sync code from github@d2dbab26ef57a01a14fbe811b084a3716af12e52

README.md

@@ -26,10 +26,10 @@ MagNET is a family of neural networks for predicting NMR chemical shifts. This r
 
 | Model | Description | Details |
 |---|---|---|
-| **MagNET** | foundation model | - for near-equilibrium geometries <br /> - trained on PBE0/pcSseg-1/gas shieldings |
-| **MagNET-Zero** | high-quality gas-phase solute shieldings | - use AIMNet2-optimized geometries <br /> - WP04/pcSseg-2 (<sup>1</sup>H shieldings) <br /> - ωB97X-D/pcSseg-2 (<sup>13</sup>C shieldings) |
-| **MagNET-PCM** | implicit solvent corrections | - shielding(PCM) - shielding(gas) <br />- use AIMNet2-optimized geometries <br /> - computed at B3LYP/pcSseg-2/chloroform |
-| **MagNET-x** | explicit solvent corrections | - shielding(solute+solvent) - shielding(solute) <br /> - supports chloroform, benzene, methanol, and water <br /> - use classical MD geometries <br /> - trained at PBE0/pcSseg-1 |
+| **MagNET** | foundation model | <ul><li>for near-equilibrium geometries</li><li>trained on PBE0/pcSseg-1/gas shieldings</li></ul> |
+| **MagNET-Zero** | high-quality gas-phase solute shieldings | <ul><li>use AIMNet2-optimized geometries</li><li>WP04/pcSseg-2 (<sup>1</sup>H shieldings)</li><li>ωB97X-D/pcSseg-2 (<sup>13</sup>C shieldings)</li></ul> |
+| **MagNET-PCM** | implicit solvent corrections | <ul><li>shielding(PCM) - shielding(gas)</li><li>use AIMNet2-optimized geometries</li><li>computed at B3LYP-D3(BJ)/pcSseg-2/chloroform</li></ul> |
+| **MagNET-x** | explicit solvent corrections | <ul><li>shielding(solute+solvent) - shielding(solute)</li><li>supports chloroform, benzene, methanol, and water</li><li>use classical MD geometries</li><li>trained at PBE0/pcSseg-1</li></ul> |
 
 All models use Equiformer-V2 and have approximately 10M weights. Separate weights are given for <sup>1</sup>H and <sup>13</sup>C prediction. Input structures can contain H, C, N, O, F, Cl, and S (inference should not be performed on structures with unsupported elements).
 
@@ -37,11 +37,11 @@ All models use Equiformer-V2 and have approximately 10M weights. Separate weight
 
 | Dataset | Details |
 |---|---|
-| **_sigma_-shake** | - 4.4M solutes from GDB-13/17 with functional group augmentation <br /> - stationary and quasiclassically perturbed structures at B3LYP-D3(BJ)/6-31G\* <br /> - PBE0/pcSseg-1/gas shieldings |
-| **_sigma_-fresh** | - 10K representative natural product, drug-like, sugar, and peptide solutes dissolved in benzene, chloroform, methanol, and water <br /> - ~10 computed poses/solute with many more solvated geometries available <br /> - 462K poses have computed PBE0/pcSseg-1 shieldings |
-| **_sigma_-pepper** (part 1) | - GDB molecules with ≤ 10 heavy atoms <br /> - AIMNet2 stationary structures <br /> - PBE0/pcSseg-1 shieldings |
-| **_sigma_-pepper** (part 2) | - GDB molecules with ≤ 9 heavy atoms <br /> - AIMNet2 stationary structures <br /> - WP04/pcSseg-2/gas (<sup>1</sup>H shieldings) <br /> - ωB97X-D/pcSseg-2/gas (<sup>13</sup>C shieldings) |
-| **_sigma_-concentrate** | - 50K random structures from *sigma*-shake <br /> - PCM(chloroform) corrections at B3LYP/pcSseg-2 |
+| **_sigma_-shake** | <ul><li>4.4M solutes from GDB-13/17 with functional group augmentation</li><li>stationary and quasiclassically perturbed structures at B3LYP-D3(BJ)/6-31G\*</li><li>PBE0/pcSseg-1/gas shieldings</li></ul> |
+| **_sigma_-fresh** | <ul><li>10K representative natural product, drug-like, sugar, and peptide solutes dissolved in benzene, chloroform, methanol, and water</li><li>~10 computed poses/solute with many more solvated geometries available</li><li>462K poses have computed PBE0/pcSseg-1 shieldings</li></ul> |
+| **_sigma_-pepper** (part 1) | <ul><li>GDB molecules with ≤ 10 heavy atoms</li><li>AIMNet2 stationary structures</li><li>PBE0/pcSseg-1 shieldings</li></ul> |
+| **_sigma_-pepper** (part 2) | <ul><li>GDB molecules with ≤ 9 heavy atoms</li><li>AIMNet2 stationary structures</li><li>WP04/pcSseg-2/gas (<sup>1</sup>H shieldings)</li><li>ωB97X-D/pcSseg-2/gas (<sup>13</sup>C shieldings)</li></ul> |
+| **_sigma_-concentrate** | <ul><li>50K random structures from *sigma*-shake</li><li>PCM(chloroform) corrections at B3LYP/pcSseg-2</li></ul> |
 
 ### Installing MagNET
 
@@ -128,8 +128,7 @@ for name, indices in sites.items():
     print(f"{name:<13} {shifts[indices].mean():6.2f} ppm")
 ```
 
-The predictions land on the measured shifts (experiment is acetone in chloroform, from the delta22
-dataset):
+The predictions are close to experiment:
 
 | site | MagNET | experiment (CDCl<sub>3</sub>) |
 |---|---|---|
@@ -140,7 +139,7 @@ dataset):
 **Notes:**
 
 - **Passes and symmetry.** `n_passes` (default 10) averages out equivariance error over the specified number of forward passes. If `symmetrize` (default True) is set, an additional `n_passes` are also performed on the mirror image of the input geometry (for a total of `2*n_passes`).
-- **Geometry.** MagNET-Zero and MagNET-PCM require AIMNet2-optimized geometries. Do not use other geometries.
+- **Geometry.** MagNET-Zero and MagNET-PCM require [AIMNet2](https://github.com/isayevlab/AIMNet2)-optimized geometries. Do not use other geometries.
 - **Components.** Pass `return_components=True` to also get the MagNET-Zero shielding, the MagNET-PCM correction, and the scaling coefficients behind each shift, as a `dict`.
 - **Other models.** For raw shieldings and solvent corrections, see the [API documentation](#api-documentation).
   
@@ -155,16 +154,16 @@ If you only want to download a single dataset, add
 
 | dataset | size | contents |
 |---|---|---|
-| `sigma-shake` | 2.2 GB | main training set: 4.4M GDB solutes, perturbed geometries, PBE0/pcSseg-1 shieldings |
-| `sigma-fresh` | 30 GB | explicit-solvent poses: 10K solutes in benzene, chloroform, methanol, and water |
+| `sigma-shake` | 2.2 GB | 4.4M GDB solutes, perturbed geometries, PBE0/pcSseg-1 shieldings |
+| `sigma-fresh` | 30 GB | 10K solutes in explicit benzene, chloroform, methanol, and water |
 | `sigma-pepper` | 501 MB | AIMNet2 GDB structures used to train MagNET-Zero and MagNET-PCM |
 | `sigma-concentrate` | 17 MB | PCM corrections on 50K *sigma*-shake structures |
 | `delta22` | 2.5 GB | experimental <sup>1</sup>H/<sup>13</sup>C shifts with matched DFT and MagNET predictions |
-| `dft8k` | 12 MB | external benchmark of 7111 organics (Guan and Paton), two levels of theory |
-| `gdb_qcd` | 291 MB | quasiclassical-dynamics benchmark of 2461 GDB molecules, for rovibrational corrections |
-| `magnet_test_predictions` | 822 MB | MagNET's raw per-atom predictions and DFT targets on the test sets |
-| `applications` | 5.1 GB | natural-product evaluation set, with MD geometries and experimental shifts |
-| `supertestset_magnet_x` | 1.2 MB | explicit-solvent test set for MagNET-x: isolated vs solvated DFT shieldings |
+| `dft8k` | 12 MB | external benchmark of 7111 organics |
+| `gdb_qcd` | 291 MB | quasiclassical dynamics of 2461 GDB molecules |
+| `magnet_test_predictions` | 822 MB | MagNET predictions and DFT targets |
+| `applications` | 5.1 GB | natural product MD geometries and experimental shifts |
+| `supertestset_magnet_x` | 1.2 MB | explicit solvent test set for MagNET-x |
 
 ### Reproducing Figures and Tables
 
@@ -186,7 +185,7 @@ pip install -r requirements.txt "pytest>=7"
 pytest
 ```
 
-The tests use small synthetic fixtures and do not need the downloaded files. Any test that does require the large model weights or datasets is skipped automatically when they are absent, and runs once they have been downloaded.
+The tests use small synthetic fixtures. If the large files are present, then more comprehensive tests will run. The `magnet` package tests need the Option B checkout with the inference stack installed (`pip install -r magnet/requirements.txt`); without PyTorch they are skipped.
 
 ### API Documentation
 
@@ -203,7 +202,7 @@ API: `predict_shifts`, `predict_shieldings`, `implicit_solvent_correction`,
 
 "Chemical Shift Prediction Beyond the Electronic Structure Limit."
 
-Adams, K.; Wagen, C.C.W.; Wolford, J.; Sak, M.H.; Saurí, J.; Feng, Z.; Bhadauria, A.S.; Bailey, M.A.; Downs, J.S.; Li, S.Z.; Liu, A.I.; Smidt, T.; Paton, R.S.; Liu, R.Y.; Coley, C.W.\*; Kwan, E.E.\*
+Adams, K.; Wagen, C.C.; Wolford, J.; Sak, M.H.; Saurí, J.; Feng, Z.; Bhadauria, A.S.; Bailey, M.A.; Downs, J.S.; Li, S.Z.; Liu, A.I.; Smidt, T.; Paton, R.S.; Liu, R.Y.; Coley, C.W.\*; Kwan, E.E.\*
 
 *submitted*, July 2026.
 
@@ -211,5 +210,5 @@ Adams, K.; Wagen, C.C.W.; Wolford, J.; Sak, M.H.; Saurí, J.; Feng, Z.; Bhadauri
 
 - The original code, model weights, and datasets in this repository are released under the MIT License (see [`LICENSE`](LICENSE)).
 - Third-party literature data redistributed here remains subject to its original publications' terms and should be
-cited accordingly: the Goodman CP3 spreadsheet (`data/cp3/`), the Schattenberg and Kaupp NS372
-spreadsheet (`data/ns372/`), and the DFT8K benchmark (`data/dft8k/`, from Guan and Paton).
+cited accordingly: CP3 (`data/cp3/`, Smith and Goodman), NS372
+ (`data/ns372/`, Schattenberg and Kaupp), and DFT8K (`data/dft8k/`, Guan and Paton).

analysis/code/build_nb_composite_models.py

@@ -54,13 +54,13 @@ N_SPLITS = 250
 
 si_composite_models_ablations = [
     md(r"""
-# SI Figure S15: composite-formula ablation workbook (`ablations.xlsx`)
+# Composite-formula ablation workbook
 
-Rebuilds the `ablations.xlsx` workbook behind ~32 of SI Figure S15's images: ~25 composite-formula
-variants (stationary geometry plus some combination of implicit PCM, explicit Desmond, and
-rovibrational QCD corrections) across all 12 delta-22 solvents, at two reference levels:
-DSD-PBEP86/pcSseg-3 (geometry PBE0/tz) and the MagNET-Zero training reference (WP04/pcSseg-2 for ¹H,
-wB97X-D/pcSseg-2 for ¹³C).
+Rebuilds the composite-formula ablation grid: ~25 composite-formula variants (stationary geometry
+plus some combination of implicit PCM, explicit Desmond, and rovibrational QCD corrections) across
+all 12 delta-22 solvents, at two reference levels: DSD-PBEP86/pcSseg-3 (geometry PBE0/tz) and the
+MagNET-Zero training reference (WP04/pcSseg-2 for ¹H, ωB97X-D/pcSseg-2 for ¹³C). Also reports the
+solvent-averaged correlations between the composite-model features.
 """),
     code(_BOOTSTRAP),
     code(_IMPORTS),
@@ -93,13 +93,13 @@ PCM-vs-Desmond table.
 """),
     code(r"""
 for nucleus, label in [("H", "Proton"), ("C", "Carbon")]:
-    corr = delta22.si_s15_feature_correlations(query_df_dft, nucleus)
+    corr = delta22.ablations_feature_correlations(query_df_dft, nucleus)
     nuc_label = "1H" if nucleus == "H" else "13C"
     nuc_title = "$^{1}$H" if nucleus == "H" else "$^{13}$C"
     composite_plots.plot_feature_correlation_heatmap(
         corr["r"], vmin=-1, vmax=1, cmap="RdBu",
         title=f"Solvent-Averaged Pearson $r$ Correlation Matrix for {nuc_title}",
-        save_path=figure_path(f"si_figure_s15_feature_corr_r_{nuc_label}.png"))
+        save_path=figure_path(f"ablations_feature_corr_r_{nuc_label}.png"))
 plt.show()
 """),
     code(r"""
@@ -107,7 +107,7 @@ pcm_desmond_corr = delta22.pcm_desmond_correlation_by_solvent(query_df_dft)
 print("PCM vs. Desmond Pearson R by nucleus and solvent:")
 display(pcm_desmond_corr.round(3))
 composite_plots.plot_pcm_desmond_correlation_table(pcm_desmond_corr,
-                                   save_path=figure_path("si_figure_s15_pcm_desmond_corr_table.png"))
+                                   save_path=figure_path("ablations_pcm_desmond_corr_table.png"))
 plt.show()
 """),
 ]

analysis/code/build_nb_leveling.py

@@ -29,7 +29,7 @@ Two independent NMR shielding datasets, analyzed separately (never pooled):
 | Dataset | Source | Methods | Nuclei | Reference |
 |---|---|---|---|---|
 | **NS372** | Schattenberg & Kaupp, *JCTC* **17**, 7602 (2021) | 44 DFT/WFT functionals | ¹H ¹¹B ¹³C ¹⁵N ¹⁷O ¹⁹F ³¹P ³³S | CCSD(T)/pcSseg-3 |
-| **delta22** | in-house `delta22.hdf5` | 18 gas-phase functionals, largest basis (pcSseg-3; mp2 → pcSseg-2), PBE0/cc-pVTZ geometry | ¹H ¹³C | DSD-PBEP86 (highest-rung in-set method, stands in for CCSD(T)) |
+| **delta22** | in-house delta-22 set | 18 gas-phase functionals, largest basis (pcSseg-3; mp2 → pcSseg-2), PBE0/cc-pVTZ geometry | ¹H ¹³C | DSD-PBEP86 (highest-rung in-set method, stands in for CCSD(T)) |
 """
 
 _CONFIG_MD = "## 1. Configuration"
@@ -69,8 +69,11 @@ def figure_path(name):
     return os.path.join("figures", name)
 
 # self-clean: this notebook builds PNG names dynamically (one per nucleus), so drop any
-# previously written figures before regenerating (glob on a missing folder returns [])
-for _stale in glob.glob(os.path.join("figures", "*.png")):
+# previously written figures before regenerating. Scope the delete to THIS figure's own prefix:
+# the figures/ folder is shared with the other notebooks in this directory, and a full reproduce
+# run executes them in name order, so a broad "figures/*.png" wipe would destroy the panels written
+# by notebooks that sort before this one. (glob on a missing folder returns [])
+for _stale in glob.glob(os.path.join("figures", "si_figure_s03_*.png")):
     os.remove(_stale)
 
 pd.set_option('display.width', 150)
@@ -120,14 +123,14 @@ _NS372_DEF_MD = r"""
 
 Conventional GIAO shieldings for 44 functionals across 8 main-group nuclei, with a
 CCSD(T)/pcSseg-3 reference. Kaupp Reduced-Set exclusions (F₃⁻, O₃, BH - multireference
-outliers) are applied. Input: `ct1c00919_si_002.xlsx`.
+outliers) are applied. Input: the Kaupp NS372 supporting-information spreadsheet.
 """
 
 _DELTA22_DEF_MD = r"""
 ## 3. delta22 - definitions
 
-Gas-phase conventional GIAO shieldings from `delta22.hdf5`: 18 functionals at their
-largest available basis (pcSseg-3; plain `mp2` only to pcSseg-2), at the PBE0/cc-pVTZ
+Gas-phase conventional GIAO shieldings from the delta-22 set: 18 functionals at their
+largest available basis (pcSseg-3; plain MP2 only to pcSseg-2), at the PBE0/cc-pVTZ
 geometry. Observations are pooled ¹H / ¹³C atom sites across all 22 solutes. There is no
 CCSD(T) reference in the file; **DSD-PBEP86 is used as the reference** for delta22 - it
 is the highest-rung double-hybrid available in this method set and stands in for CCSD(T)
@@ -138,7 +141,7 @@ _LOAD_MD = r"""
 ## 4. Load datasets + global colour scale
 
 Run the loaders, analyse every nucleus, and compute the figure-wide
-`-log10(1-|r|)` maximum (`GLOBAL_VMAX`) used as the colour scale on every correlation matrix
+-log10(1-|r|) maximum used as the colour scale on every correlation matrix
 below, for NS372 and delta22 alike.
 """
 
@@ -190,8 +193,9 @@ sum_ns372.round(5)
     md(r"""
 ### 5.2  Per-method scaled RMSE vs CCSD(T)
 
-`scaled_rmse` = RMSE of residuals after a per-method linear fit
-`sigma_method ~ a*sigma_CCSD(T) + b` - the error that survives empirical linear scaling.
+Scaled RMSE = RMSE of residuals after a per-method linear fit of each method's shieldings
+against the CCSD(T) reference (slope and intercept) - the error that survives empirical
+linear scaling.
 """),
     code(r"""
 rmse_ns372 = pd.DataFrame({nuc: res_ns372[nuc]['scaled_rmse'] for nuc in res_ns372})

analysis/code/build_nb_scaling.py

@@ -92,9 +92,9 @@ print("Symmetrized deployment Table S11 (13C):"); display(symmetrized["C"])
     md(r"""
 ## Composite-model coefficients (Figure 5C/5D, SI S15)
 
-The natural-products figures use a larger family of per-solvent coefficients fit on delta-22 (the
-`composite_model` group in `applications.hdf5`): OLS fits, 1000-seed bootstrap resamples, their RMSE
-distributions, and the PCM conversion factors. The group regenerates from released inputs alone.
+The natural-products figures use a larger family of per-solvent coefficients fit on delta-22: OLS
+fits, 1000-seed bootstrap resamples, their RMSE distributions, and the PCM conversion factors. These
+regenerate from released inputs alone.
 """),
     code(r"""
 reader = Applications(paths.dataset_file("applications", root=REPO))

analysis/code/composite_models.py

@@ -1,4 +1,4 @@
-"""Reproduces the delta-22 composite-formula ablation study behind SI Figure S15's Excel screenshots.
+"""Reproduces the delta-22 composite-formula ablation study (the ablations.xlsx workbook).
 
 Rebuilds `ablations.xlsx`: ~25 composite-formula variants (stationary geometry plus some combination
 of implicit PCM, explicit Desmond, and rovibrational QCD corrections) across all 12 delta-22 solvents,
@@ -110,8 +110,8 @@ CARBON_FORMULA_CONFIG = {
 _FORMULA_CONFIG = {"H": PROTON_FORMULA_CONFIG, "C": CARBON_FORMULA_CONFIG}
 
 # This ablation study fits plain OLS composite formulas, not an ML model. The canonical SI text
-# for Figure S15 states that "since no ML was used in this particular analysis, all 22 molecules
-# were considered, including nitromethane" -- other, ML-related analyses (e.g. scaling_factors.py's
+# for this ablation analysis states that "since no ML was used in this particular analysis, all 22
+# molecules were considered, including nitromethane" -- other, ML-related analyses (e.g. scaling_factors.py's
 # EXCLUDE_SOLUTES) drop nitromethane as an ML-training outlier; this module's default excludes nothing.
 EXCLUDE_SOLUTES = ()
 

analysis/code/composite_plots.py

@@ -1,6 +1,6 @@
-"""Correlation-heatmap panels for the composite-model ablation SI figure (Figure S15).
+"""Correlation-heatmap panels for the composite-formula ablation workbook.
 
-Two seaborn heatmaps behind Figure S15's feature-correlation panels: a per-nucleus lower-triangle
+Two seaborn heatmaps behind the ablation workbook's feature-correlation panels: a per-nucleus lower-triangle
 correlation matrix among the five composite-model features (stationary shielding, PCM, Desmond, its
 vibrational analogue, QCD), and a PCM-vs-Desmond Pearson r table by nucleus and solvent. The notebook
 supplies the correlation table(s) computed by delta22.py and calls these functions directly.

analysis/code/delta22.py

@@ -721,15 +721,15 @@ def correlation_matrix(query_df, value_col, sites_index, column_var):
     return pivot.corr()
 
 
-# SI Figure S15 "Correlations Between Features": correlations among the composite features within
-# one nucleus and solvent.
+# Composite-formula ablation workbook, "Correlations Between Features": correlations among the
+# composite features within one nucleus and solvent.
 FEATURE_CORRELATION_COLUMNS = ("stationary", "pcm", "desmond", "desmond_vib", "qcd")
 
 
 def feature_correlation_matrix(query_df, nucleus, solvent, cols=FEATURE_CORRELATION_COLUMNS, squared=False):
     """Pearson correlation matrix (R^2 if squared=True) between the feature columns, within one
     nucleus and solvent, over sites where both features have a value. query_df must be filtered to a
-    single (sap_nmr_method, sap_basis, sap_geometry_type); si_s15_feature_correlations does that."""
+    single (sap_nmr_method, sap_basis, sap_geometry_type); ablations_feature_correlations does that."""
     sub = query_df[(query_df["nucleus"] == nucleus) & (query_df["solvent"] == solvent)][list(cols)]
     sub = sub.dropna(how="all")
     if len(sub) < 2:
@@ -740,7 +740,7 @@ def feature_correlation_matrix(query_df, nucleus, solvent, cols=FEATURE_CORRELAT
 
 def average_feature_correlation_matrix(query_df, nucleus, solvents, cols=FEATURE_CORRELATION_COLUMNS,
                                        squared=False):
-    """Cell-by-cell nanmean of feature_correlation_matrix over the solvents (SI Figure S15's
+    """Cell-by-cell nanmean of feature_correlation_matrix over the solvents (the ablation workbook's
     solvent-averaged matrix). nanmean so a solvent with too few sites for one feature pair drops out
     of that cell alone, not the whole average. Same single-level-of-theory requirement."""
     mats = [feature_correlation_matrix(query_df, nucleus, s, cols, squared) for s in solvents]
@@ -750,7 +750,7 @@ def average_feature_correlation_matrix(query_df, nucleus, solvents, cols=FEATURE
     return pd.DataFrame(avg, index=mats[0].index, columns=mats[0].columns)
 
 
-def si_s15_feature_correlations(query_df_dft, nucleus, solvents=None):
+def ablations_feature_correlations(query_df_dft, nucleus, solvents=None):
     """Solvent-averaged feature correlations for one nucleus from the unfiltered
     load_query_df_dft(...) output: filters to the reference level
     (MAGNET_PCM_OUTPUT_METHODS[nucleus], pcSseg2, aimnet2), then returns {"r": Pearson R matrix,
@@ -765,9 +765,9 @@ def si_s15_feature_correlations(query_df_dft, nucleus, solvents=None):
 
 
 def pcm_desmond_correlation_by_solvent(query_df_dft, solvents=None):
-    """SI Figure S15's PCM-vs-Desmond table: per nucleus and solvent, the Pearson correlation
+    """The ablation workbook's PCM-vs-Desmond table: per nucleus and solvent, the Pearson correlation
     between the PCM (implicit) and Desmond (explicit) corrections. Filters to the reference level
-    per nucleus itself (same as si_s15_feature_correlations). Returns a DataFrame indexed by
+    per nucleus itself (same as ablations_feature_correlations). Returns a DataFrame indexed by
     nucleus, one column per solvent."""
     solvents = list(solvents) if solvents is not None else sorted(query_df_dft["solvent"].unique())
     rows = {}

analysis/code/leveling.py

@@ -274,43 +274,42 @@ DELTA22_REF = 'DSD-PBEP86'
 
 def load_delta22(path):
     """Load the delta22 shielding matrix, sorted by family, for 1H and 13C sites."""
-    f = h5py.File(path, 'r')
-    mn = [m.decode() for m in f['conventional_nmr_method_names'][:]]
-    order = ['hf', 'blyp_d3bj', 'bp86_d3bj', 'b97d3_d3bj', 'tpsstpss_d3bj',
-             'b3lyp_d3bj', 'pbe0_d3bj', 'm062x_d3', 'wb97xd', 'wp04', 'wc04',
-             'B2PLYP', 'mPW2PLYP', 'B2GP_PLYP', 'dsd_pbep86', 'revdsd_pbep86',
-             'mp2', 'dlpno_mp2']
-    disp = {'hf': 'HF', 'blyp_d3bj': 'BLYP-D3', 'bp86_d3bj': 'BP86-D3',
-            'b97d3_d3bj': 'B97D3', 'tpsstpss_d3bj': 'TPSS-D3', 'b3lyp_d3bj': 'B3LYP-D3',
-            'pbe0_d3bj': 'PBE0-D3', 'm062x_d3': 'M06-2X-D3', 'wb97xd': 'ωB97X-D',
-            'wp04': 'WP04', 'wc04': 'WC04', 'B2PLYP': 'B2PLYP', 'mPW2PLYP': 'mPW2PLYP',
-            'B2GP_PLYP': 'B2GP-PLYP', 'dsd_pbep86': 'DSD-PBEP86',
-            'revdsd_pbep86': 'revDSD-PBEP86', 'mp2': 'MP2', 'dlpno_mp2': 'DLPNO-MP2'}
-    fam = {'hf': 'WFT', 'blyp_d3bj': 'GGA', 'bp86_d3bj': 'GGA', 'b97d3_d3bj': 'GGA',
-           'tpsstpss_d3bj': 'mGGA', 'b3lyp_d3bj': 'GH', 'pbe0_d3bj': 'GH',
-           'm062x_d3': 'GH', 'wb97xd': 'RSH', 'wp04': 'GH', 'wc04': 'GH',
-           'B2PLYP': 'DH', 'mPW2PLYP': 'DH', 'B2GP_PLYP': 'DH', 'dsd_pbep86': 'DH',
-           'revdsd_pbep86': 'DH', 'mp2': 'WFT', 'dlpno_mp2': 'WFT'}
-    gas = {}
-    for i, m in enumerate(mn):
-        func, basis, smodel, _ = m.split(',')
-        if smodel == 'gas':
-            gas.setdefault(func, {})[basis] = i
-    sel, labels, families = [], [], []
-    for func in order:
-        b = gas[func]
-        big = 'pcSseg3' if 'pcSseg3' in b else ('pcSseg2' if 'pcSseg2' in b else 'pcSseg1')
-        sel.append(b[big]); labels.append(disp[func]); families.append(fam[func])
-    GEOM = 1  # geometries axis: 0 = AIMNet2, 1 = PBE0/cc-pVTZ
-    rows = {1: [], 6: []}
-    for s in f['solutes']:
-        an = f['solutes'][s]['atomic_numbers'][:]
-        cs = _decode_fixed_point(
-            f['solutes'][s]['stationary_and_pcm']['conventional_shieldings'][:, GEOM, :])
-        for ai, z in enumerate(an):
-            if z in (1, 6):
-                rows[int(z)].append(cs[sel, ai])
-    f.close()
+    with h5py.File(path, 'r') as f:
+        mn = [m.decode() for m in f['conventional_nmr_method_names'][:]]
+        order = ['hf', 'blyp_d3bj', 'bp86_d3bj', 'b97d3_d3bj', 'tpsstpss_d3bj',
+                 'b3lyp_d3bj', 'pbe0_d3bj', 'm062x_d3', 'wb97xd', 'wp04', 'wc04',
+                 'B2PLYP', 'mPW2PLYP', 'B2GP_PLYP', 'dsd_pbep86', 'revdsd_pbep86',
+                 'mp2', 'dlpno_mp2']
+        disp = {'hf': 'HF', 'blyp_d3bj': 'BLYP-D3', 'bp86_d3bj': 'BP86-D3',
+                'b97d3_d3bj': 'B97D3', 'tpsstpss_d3bj': 'TPSS-D3', 'b3lyp_d3bj': 'B3LYP-D3',
+                'pbe0_d3bj': 'PBE0-D3', 'm062x_d3': 'M06-2X-D3', 'wb97xd': 'ωB97X-D',
+                'wp04': 'WP04', 'wc04': 'WC04', 'B2PLYP': 'B2PLYP', 'mPW2PLYP': 'mPW2PLYP',
+                'B2GP_PLYP': 'B2GP-PLYP', 'dsd_pbep86': 'DSD-PBEP86',
+                'revdsd_pbep86': 'revDSD-PBEP86', 'mp2': 'MP2', 'dlpno_mp2': 'DLPNO-MP2'}
+        fam = {'hf': 'WFT', 'blyp_d3bj': 'GGA', 'bp86_d3bj': 'GGA', 'b97d3_d3bj': 'GGA',
+               'tpsstpss_d3bj': 'mGGA', 'b3lyp_d3bj': 'GH', 'pbe0_d3bj': 'GH',
+               'm062x_d3': 'GH', 'wb97xd': 'RSH', 'wp04': 'GH', 'wc04': 'GH',
+               'B2PLYP': 'DH', 'mPW2PLYP': 'DH', 'B2GP_PLYP': 'DH', 'dsd_pbep86': 'DH',
+               'revdsd_pbep86': 'DH', 'mp2': 'WFT', 'dlpno_mp2': 'WFT'}
+        gas = {}
+        for i, m in enumerate(mn):
+            func, basis, smodel, _ = m.split(',')
+            if smodel == 'gas':
+                gas.setdefault(func, {})[basis] = i
+        sel, labels, families = [], [], []
+        for func in order:
+            b = gas[func]
+            big = 'pcSseg3' if 'pcSseg3' in b else ('pcSseg2' if 'pcSseg2' in b else 'pcSseg1')
+            sel.append(b[big]); labels.append(disp[func]); families.append(fam[func])
+        GEOM = 1  # geometries axis: 0 = AIMNet2, 1 = PBE0/cc-pVTZ
+        rows = {1: [], 6: []}
+        for s in f['solutes']:
+            an = f['solutes'][s]['atomic_numbers'][:]
+            cs = _decode_fixed_point(
+                f['solutes'][s]['stationary_and_pcm']['conventional_shieldings'][:, GEOM, :])
+            for ai, z in enumerate(an):
+                if z in (1, 6):
+                    rows[int(z)].append(cs[sel, ai])
     out = {}
     for nuc, z in (('1H', 1), ('13C', 6)):
         M = np.array(rows[z])

analysis/code/test_delta22_analysis.py

@@ -790,7 +790,7 @@ def test_fig2a_reproduces_published_magnet_point():
         assert got == pytest.approx(expected, abs=1e-4), f"{nucleus}: {got} vs {expected}"
 
 
-# --- SI Figure S15's "Correlations Between Features" ---------------------------------------------
+# --- Composite-formula ablation workbook's "Correlations Between Features" -----------------------
 
 def test_feature_correlation_matrix_synthetic():
     # a and b perfectly correlated, c independent noise uncorrelated with either
@@ -841,9 +841,10 @@ def test_average_feature_correlation_matrix_ignores_a_solvent_with_too_few_sites
 @pytest.mark.skipif(
     not (os.path.exists(REAL_H5) and os.path.exists(REAL_XLSX)),
     reason="real delta22.hdf5 / experimental xlsx not present")
-def test_si_s15_feature_correlations_reproduces_published_matrices():
+def test_ablations_feature_correlations_reproduces_published_matrices():
     """Every entry of both nuclei's published Pearson R and R^2 "Correlations Between Features"
-    matrices (SI Figure S15), reproduced from the released data to 3 decimal places. Guards against
+    matrices (the composite-formula ablation workbook), reproduced from the released data to 3
+    decimal places. Guards against
     pooling all DFT methods/bases/geometries together instead of filtering to the MagNET-Zero
     reference level, which throws entries off by up to 0.03."""
     dft = D.load_query_df_dft(REAL_H5, REAL_XLSX, verbose=False)
@@ -866,7 +867,7 @@ def test_si_s15_feature_correlations_reproduces_published_matrices():
               ("qcd", "stationary"): 0.380, ("qcd", "pcm"): 0.404, ("qcd", "desmond"): 0.189, ("qcd", "desmond_vib"): 0.021},
     }
     for nucleus in ("H", "C"):
-        corr = D.si_s15_feature_correlations(dft, nucleus)
+        corr = D.ablations_feature_correlations(dft, nucleus)
         for (a, b), expected in published_r[nucleus].items():
             assert corr["r"].loc[a, b] == pytest.approx(expected, abs=5e-4), f"{nucleus} r[{a},{b}]"
         for (a, b), expected in published_r2[nucleus].items():

analysis/manuscript_figures/fig5b_dft8k.ipynb

@@ -144,13 +144,7 @@
    "cell_type": "markdown",
    "id": "5f4276e7",
    "metadata": {},
-   "source": [
-    "## Functional-group error breakdown\n",
-    "\n",
-    "Mean absolute residual for six functional groups (Carbonyls, Amines, Sulfonyl, Pyridines, Furans,\n",
-    "Nitroso), via RDKit SMARTS matching over every molecule's SMILES\n",
-    "(`dft8k_residuals.FUNCTIONAL_GROUP_SMARTS`)."
-   ]
+   "source": "## Functional-group error breakdown\n\nMean absolute residual for six functional groups (Carbonyls, Amines, Sulfonyl, Pyridines, Furans,\nNitroso), via RDKit SMARTS matching over every molecule's SMILES."
   },
   {
    "cell_type": "code",

analysis/si_figures/si_composite_models_ablations.ipynb

@@ -2,22 +2,22 @@
  "cells": [
   {
    "cell_type": "markdown",
-   "id": "bfdd4bf5",
+   "id": "77f6c813",
    "metadata": {},
    "source": [
-    "# SI Figure S15: composite-formula ablation workbook (`ablations.xlsx`)\n",
+    "# Composite-formula ablation workbook\n",
     "\n",
-    "Rebuilds the `ablations.xlsx` workbook behind ~32 of SI Figure S15's images: ~25 composite-formula\n",
-    "variants (stationary geometry plus some combination of implicit PCM, explicit Desmond, and\n",
-    "rovibrational QCD corrections) across all 12 delta-22 solvents, at two reference levels:\n",
-    "DSD-PBEP86/pcSseg-3 (geometry PBE0/tz) and the MagNET-Zero training reference (WP04/pcSseg-2 for ¹H,\n",
-    "wB97X-D/pcSseg-2 for ¹³C)."
+    "Rebuilds the composite-formula ablation grid: ~25 composite-formula variants (stationary geometry\n",
+    "plus some combination of implicit PCM, explicit Desmond, and rovibrational QCD corrections) across\n",
+    "all 12 delta-22 solvents, at two reference levels: DSD-PBEP86/pcSseg-3 (geometry PBE0/tz) and the\n",
+    "MagNET-Zero training reference (WP04/pcSseg-2 for ¹H, ωB97X-D/pcSseg-2 for ¹³C). Also reports the\n",
+    "solvent-averaged correlations between the composite-model features."
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "cb68d1d6",
+   "id": "e00f8b39",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -32,7 +32,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "20914013",
+   "id": "b3acc33b",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -47,7 +47,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "26ab5657",
+   "id": "a909ba74",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -69,7 +69,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "76d31292",
+   "id": "5df608e9",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -80,7 +80,7 @@
   },
   {
    "cell_type": "markdown",
-   "id": "2c351f43",
+   "id": "b7bc63a7",
    "metadata": {},
    "source": [
     "## Preview: one formula's mean test RMSE at the DSD-PBEP86 reference level"
@@ -89,7 +89,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "eb03ec52",
+   "id": "ad79f75e",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -103,7 +103,7 @@
   },
   {
    "cell_type": "markdown",
-   "id": "ee9136d7",
+   "id": "7ae7d282",
    "metadata": {},
    "source": [
     "## Build the full ablations workbook (both reference levels, both nuclei)"
@@ -112,7 +112,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "5c9fbbd3",
+   "id": "f9acdb9e",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -122,7 +122,7 @@
   },
   {
    "cell_type": "markdown",
-   "id": "78a693c7",
+   "id": "66102dfb",
    "metadata": {},
    "source": [
     "## Correlations Between Features\n",
@@ -135,25 +135,25 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "9ef593a6",
+   "id": "208d343b",
    "metadata": {},
    "outputs": [],
    "source": [
     "for nucleus, label in [(\"H\", \"Proton\"), (\"C\", \"Carbon\")]:\n",
-    "    corr = delta22.si_s15_feature_correlations(query_df_dft, nucleus)\n",
+    "    corr = delta22.ablations_feature_correlations(query_df_dft, nucleus)\n",
     "    nuc_label = \"1H\" if nucleus == \"H\" else \"13C\"\n",
     "    nuc_title = \"$^{1}$H\" if nucleus == \"H\" else \"$^{13}$C\"\n",
     "    composite_plots.plot_feature_correlation_heatmap(\n",
     "        corr[\"r\"], vmin=-1, vmax=1, cmap=\"RdBu\",\n",
     "        title=f\"Solvent-Averaged Pearson $r$ Correlation Matrix for {nuc_title}\",\n",
-    "        save_path=figure_path(f\"si_figure_s15_feature_corr_r_{nuc_label}.png\"))\n",
+    "        save_path=figure_path(f\"ablations_feature_corr_r_{nuc_label}.png\"))\n",
     "plt.show()"
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "5bc382f2",
+   "id": "4e3701cd",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -161,7 +161,7 @@
     "print(\"PCM vs. Desmond Pearson R by nucleus and solvent:\")\n",
     "display(pcm_desmond_corr.round(3))\n",
     "composite_plots.plot_pcm_desmond_correlation_table(pcm_desmond_corr,\n",
-    "                                   save_path=figure_path(\"si_figure_s15_pcm_desmond_corr_table.png\"))\n",
+    "                                   save_path=figure_path(\"ablations_pcm_desmond_corr_table.png\"))\n",
     "plt.show()"
    ]
   }

analysis/si_figures/si_figure_s03_leveling.ipynb

@@ -2,7 +2,7 @@
  "cells": [
   {
    "cell_type": "markdown",
-   "id": "b61775ea",
+   "id": "3a2b8d51",
    "metadata": {},
    "source": [
     "# SI Figure S3: inter-method correlation and PCA of NS372 and delta-22 shieldings\n",
@@ -13,7 +13,7 @@
   },
   {
    "cell_type": "markdown",
-   "id": "aeb46df4",
+   "id": "0fb6f798",
    "metadata": {},
    "source": [
     "Two independent NMR shielding datasets, analyzed separately (never pooled):\n",
@@ -21,12 +21,12 @@
     "| Dataset | Source | Methods | Nuclei | Reference |\n",
     "|---|---|---|---|---|\n",
     "| **NS372** | Schattenberg & Kaupp, *JCTC* **17**, 7602 (2021) | 44 DFT/WFT functionals | ¹H ¹¹B ¹³C ¹⁵N ¹⁷O ¹⁹F ³¹P ³³S | CCSD(T)/pcSseg-3 |\n",
-    "| **delta22** | in-house `delta22.hdf5` | 18 gas-phase functionals, largest basis (pcSseg-3; mp2 → pcSseg-2), PBE0/cc-pVTZ geometry | ¹H ¹³C | DSD-PBEP86 (highest-rung in-set method, stands in for CCSD(T)) |"
+    "| **delta22** | in-house delta-22 set | 18 gas-phase functionals, largest basis (pcSseg-3; mp2 → pcSseg-2), PBE0/cc-pVTZ geometry | ¹H ¹³C | DSD-PBEP86 (highest-rung in-set method, stands in for CCSD(T)) |"
    ]
   },
   {
    "cell_type": "markdown",
-   "id": "d0ce838b",
+   "id": "3ff927f7",
    "metadata": {},
    "source": [
     "## 1. Configuration"
@@ -35,7 +35,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "62157a75",
+   "id": "9e5a25b8",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -49,7 +49,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "8ea46824",
+   "id": "06e0333d",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -65,7 +65,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "34796b33",
+   "id": "fbd3d68f",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -85,8 +85,11 @@
     "    return os.path.join(\"figures\", name)\n",
     "\n",
     "# self-clean: this notebook builds PNG names dynamically (one per nucleus), so drop any\n",
-    "# previously written figures before regenerating (glob on a missing folder returns [])\n",
-    "for _stale in glob.glob(os.path.join(\"figures\", \"*.png\")):\n",
+    "# previously written figures before regenerating. Scope the delete to THIS figure's own prefix:\n",
+    "# the figures/ folder is shared with the other notebooks in this directory, and a full reproduce\n",
+    "# run executes them in name order, so a broad \"figures/*.png\" wipe would destroy the panels written\n",
+    "# by notebooks that sort before this one. (glob on a missing folder returns [])\n",
+    "for _stale in glob.glob(os.path.join(\"figures\", \"si_figure_s03_*.png\")):\n",
     "    os.remove(_stale)\n",
     "\n",
     "pd.set_option('display.width', 150)\n",
@@ -96,7 +99,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "a402f2fc",
+   "id": "84ddcb3f",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -139,25 +142,25 @@
   },
   {
    "cell_type": "markdown",
-   "id": "52b5cdac",
+   "id": "4fbeeddb",
    "metadata": {},
    "source": [
     "## 2. NS372 (Kaupp) - definitions\n",
     "\n",
     "Conventional GIAO shieldings for 44 functionals across 8 main-group nuclei, with a\n",
     "CCSD(T)/pcSseg-3 reference. Kaupp Reduced-Set exclusions (F₃⁻, O₃, BH - multireference\n",
-    "outliers) are applied. Input: `ct1c00919_si_002.xlsx`."
+    "outliers) are applied. Input: the Kaupp NS372 supporting-information spreadsheet."
    ]
   },
   {
    "cell_type": "markdown",
-   "id": "f87c7141",
+   "id": "34326933",
    "metadata": {},
    "source": [
     "## 3. delta22 - definitions\n",
     "\n",
-    "Gas-phase conventional GIAO shieldings from `delta22.hdf5`: 18 functionals at their\n",
-    "largest available basis (pcSseg-3; plain `mp2` only to pcSseg-2), at the PBE0/cc-pVTZ\n",
+    "Gas-phase conventional GIAO shieldings from the delta-22 set: 18 functionals at their\n",
+    "largest available basis (pcSseg-3; plain MP2 only to pcSseg-2), at the PBE0/cc-pVTZ\n",
     "geometry. Observations are pooled ¹H / ¹³C atom sites across all 22 solutes. There is no\n",
     "CCSD(T) reference in the file; **DSD-PBEP86 is used as the reference** for delta22 - it\n",
     "is the highest-rung double-hybrid available in this method set and stands in for CCSD(T)\n",
@@ -166,20 +169,20 @@
   },
   {
    "cell_type": "markdown",
-   "id": "be87913f",
+   "id": "eedf9b3d",
    "metadata": {},
    "source": [
     "## 4. Load datasets + global colour scale\n",
     "\n",
     "Run the loaders, analyse every nucleus, and compute the figure-wide\n",
-    "`-log10(1-|r|)` maximum (`GLOBAL_VMAX`) used as the colour scale on every correlation matrix\n",
+    "-log10(1-|r|) maximum used as the colour scale on every correlation matrix\n",
     "below, for NS372 and delta22 alike."
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "d0a24460",
+   "id": "4bec0cf4",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -203,7 +206,7 @@
   },
   {
    "cell_type": "markdown",
-   "id": "0d07150e",
+   "id": "667fef3c",
    "metadata": {},
    "source": [
     "## 5. NS372 results"
@@ -211,7 +214,7 @@
   },
   {
    "cell_type": "markdown",
-   "id": "747989ff",
+   "id": "c5f2cbda",
    "metadata": {},
    "source": [
     "### 5.1  Leveling diagnostics - NS372"
@@ -220,7 +223,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "52a133ee",
+   "id": "8b9b9449",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -231,19 +234,20 @@
   },
   {
    "cell_type": "markdown",
-   "id": "a385e6f5",
+   "id": "2b666d04",
    "metadata": {},
    "source": [
     "### 5.2  Per-method scaled RMSE vs CCSD(T)\n",
     "\n",
-    "`scaled_rmse` = RMSE of residuals after a per-method linear fit\n",
-    "`sigma_method ~ a*sigma_CCSD(T) + b` - the error that survives empirical linear scaling."
+    "Scaled RMSE = RMSE of residuals after a per-method linear fit of each method's shieldings\n",
+    "against the CCSD(T) reference (slope and intercept) - the error that survives empirical\n",
+    "linear scaling."
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "319abc43",
+   "id": "4020359a",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -256,7 +260,7 @@
   },
   {
    "cell_type": "markdown",
-   "id": "59cb8672",
+   "id": "ba763232",
    "metadata": {},
    "source": [
     "### 5.3  Inter-method correlation matrices - NS372 (one nucleus per file)"
@@ -265,7 +269,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "73c90f1e",
+   "id": "133a5053",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -282,7 +286,7 @@
   },
   {
    "cell_type": "markdown",
-   "id": "33da4dd0",
+   "id": "0fa55480",
    "metadata": {},
    "source": [
     "### 5.4  PC1/PC2 structure - NS372 (one nucleus per file)"
@@ -291,7 +295,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "034bf193",
+   "id": "a37b38c6",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -307,7 +311,7 @@
   },
   {
    "cell_type": "markdown",
-   "id": "3ec202cf",
+   "id": "bc460701",
    "metadata": {},
    "source": [
     "### 5.5  delta22 panels (lead the published figure: corr ¹H/¹³C + PCA)\n",
@@ -319,7 +323,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "fc1d57ab",
+   "id": "74e02974",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -345,7 +349,7 @@
   },
   {
    "cell_type": "markdown",
-   "id": "4c1e4350",
+   "id": "49825011",
    "metadata": {},
    "source": [
     "## 6. Combined summary (both datasets)"
@@ -354,7 +358,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "185ce298",
+   "id": "bffe5ff3",
    "metadata": {},
    "outputs": [],
    "source": [

analysis/si_tables/si_table_s10_s11_scaling.ipynb

@@ -2,7 +2,7 @@
  "cells": [
   {
    "cell_type": "markdown",
-   "id": "badbc639",
+   "id": "41675f1b",
    "metadata": {},
    "source": [
     "# Tables S10 and S11: MagNET-Zero/MagNET-PCM scaling factors\n",
@@ -14,7 +14,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "672ff802",
+   "id": "fb163d74",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -30,7 +30,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "1a1fa320",
+   "id": "177bdc88",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -46,7 +46,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "deba4a1e",
+   "id": "7f56b61d",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -61,7 +61,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "e408fac2",
+   "id": "44a384c9",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -78,7 +78,7 @@
   },
   {
    "cell_type": "markdown",
-   "id": "b6094980",
+   "id": "219c400a",
    "metadata": {},
    "source": [
     "## Reproducibility check: re-derive from delta-22"
@@ -87,7 +87,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "19f08d14",
+   "id": "b525b479",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -103,7 +103,7 @@
   },
   {
    "cell_type": "markdown",
-   "id": "8edf48b3",
+   "id": "14d0ed95",
    "metadata": {},
    "source": [
     "## Deployment tables with reflection symmetrization\n",
@@ -116,7 +116,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "448f2e9f",
+   "id": "111de099",
    "metadata": {},
    "outputs": [],
    "source": [
@@ -127,20 +127,20 @@
   },
   {
    "cell_type": "markdown",
-   "id": "ca457bc6",
+   "id": "31089f6c",
    "metadata": {},
    "source": [
     "## Composite-model coefficients (Figure 5C/5D, SI S15)\n",
     "\n",
-    "The natural-products figures use a larger family of per-solvent coefficients fit on delta-22 (the\n",
-    "`composite_model` group in `applications.hdf5`): OLS fits, 1000-seed bootstrap resamples, their RMSE\n",
-    "distributions, and the PCM conversion factors. The group regenerates from released inputs alone."
+    "The natural-products figures use a larger family of per-solvent coefficients fit on delta-22: OLS\n",
+    "fits, 1000-seed bootstrap resamples, their RMSE distributions, and the PCM conversion factors. These\n",
+    "regenerate from released inputs alone."
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "47c3d1e7",
+   "id": "d2d9f7b4",
    "metadata": {},
    "outputs": [],
    "source": [

data/applications/applications_reader.py

@@ -28,7 +28,7 @@ SHELL_SIZES = [50, 150, 250, 350, 450, 550, 650]
 def _decode(values):
     """Turn stored whole numbers back into real values; blanks where data is missing."""
     values = np.asarray(values)
-    if values.dtype == np.int32:
+    if np.issubdtype(values.dtype, np.integer):
         out = values.astype(np.float64) / _SCALE
         out[values == _MISSING_MARKER] = np.nan
         return out

data/applications/test_applications.py

@@ -102,6 +102,17 @@ def test_decode_roundtrip_and_missing():
     assert np.isnan(_decode(enc_with_miss)[-1])
 
 
+def test_decode_int64_is_still_scaled():
+    # a rebuild that stored the columns as int64 (any integer width) must decode the same way as
+    # int32; the old exact `dtype == np.int32` check fell through and returned raw scaled integers.
+    vals = np.array([1.2345, -50.0, 0.0], dtype=np.float64)
+    enc = _enc(vals).astype(np.int64)
+    out = _decode(enc)
+    assert np.allclose(out, vals, atol=5e-5)
+    enc_with_miss = np.append(enc, np.int64(-2147483648))
+    assert np.isnan(_decode(enc_with_miss)[-1])
+
+
 def test_solutes_and_solvents_group_excluded(fake_release):
     assert set(fake_release.solutes()) == {"vomicine", "isomer_1E"}
 

data/delta22/delta22_reader.py

@@ -39,7 +39,7 @@ _MISSING_MARKER = -2147483648
 
 def _decode_fixed_point(values):
     values = np.asarray(values)
-    if values.dtype == np.int32:
+    if np.issubdtype(values.dtype, np.integer):
         out = values.astype(np.float64) / _FIXED_POINT_SCALE
         out[values == _MISSING_MARKER] = np.nan
         return out
@@ -74,7 +74,7 @@ def _sort_and_parse_raw_shieldings(shield_df):
         value_list = [v for v in value_list if v]
         try:
             return [float(i) for i in value_list]
-        except:
+        except (ValueError, TypeError):
             return []
 
     def _to_float_list(x):

data/delta22/test_delta22.py

@@ -144,6 +144,17 @@ def test_decode_passes_floats_through():
     np.testing.assert_array_equal(_decode_fixed_point(f), f)
 
 
+def test_decode_int64_is_still_scaled():
+    # a rebuild that stored the fixed-point columns as int64 (any integer width) must decode the
+    # same way as int32; the old exact `dtype == np.int32` check fell through and returned the raw
+    # scaled integers as if they were ppm.
+    enc = np.array([2070700, _MISSING_MARKER, -150000], np.int64)
+    out = _decode_fixed_point(enc)
+    assert out[0] == pytest.approx(207.07)
+    assert np.isnan(out[1])
+    assert out[2] == pytest.approx(-15.0)
+
+
 # ---- the full reader against the synthetic file ----
 EXPECTED_COLUMNS = ["experimental", "stationary", "qcd", "pcm",
                     "desmond", "openMM", "desmond_vib", "openMM_vib"]

magnet/eqV2/edge_rot_mat.py

@@ -5,15 +5,17 @@ def init_edge_rot_mat(edge_distance_vec):
     edge_vec_0 = edge_distance_vec
     edge_vec_0_distance = torch.sqrt(torch.sum(edge_vec_0**2, dim=1))
 
-    # Make sure the atoms are far enough apart
-    #assert torch.min(edge_vec_0_distance) < 0.0001
-    if torch.min(edge_vec_0_distance) < 0.0001:
-        print(
-            "Error edge_vec_0_distance: {}".format(
-                torch.min(edge_vec_0_distance)
-            )
+    # Two atoms at (nearly) the same position cannot define a local frame: the normalization below
+    # would divide by ~zero and silently produce NaN or a garbage frame that propagates through the
+    # whole forward pass. Refuse the input instead. This is reachable from raw geometries, e.g.
+    # overlapping atoms in an unrelaxed explicit-solvent snapshot.
+    min_distance = torch.min(edge_vec_0_distance)
+    if min_distance < 0.0001:
+        raise ValueError(
+            f"overlapping atoms: smallest interatomic distance {float(min_distance):.2e} is too "
+            "small to define a local frame; check the input geometry for coincident atoms"
         )
-        
+
     norm_x = edge_vec_0 / (edge_vec_0_distance.view(-1, 1))
 
     edge_vec_2 = torch.rand_like(edge_vec_0) - 0.5

magnet/eqV2/equiformer_v2_NMR.py

@@ -4,7 +4,6 @@ import math
 import numpy as np
 import torch
 import torch.nn as nn
-from pyexpat.model import XML_CQUANT_OPT
 
 from .ocpmodels.common.registry import registry
 from .ocpmodels.common.utils import conditional_grad
@@ -422,8 +421,7 @@ class EquiformerV2_NMR(BaseModel):
             edge_index = edge_index[:, ~remove_solvent_edge]
             edge_distance = edge_distance[~remove_solvent_edge]
             edge_distance_vec = edge_distance_vec[~remove_solvent_edge]
-            neighbors = torch.ones_like(neighbors) * edge_index.shape[1]
-        
+
         #print(edge_index.shape)
 
         

magnet/test_api.py

@@ -14,8 +14,11 @@ import pytest
 
 warnings.filterwarnings("ignore")
 
+import torch
+
 import magnet.api as api
 from magnet import scaling
+from magnet.eqV2.edge_rot_mat import init_edge_rot_mat
 
 _HERE = os.path.dirname(os.path.abspath(__file__))
 _REPO = os.path.dirname(_HERE)                       # magnet/ -> repo root
@@ -461,3 +464,22 @@ def test_real_atoms_beyond_cutoff_crash():
     # two atoms 50 A apart: no edges within cutoff -> same zero-edge crash as one atom.
     api.predict_shieldings(np.array([6, 1]), np.array([[0., 0., 0.], [0., 0., 50.]]),
                            n_passes=1, symmetrize=False)
+
+
+# ---- degenerate-geometry guard in the local-frame construction (no weights needed) ----
+def test_edge_rot_mat_rejects_coincident_atoms():
+    # an edge between two atoms at the same position has ~zero length; the old code printed a
+    # warning and then divided by ~zero, silently producing NaN frames that poison the whole
+    # forward pass. It must raise instead.
+    edge_vecs = torch.tensor([[1.0, 0.0, 0.0], [0.0, 0.0, 0.0]])
+    with pytest.raises(ValueError, match="overlapping atoms"):
+        init_edge_rot_mat(edge_vecs)
+
+
+def test_edge_rot_mat_accepts_separated_atoms():
+    edge_vecs = torch.tensor([[1.0, 0.0, 0.0], [0.0, 1.5, 0.0], [0.0, 0.0, 2.0]])
+    rot = init_edge_rot_mat(edge_vecs)
+    assert rot.shape == (3, 3, 3)
+    # each frame is a proper rotation: R @ R^T == I
+    identity = torch.eye(3).expand(3, 3, 3)
+    assert torch.allclose(torch.bmm(rot, rot.transpose(1, 2)), identity, atol=1e-4)