Rename plots/ to analysis/ and add README with embedded figures

- Rename directory from plots/ to analysis/
- Add README.md with inline PNG figures and narrative text
- Update cache paths in comparison_workflow
- Remove stale file references from model_benchmarks/README.md

Author: bobbypaton

plots/250k_rndm_zinc_drugs_clean_3.csv analysis/250k_rndm_zinc_drugs_clean_3.csvplots/250k_rndm_zinc_drugs_clean_3.csv renamed to analysis/250k_rndm_zinc_drugs_clean_3.csv

@@ -0,0 +1,67 @@
+# Analysis
+
+Scripts and data for visualising the deltahf training set and comparing predicted heats of formation across chemical domains.
+
+---
+
+## Training Data Distribution
+
+**Script:** `plot_training_data.py`
+
+```bash
+python plot_training_data.py
+```
+
+The training data (`deltahf/data/training_data.csv`) contains 313 molecules split into energetic (n=45) and non-energetic (n=268) categories. The histograms below show absolute ΔHf°, ΔHf° per heavy atom, and ΔHf° per atom (including H). The energetic molecules are notably shifted toward positive ΔHf° values, and the per-atom normalisation reveals a clearer separation between the two groups.
+
+![Training data ΔHf° distributions](training_data_histogram.png)
+
+---
+
+## ZINC vs Cawkwell Comparison
+
+**Script:** `plot_zinc_vs_cawkwell.py`
+
+This script compares predicted ΔHf° distributions for two molecule sets:
+
+- **Cawkwell training set** (531 molecules) — the full deltahf training data
+- **ZINC drug-like sample** (1,000 molecules) — randomly sampled from the ZINC 250k drug-like dataset, filtered to supported elements and neutralised
+
+The comparison assesses whether the training set covers the chemical space of typical drug-like molecules and how the predicted ΔHf° distributions differ.
+
+### Usage
+
+```bash
+# Full pipeline: prepare inputs, run predictions, plot
+python plot_zinc_vs_cawkwell.py
+
+# Plot only (if predictions already exist)
+python plot_zinc_vs_cawkwell.py --plot-only
+```
+
+By default the script uses gXTB + `bondorder_ext`. The `comparison_workflow` file contains commands for running the xTB variant separately.
+
+### gXTB predictions
+
+Using gXTB + `bondorder_ext`, the ZINC drug-like molecules tend toward more negative predicted ΔHf° than the training set. When normalised per heavy atom, the distributions overlap more substantially, suggesting the atom equivalent model can extrapolate reasonably to drug-like chemical space.
+
+![gXTB: ZINC vs Cawkwell ΔHf° distributions](zinc_vs_cawkwell_gxtb_histogram.png)
+
+### xTB predictions
+
+The same comparison using xTB + `bondorder_ext` shows a similar pattern. The distributions are broader due to the lower accuracy of xTB energies, but the relative shift between training set and drug-like molecules is consistent.
+
+![xTB: ZINC vs Cawkwell ΔHf° distributions](zinc_vs_cawkwell_xtb_histogram.png)
+
+---
+
+## Data Files
+
+| File | Description |
+|------|-------------|
+| `250k_rndm_zinc_drugs_clean_3.csv` | ZINC 250k drug-like dataset (source data) |
+| `zinc_sample_1000.csv` | 1,000-molecule random sample (neutralised, supported elements only) |
+| `cawkwell_input.csv` | Cawkwell training set formatted for prediction |
+| `cawkwell_gxtb_predictions.csv` | gXTB + bondorder_ext predictions for Cawkwell set |
+| `zinc_gxtb_predictions.csv` | gXTB + bondorder_ext predictions for ZINC sample |
+| `comparison_workflow` | Shell commands for the xTB comparison workflow |

analysis/README.md

@@ -3,13 +3,13 @@ conda activate deltahf
 python -m deltahf predict \
   -i cawkwell_input.csv \
   --model bondorder_ext \
-  --cache-dir plots/.cache_zinc_cawkwell/cawkwell \
+  --cache-dir analysis/.cache_zinc_cawkwell/cawkwell \
   -o cawkwell_xtb_predictions.csv
 
 python -m deltahf predict \
   -i zinc_sample_1000.csv \
   --model bondorder_ext \
-  --cache-dir plots/.cache_zinc_cawkwell/zinc \
+  --cache-dir analysis/.cache_zinc_cawkwell/zinc \
   -o zinc_xtb_predictions.csv
 
 python plot_zinc_vs_cawkwell.py --plot-only