First Ascent by
Ayush Tiwary
The Perovskite Stability Ascent
Materials Science / Solar Energy
The Proposed Route
A high-altitude traverse across the perovskite composition landscape. The climber maps 43,000+ devices across a 410-feature face, learning which chemical handholds (composition, fabrication) lead to stable footing at the summit (long-term device stability). Temporal splits keep the ascent honest.
🧗 The Crux
Sparse stability labels (only 17% coverage) make the upper section of the route thin and runout. Composition parsing of complex ABX3 formulas is the technical crux — if it fails, the whole rack is useless. Temporal train/test split adds route-finding difficulty.
⚠️ Pre-Climb Checklist
✅ Dataset verified locally (83MB). ✅ Colab feasibility tested. ⚠️ Descope multi-task learning if time-limited — complete Exercises 0–5 first. ⚠️ Validate composition parser on top-10 formulas before committing to full dataset.
Guidance
- Temporal split is the right call — shows methodological maturity
- SHAP interpretability validates against known perovskite chemistry
- Connect SHAP results to physical mechanisms (Cs stabilization, halide phase segregation)
- That connection elevates this from model exercise to research contribution
Source proposal: AyushTiwary_CHEM169_269_FinalProjectProposal.pdf
CHEM 169/269 · Applied AI & Machine Learning for Biochemistry