Switch TPD to Summary-ECFlow model (21-dim hand-crafted features)
Bing Yan
ad63797
- Replace v8 full-signal TPD model with Summary-ECFlow ablation checkpoint
(better classification: LH_Surface 99.5% vs 78.1%, all 6 mechs correct)
- Add extract_tpd_summary_stats() to preprocessing.py for feature extraction
- Update inference.py to auto-detect use_summary_features from checkpoint args
and compute/pass summary features during prediction
- Regenerate FirstOrder TPD demo data (sample_000090) for better compatibility
Made-with: Cursor
Switch to v9 EC + v8 TPD noise-augmented models for robustness
Bing Yan
0da091b
EC: v9 noise-augmented (93% clean, 87% noisy accuracy) — best balance
of precision and robustness for real user data. Supports all 6 mechanisms.
TPD: v8 noise-augmented (93% clean, 92.5% noisy accuracy) — best
noise-robust classification across all 6 TPD mechanisms.
Also regenerated FirstOrder TPD demo data (previous sample was
confusable with LH_Surface across all model versions).
Made-with: Cursor
Switch to v7 model checkpoints to support all 6 EC mechanisms including Nernst
Bing Yan
41896af
v8 model was not trained on Nernst data. Reverting to v7 which supports
BV, MHC, Ads, EC, LH, and Nernst. Added Nernst demo data (96.2% confidence).
Made-with: Cursor
Fix CSV parser column detection and regenerate LH demo data
Bing Yan
c2d475b
1. Fix parse_cv_csv: the generic pattern ("i", 1.0) was matching
"potential" as a substring, causing both e_col and i_col to point
to column 0. Now tracks assigned columns to prevent double-matching.
2. Regenerate LH demo data: the previous LH sample had asymmetric
peaks that caused estimate_E0 to be off by 30mV, shifting theta
enough to confuse the classifier. New sample (sample_125030) has
<3mV E0 estimation error and classifies as LH with 99.7% confidence.
3. Remove Nernst demo data: Nernst mechanism was not included in the
v8 training dataset (0 samples), so the model cannot classify it.
All 5 EC mechanisms now classify correctly:
BV 60%, MHC 99.3%, Ads 100%, EC 100%, LH 99.7%
Made-with: Cursor
Initial ECFlow deployment
Bing Yan
d6b782e
Amortized Bayesian inference for mechanism identification and parameter
estimation in electrochemistry (CV) and catalysis (TPD).
- v7 targeted calibration checkpoints (EC + TPD)
- Gradio web UI with CSV and image upload
- Demo data for all 12 mechanisms
Made-with: Cursor
