NAGL2 Approach 1: Initial plan
- Alexandra McIsaac
Initial approach as suggested stakeholder consensus.
Overview
Summary |
|
|---|---|
GitHub link |
|
Status | Not started In progress Completed Won't progress |
Milestones and metrics
Stage | Milestone/Benchmark | Contributors | Deadline | Status |
|---|---|---|---|---|
Curate training dataset | Work out best level of theory for the training dataset | @Alexandra McIsaac , @Lily Wang | November 10, 2024 | In progress |
| Compute training dataset | @Alexandra McIsaac | December 31, 2024 | Not started |
Curate testing dataset | Compile QM dataset |
| November 30, 2024 | Not started |
| Compute QM dataset |
| January 31, 2025 | Not started |
| Compile simulation test set (Free Solv, maybe non-hydration solvation free energy sets that are harder to reproduce) |
| April 15, 2025 | Not started |
Determine best NN architecture | Implement attention-based GNN | @Alexandra McIsaac @Brent Westbrook (Unlicensed) @Lily Wang | December 31, 2024 | Not started |
| Implement bond features in GraphSAGE (?) | @Alexandra McIsaac @Brent Westbrook (Unlicensed) @Lily Wang | December 31, 2024 | Not started |
| Determine best architecture | @Alexandra McIsaac @Brent Westbrook (Unlicensed) @Lily Wang | January 31, 2025 | Not started |
First pass at NN training | Train using just ESPs, dipoles, quadrupoles |
| Feb 28, 2025 | Not started |
| Regularize to RESP charges or MBIS charges if buried atoms are a problem |
|
| Not started |
| Train directly to charge model if still having issues |
|
| Not started |
Benchmark 1: QM | Neural network charge model with low testing error on QM data (ESPs, dipoles) |
| March 15, 2025 |
|
Re-train VDW terms |
|
| March 30, 2025 | Not started |
Re-train valence terms |
|
| April 15, 2025 | Not started |
Benchmark 2: Simulation | Neural network charge model with equivalent or better performance to NAGL in simulations |
| April 30, 2025 |
|
Progress and findings
Curated data (or similar title)