reliability of AI-driven synthesis planning and provide a foundation for developing more practical and trustworthy next-generation CASP systems.
- Collections
- Language
-
- English
- Classification
- Computer science and technology
- License
-
License undefined
- Open access status
- green
- Identifiers
-
- NDP-USI 2026INF006
- ARK ark:/12658/srd1336431
- URN urn:nbn:ch:rero-006-124146
- Persistent URL
- https://n2t.net/ark:/12658/srd1336431
Statistics
Document views: 2
File downloads:
- 2026INF006.pdf: 0