diskmelts 2.2.0__tar.gz → 2.2.2__tar.gz

diskmelts-2.2.2/MANIFEST.in

@@ -0,0 +1,2 @@
+# Tests need the repository checkout (fixtures, Realobs_data); keep them out of the sdist.
+prune tests

{diskmelts-2.2.0/src/diskmelts.egg-info → diskmelts-2.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: diskmelts
-Version: 2.2.0
+Version: 2.2.2
 Summary: Machine-learning Enabled Line fitting Tool for Spectra of protoplanetary disks
 Author: Chengyan Xie, Dingshan Deng
 License-Expression: MIT
@@ -15,12 +15,12 @@ Classifier: Topic :: Scientific/Engineering :: Astronomy
 Requires-Python: >=3.9
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: torch
+Requires-Dist: torch>=1.13
 Requires-Dist: numpy<2,>=1.26; sys_platform == "darwin" and platform_machine == "x86_64"
 Requires-Dist: numpy; sys_platform != "darwin" or platform_machine != "x86_64"
 Requires-Dist: pandas
 Requires-Dist: scipy
-Requires-Dist: scikit-learn
+Requires-Dist: scikit-learn>=1.0
 Requires-Dist: matplotlib
 Provides-Extra: io
 Requires-Dist: astropy; extra == "io"
@@ -49,10 +49,11 @@ density (`logN`), and emitting area scaling (`A`) from observed spectra.
 
 ## Installation
 
-DiskMELTS is available on [PyPI](https://pypi.org/project/diskmelts/).
-Version 2.2.0 bundles 44 production checkpoints, including sigma=2 models for
-18 species and CH3 in both sigma versions.
-With Python 3.9 or newer, install it from any directory. These commands use
+DiskMELTS is available on [PyPI](https://pypi.org/project/diskmelts/) and
+requires Python 3.9 or newer. It bundles 46 production checkpoints, 23 each
+for the sigma=0 and sigma=2 grids, covering the same 18 species (see
+[Available molecules](#available-molecules)).
+Install it from any directory. These commands use
 `python` in an activated environment; the
 [installation guide](https://diskmelts.readthedocs.io/en/latest/installation.html)
 shows how to create one on macOS, Linux, Windows, or Conda:
@@ -87,6 +88,36 @@ conda activate diskmelts
 Use `environment-notebooks.yaml` or `environment-dev.yaml` for the corresponding
 notebook or contributor dependencies.
 
+## Minimal workflow
+
+Subtract the continuum, fit one molecule, and print the result:
+
+```python
+from diskmelts import consub, fit_stage, load_fitting_models
+
+# 1. Raw spectrum: CSV with wave (µm), flux (Jy), and optional err columns.
+raw_wave, raw_flux, raw_err = consub.read_spectrum("my_raw_spectrum.csv")
+wave, flux, err, cont, cont_unshifted = consub.calculate(
+    raw_wave, raw_flux, raw_err, preset="ttauri",
+)
+
+# 2. Fit C2H2 near 13.7 µm with the bundled surrogate model.
+models = load_fitting_models("C2H2")
+fit = fit_stage(
+    wave, flux, molecule="C2H2", pretrained=models,
+    wavelength_range=(13.0, 15.0), fit_ranges=(13.5, 14.1),
+    distance_pc=140.0, n_samples=1024, n_refine=4, n_top=4,
+)
+
+# 3. Best-fit T (K), logN (cm^-2), A (au^2), and the spread of top solutions.
+print(fit["params"]["C2H2"])
+print(fit["uncertainty"]["C2H2"])
+```
+
+The sections below explain each step, and the
+[first-spectrum guide](https://diskmelts.readthedocs.io/en/latest/quickstart.html)
+covers multi-stage fits, saved outputs, and plots.
+
 ## Runtime dependencies
 
 | Package | Purpose |
@@ -95,7 +126,7 @@ notebook or contributor dependencies.
 | `numpy` | Array operations |
 | `pandas` | CSV I/O |
 | `scipy` | L-BFGS-B optimizer, NNLS, Sobol sampling |
-| `scikit-learn` | Model scaling and PCA metadata |
+| `scikit-learn` | PCA and scaling when training new models |
 | `matplotlib` | Plotting |
 
 ## Continuum subtraction
@@ -130,12 +161,6 @@ With `preset=None, linefreefile=None` (or `""`), fitting uses all nonexcluded
 pixels and skips the line-free correction, so `cont` equals `contOrig`.
 Reference tables are bundled; normal continuum subtraction needs no model grids.
 
-The private `0-Consub/` experiment directory is excluded from Git and is not
-required by the installed package. Its runtime code and reference tables are
-maintained under `src/diskmelts/consub/`.
-The distributed tests in `tests/` exercise the installed continuum API and
-bundled references without that private directory.
-
 Returned `err` represents measurement/noise error, not correlated continuum
 uncertainty. Broad excluded intervals require inspection. See the
 [continuum guide](https://diskmelts.readthedocs.io/en/latest/continuum.html) for presets, regional box sizes,
@@ -182,31 +207,35 @@ You can also read spectra with your own code. DiskMELTS only requires
 one-dimensional `obs_wav` and `obs_flux` arrays. Replace the sample CSV with
 your own file to fit your source; update `distance_pc` and fit intervals.
 
-For a complete staged fit-and-subtract workflow from a checkout, edit and run:
+For a complete staged fit-and-subtract workflow from a clone of the
+repository, run:
 
 ```bash
 python examples/dev_v2_realobs.py
+python examples/dev_v2_realobs.py --input /path/to/my_spectrum.csv --name my_source --distance-pc 140 --sigma 2
 ```
 
+Adjust that script's stage windows for your spectrum.
+`python examples/synthetic_array_fit.py` generates an HCN spectrum and fits it
+from in-memory arrays, without an observation file.
+
 The v2 fitting API is `load_fitting_models`, `fit_stage`, `fit_observation`, and
-`plot_saved_observation`. The production checkpoints are packaged under
-`src/diskmelts/models/` and included in the wheel. Model fitting requires no
-`Model_grids/` or `Pretrain_grid/` directory. `diskmelts.MODEL_SPECS` records
-PCA counts, network layers, and `grid_sigma=0` for the original checkpoints.
-In version 2.2.0, sigma=2 checkpoints are available for H2O,
-C2H2, HCN, CO2, 13C12CH2, 13CO2, C4H2, HC3N, C2H6, C2H4, CH4,
-C2HD, CO17O, CO18O, C3H4, CH3, C6H6, and OH. Select them with
-`load_fitting_models(['H2O', 'C2H2'], sigma=2)` or
-`fit_observation(..., sigma=2)`. The C2H2 bank covers 6–9 and 11.5–17 µm;
-choose fit windows within those bands. OH covers 15–26 µm. Both C6H6
-checkpoints cover 13–17 µm and 100–675 K; the sigma=2 checkpoint is selected
-with `load_fitting_models('C6H6', sigma=2)`.
-The checkpoint settings are checked when
-loading. The checkpoint flux convention is 1 au² at 140 pc; `distance_pc`
-sets the source distance and makes the returned emitting area physical.
-The [fitting notebook](https://github.com/wddlx/DiskMELTS/blob/main/notebooks/Example_Fitting.ipynb) requires 13C12CH2 to
-be no hotter than C2H2 and gives them the same emitting area. It ties 13CO2
-to CO2 in both temperature and area. Set
+`plot_saved_observation`. The production checkpoints are included in the
+wheel, so model fitting needs no local model grids. The checkpoint flux
+convention is 1 au² at 140 pc; `distance_pc` sets the source distance and
+makes the returned emitting area physical. `fit_observation` accepts either a
+CSV path or in-memory `obs_wav`, `obs_flux`, and optional `obs_err` arrays
+(new in 2.2.1).
+
+For a synthetic spectrum, `wave, flux = diskmelts.synthesize_spectrum('HCN',
+T=300, logN=18, A=1)` returns the native wavelength axis and model flux (new
+in 2.2.1). See the
+[generation guide](https://diskmelts.readthedocs.io/en/latest/generate_spectra.html).
+
+The [fitting notebook](https://github.com/wddlx/DiskMELTS/blob/main/notebooks/Example_Fitting.ipynb)
+requires 13C12CH2 to be no hotter than C2H2 and gives them the same emitting
+area; it ties 13CO2 to CO2 in both temperature and area. Point its input at
+your own continuum-subtracted CSV, or pass arrays. Set
 `isotopologue_groups`, `Tshare` (`free`, `equal`, or `less`), and `Ashare` in a
 joint stage to choose those constraints. See the
 [fitting guide](https://diskmelts.readthedocs.io/en/latest/quickstart.html) for examples.
@@ -218,6 +247,32 @@ separate method does require the full local grids. See the
 plotting.
 For complete sigma=2 grids, pass `grid_root='Model_grids/sigma2'`.
 
+## Available molecules
+
+Each molecule is available for two slab-model grids: sigma=0 (default) assumes
+no turbulent broadening, so lines are only thermally broadened, and sigma=2
+assumes 2 km/s turbulent broadening. Select sigma=2 with `load_fitting_models(..., sigma=2)` or
+`fit_observation(..., sigma=2)`. Fit windows must lie inside a model's band.
+
+| Molecule | Band (µm) | T (K) | logN (cm⁻²) |
+|---|---|---|---|
+| H2O | 4.9–25.0 (five overlapping tiles) | 100–1400 | 13–19 |
+| C2H2 | 6–9 and 11.5–17 | 100–1400 | 13–19.5 |
+| HCN, CO2, 13C12CH2, 13CO2 | 11.5–17 | 100–1400 | 13–19.5 |
+| HC3N, CO17O, CO18O | 13–18 | 100–1400 | 13–19.5 |
+| C4H2 | 14–18 | 100–1400 | 13–19.5 |
+| C3H4, CH3 | 15–18 | 100–1400 | 13–19.5 |
+| C6H6 | 13–17 | 100–675 | 13–19.5 |
+| C2H6 | 9–15 | 100–1400 | 13–19.5 |
+| C2H4 | 8–14 | 100–1400 | 13–19.5 |
+| CH4 | 5–9 | 100–1400 | 13–19.5 |
+| C2HD | 11.5–20.5 | 100–1400 | 13–19.5 |
+| OH | 15–26 | 100–1400 | 13–19.5 |
+
+`diskmelts.MODEL_SPECS` records the exact domains, PCA sizes, and network
+layers. Version 2.2.0 lacks the sigma=0 OH and 6–9 µm C2H2 checkpoints;
+upgrade with `python -m pip install -U diskmelts`.
+
 ## Training and validation data
 
 Generated data under `Model_grids/` and `Pretrain_grid/` are intentionally
@@ -250,7 +305,7 @@ uses the same names with local `Model_grids/sigma0/C6H6/` and `Model_grids/sigma
 files. See the [training guide](https://diskmelts.readthedocs.io/en/latest/training.html)
 for peak audits and retraining commands.
 
-Version 2.2.0 also includes a CH3 checkpoint for 15–18 µm,
+The package also includes a CH3 checkpoint for 15–18 µm,
 `T=100–1400 K`, and `logN=13–19.5`. Its strongest grid feature in this window
 is near 16.4–16.6 µm. Use `load_fitting_models('CH3')` after installing this
 package; direct-grid fitting uses local `Model_grids/sigma0/CH3/` files.

{diskmelts-2.2.0 → diskmelts-2.2.2}/README.md

@@ -10,10 +10,11 @@ density (`logN`), and emitting area scaling (`A`) from observed spectra.
 
 ## Installation
 
-DiskMELTS is available on [PyPI](https://pypi.org/project/diskmelts/).
-Version 2.2.0 bundles 44 production checkpoints, including sigma=2 models for
-18 species and CH3 in both sigma versions.
-With Python 3.9 or newer, install it from any directory. These commands use
+DiskMELTS is available on [PyPI](https://pypi.org/project/diskmelts/) and
+requires Python 3.9 or newer. It bundles 46 production checkpoints, 23 each
+for the sigma=0 and sigma=2 grids, covering the same 18 species (see
+[Available molecules](#available-molecules)).
+Install it from any directory. These commands use
 `python` in an activated environment; the
 [installation guide](https://diskmelts.readthedocs.io/en/latest/installation.html)
 shows how to create one on macOS, Linux, Windows, or Conda:
@@ -48,6 +49,36 @@ conda activate diskmelts
 Use `environment-notebooks.yaml` or `environment-dev.yaml` for the corresponding
 notebook or contributor dependencies.
 
+## Minimal workflow
+
+Subtract the continuum, fit one molecule, and print the result:
+
+```python
+from diskmelts import consub, fit_stage, load_fitting_models
+
+# 1. Raw spectrum: CSV with wave (µm), flux (Jy), and optional err columns.
+raw_wave, raw_flux, raw_err = consub.read_spectrum("my_raw_spectrum.csv")
+wave, flux, err, cont, cont_unshifted = consub.calculate(
+    raw_wave, raw_flux, raw_err, preset="ttauri",
+)
+
+# 2. Fit C2H2 near 13.7 µm with the bundled surrogate model.
+models = load_fitting_models("C2H2")
+fit = fit_stage(
+    wave, flux, molecule="C2H2", pretrained=models,
+    wavelength_range=(13.0, 15.0), fit_ranges=(13.5, 14.1),
+    distance_pc=140.0, n_samples=1024, n_refine=4, n_top=4,
+)
+
+# 3. Best-fit T (K), logN (cm^-2), A (au^2), and the spread of top solutions.
+print(fit["params"]["C2H2"])
+print(fit["uncertainty"]["C2H2"])
+```
+
+The sections below explain each step, and the
+[first-spectrum guide](https://diskmelts.readthedocs.io/en/latest/quickstart.html)
+covers multi-stage fits, saved outputs, and plots.
+
 ## Runtime dependencies
 
 | Package | Purpose |
@@ -56,7 +87,7 @@ notebook or contributor dependencies.
 | `numpy` | Array operations |
 | `pandas` | CSV I/O |
 | `scipy` | L-BFGS-B optimizer, NNLS, Sobol sampling |
-| `scikit-learn` | Model scaling and PCA metadata |
+| `scikit-learn` | PCA and scaling when training new models |
 | `matplotlib` | Plotting |
 
 ## Continuum subtraction
@@ -91,12 +122,6 @@ With `preset=None, linefreefile=None` (or `""`), fitting uses all nonexcluded
 pixels and skips the line-free correction, so `cont` equals `contOrig`.
 Reference tables are bundled; normal continuum subtraction needs no model grids.
 
-The private `0-Consub/` experiment directory is excluded from Git and is not
-required by the installed package. Its runtime code and reference tables are
-maintained under `src/diskmelts/consub/`.
-The distributed tests in `tests/` exercise the installed continuum API and
-bundled references without that private directory.
-
 Returned `err` represents measurement/noise error, not correlated continuum
 uncertainty. Broad excluded intervals require inspection. See the
 [continuum guide](https://diskmelts.readthedocs.io/en/latest/continuum.html) for presets, regional box sizes,
@@ -143,31 +168,35 @@ You can also read spectra with your own code. DiskMELTS only requires
 one-dimensional `obs_wav` and `obs_flux` arrays. Replace the sample CSV with
 your own file to fit your source; update `distance_pc` and fit intervals.
 
-For a complete staged fit-and-subtract workflow from a checkout, edit and run:
+For a complete staged fit-and-subtract workflow from a clone of the
+repository, run:
 
 ```bash
 python examples/dev_v2_realobs.py
+python examples/dev_v2_realobs.py --input /path/to/my_spectrum.csv --name my_source --distance-pc 140 --sigma 2
 ```
 
+Adjust that script's stage windows for your spectrum.
+`python examples/synthetic_array_fit.py` generates an HCN spectrum and fits it
+from in-memory arrays, without an observation file.
+
 The v2 fitting API is `load_fitting_models`, `fit_stage`, `fit_observation`, and
-`plot_saved_observation`. The production checkpoints are packaged under
-`src/diskmelts/models/` and included in the wheel. Model fitting requires no
-`Model_grids/` or `Pretrain_grid/` directory. `diskmelts.MODEL_SPECS` records
-PCA counts, network layers, and `grid_sigma=0` for the original checkpoints.
-In version 2.2.0, sigma=2 checkpoints are available for H2O,
-C2H2, HCN, CO2, 13C12CH2, 13CO2, C4H2, HC3N, C2H6, C2H4, CH4,
-C2HD, CO17O, CO18O, C3H4, CH3, C6H6, and OH. Select them with
-`load_fitting_models(['H2O', 'C2H2'], sigma=2)` or
-`fit_observation(..., sigma=2)`. The C2H2 bank covers 6–9 and 11.5–17 µm;
-choose fit windows within those bands. OH covers 15–26 µm. Both C6H6
-checkpoints cover 13–17 µm and 100–675 K; the sigma=2 checkpoint is selected
-with `load_fitting_models('C6H6', sigma=2)`.
-The checkpoint settings are checked when
-loading. The checkpoint flux convention is 1 au² at 140 pc; `distance_pc`
-sets the source distance and makes the returned emitting area physical.
-The [fitting notebook](https://github.com/wddlx/DiskMELTS/blob/main/notebooks/Example_Fitting.ipynb) requires 13C12CH2 to
-be no hotter than C2H2 and gives them the same emitting area. It ties 13CO2
-to CO2 in both temperature and area. Set
+`plot_saved_observation`. The production checkpoints are included in the
+wheel, so model fitting needs no local model grids. The checkpoint flux
+convention is 1 au² at 140 pc; `distance_pc` sets the source distance and
+makes the returned emitting area physical. `fit_observation` accepts either a
+CSV path or in-memory `obs_wav`, `obs_flux`, and optional `obs_err` arrays
+(new in 2.2.1).
+
+For a synthetic spectrum, `wave, flux = diskmelts.synthesize_spectrum('HCN',
+T=300, logN=18, A=1)` returns the native wavelength axis and model flux (new
+in 2.2.1). See the
+[generation guide](https://diskmelts.readthedocs.io/en/latest/generate_spectra.html).
+
+The [fitting notebook](https://github.com/wddlx/DiskMELTS/blob/main/notebooks/Example_Fitting.ipynb)
+requires 13C12CH2 to be no hotter than C2H2 and gives them the same emitting
+area; it ties 13CO2 to CO2 in both temperature and area. Point its input at
+your own continuum-subtracted CSV, or pass arrays. Set
 `isotopologue_groups`, `Tshare` (`free`, `equal`, or `less`), and `Ashare` in a
 joint stage to choose those constraints. See the
 [fitting guide](https://diskmelts.readthedocs.io/en/latest/quickstart.html) for examples.
@@ -179,6 +208,32 @@ separate method does require the full local grids. See the
 plotting.
 For complete sigma=2 grids, pass `grid_root='Model_grids/sigma2'`.
 
+## Available molecules
+
+Each molecule is available for two slab-model grids: sigma=0 (default) assumes
+no turbulent broadening, so lines are only thermally broadened, and sigma=2
+assumes 2 km/s turbulent broadening. Select sigma=2 with `load_fitting_models(..., sigma=2)` or
+`fit_observation(..., sigma=2)`. Fit windows must lie inside a model's band.
+
+| Molecule | Band (µm) | T (K) | logN (cm⁻²) |
+|---|---|---|---|
+| H2O | 4.9–25.0 (five overlapping tiles) | 100–1400 | 13–19 |
+| C2H2 | 6–9 and 11.5–17 | 100–1400 | 13–19.5 |
+| HCN, CO2, 13C12CH2, 13CO2 | 11.5–17 | 100–1400 | 13–19.5 |
+| HC3N, CO17O, CO18O | 13–18 | 100–1400 | 13–19.5 |
+| C4H2 | 14–18 | 100–1400 | 13–19.5 |
+| C3H4, CH3 | 15–18 | 100–1400 | 13–19.5 |
+| C6H6 | 13–17 | 100–675 | 13–19.5 |
+| C2H6 | 9–15 | 100–1400 | 13–19.5 |
+| C2H4 | 8–14 | 100–1400 | 13–19.5 |
+| CH4 | 5–9 | 100–1400 | 13–19.5 |
+| C2HD | 11.5–20.5 | 100–1400 | 13–19.5 |
+| OH | 15–26 | 100–1400 | 13–19.5 |
+
+`diskmelts.MODEL_SPECS` records the exact domains, PCA sizes, and network
+layers. Version 2.2.0 lacks the sigma=0 OH and 6–9 µm C2H2 checkpoints;
+upgrade with `python -m pip install -U diskmelts`.
+
 ## Training and validation data
 
 Generated data under `Model_grids/` and `Pretrain_grid/` are intentionally
@@ -211,7 +266,7 @@ uses the same names with local `Model_grids/sigma0/C6H6/` and `Model_grids/sigma
 files. See the [training guide](https://diskmelts.readthedocs.io/en/latest/training.html)
 for peak audits and retraining commands.
 
-Version 2.2.0 also includes a CH3 checkpoint for 15–18 µm,
+The package also includes a CH3 checkpoint for 15–18 µm,
 `T=100–1400 K`, and `logN=13–19.5`. Its strongest grid feature in this window
 is near 16.4–16.6 µm. Use `load_fitting_models('CH3')` after installing this
 package; direct-grid fitting uses local `Model_grids/sigma0/CH3/` files.

{diskmelts-2.2.0 → diskmelts-2.2.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "diskmelts"
-version = "2.2.0"
+version = "2.2.2"
 description = "Machine-learning Enabled Line fitting Tool for Spectra of protoplanetary disks"
 readme = "README.md"
 requires-python = ">=3.9"
@@ -21,13 +21,15 @@ classifiers = [
     "Topic :: Scientific/Engineering :: Astronomy",
 ]
 dependencies = [
-    "torch",
+    # Checkpoints load with torch.load(weights_only=True), added in 1.13.
+    "torch>=1.13",
     # Intel macOS wheels stop at PyTorch 2.2, whose NumPy bridge needs NumPy 1.x.
     "numpy>=1.26,<2; sys_platform == 'darwin' and platform_machine == 'x86_64'",
     "numpy; sys_platform != 'darwin' or platform_machine != 'x86_64'",
     "pandas",
     "scipy",
-    "scikit-learn",
+    # Used only to fit PCA/scalers when training; checkpoints store plain arrays.
+    "scikit-learn>=1.0",
     "matplotlib",
 ]
 

{diskmelts-2.2.0 → diskmelts-2.2.2}/src/diskmelts/__init__.py

@@ -12,6 +12,7 @@ from diskmelts.trainmodel import (
     train_model_bank,
     select_grid_holdout,
     tune_pca_components,
+    convert_legacy_checkpoint,
 )
 from diskmelts.fitting import (
     load_models,
@@ -41,6 +42,7 @@ from diskmelts.v2_fitting import (
     MODEL_SPECS,
     SIGMA2_MODEL_SPECS,
     load_fitting_models,
+    synthesize_spectrum,
     fit_stage,
     fit_observation,
 )
@@ -61,6 +63,7 @@ __all__ = [
     "MODEL_SPECS",
     "SIGMA2_MODEL_SPECS",
     "load_fitting_models",
+    "synthesize_spectrum",
     "fit_stage",
     "fit_observation",
     "fit_grid_stage",
@@ -76,6 +79,7 @@ __all__ = [
     "train_model_bank",
     "select_grid_holdout",
     "tune_pca_components",
+    "convert_legacy_checkpoint",
     # fitting
     "load_models",
     "load_model_bank",

{diskmelts-2.2.0 → diskmelts-2.2.2}/src/diskmelts/consub/_core.py

@@ -41,6 +41,8 @@ def _regions(regions):
     a = np.asarray(regions, dtype=float)
     if a.size == 0:
         return np.empty((0, 2))
+    if a.shape == (2,):
+        a = a[np.newaxis]
     if a.ndim != 2 or a.shape[1] != 2 or not np.all(np.isfinite(a)) or np.any(a[:, 1] <= a[:, 0]):
         raise ValueError("Regions must be finite (xmin, xmax) pairs with xmin < xmax")
     return a
@@ -128,11 +130,21 @@ def _mad(a):
     return 1.4826 * np.median(np.abs(a - np.median(a)))
 
 
-def _interp(x, xp, yp):
+def _interp(x, xp, yp, linear_ends=False):
     if len(xp) == 1:
         return np.full(len(x), yp[0])
     # Constant end corrections avoid unstable polynomial extrapolation.
-    return PchipInterpolator(xp, yp)(np.clip(x, xp[0], xp[-1]))
+    out = PchipInterpolator(xp, yp)(np.clip(x, xp[0], xp[-1]))
+    if linear_ends:
+        # A sloped continuum beyond the last knot needs a slope, not a constant;
+        # fit it over about 1 um of knots so one noisy knot cannot set it.
+        xp, yp = np.asarray(xp), np.asarray(yp)
+        for outside, end, near in ((x < xp[0], 0, xp <= xp[0] + 1.0),
+                                   (x > xp[-1], -1, xp >= xp[-1] - 1.0)):
+            if outside.any() and near.sum() >= 2:
+                slope = np.polyfit(xp[near], yp[near], 1)[0]
+                out[outside] = yp[end] + slope * (x[outside] - xp[end])
+    return out
 
 
 def _widths(x, good, boxsizes):
@@ -143,7 +155,7 @@ def _widths(x, good, boxsizes):
         n = np.convolve(good.astype(float), np.ones(min(101, len(x))), "same")
         fraction = np.maximum(n / min(101, len(x)), 0.05)
         return np.clip(31 * dx / fraction, 0.25, 1.0)
-    if np.isscalar(boxsizes):
+    if np.ndim(boxsizes) == 0:
         if not np.isfinite(boxsizes) or boxsizes < 5:
             raise ValueError("boxsizes must be at least 5 pixels")
         return dx * float(boxsizes)
@@ -188,7 +200,7 @@ def _local_baseline(x, y, noise, good, widths):
     if len(positions) < 3:
         raise ValueError("Too few continuum windows; widen boxsizes or revise masks")
     knots = pd.DataFrame({"x": positions, "y": values}).groupby("x", sort=True).median()
-    return _interp(x, knots.index.to_numpy(), knots.y.to_numpy())
+    return _interp(x, knots.index.to_numpy(), knots.y.to_numpy(), linear_ends=True)
 
 
 def _arpls(x, y, noise, good, widths):
@@ -203,10 +215,12 @@ def _arpls(x, y, noise, good, widths):
     weights = base.copy()
     converged = False
     for _ in range(40):
-        z = spsolve(diags(weights) + penalty, weights*y)
+        z = spsolve((diags(weights) + penalty).tocsc(), weights*y)
         residual = y-z
         negative = residual[good & (residual < 0)]
         if len(negative) < 3:
+            # Nothing left below the baseline to down-weight: a fixed point.
+            converged = True
             break
         scale = max(np.std(negative), np.median(noise[good])*0.1)
         new = base * expit(-2*(residual - (2*scale-np.mean(negative)))/scale)
@@ -354,8 +368,11 @@ def subtract_continuum(wav, flux, *, error=None, exclude=None, boxsizes=None,
     unsupported |= (x < x[reference][0]) | (x > x[reference][-1])
     cont_error[unsupported] = np.nan
     if unsupported.any():
-        warnings.warn("Continuum crosses regions with weak support; inspect the plot or "
-                      "subtract_continuum() diagnostics before using those regions", stacklevel=2)
+        edges = np.flatnonzero(np.diff(np.r_[0, unsupported.astype(int), 0]))
+        spans = ", ".join(f"{x[a]:.2f}-{x[b-1]:.2f}" for a, b in zip(edges[::2], edges[1::2]))
+        warnings.warn(f"Continuum is interpolated or extrapolated without nearby reference "
+                      f"pixels at {spans} um (weak_support column); inspect the plot before "
+                      "using those regions", stacklevel=2)
     result = pd.DataFrame({"wave": x, "flux": original-continuum,
                            "err": measurement, "cont": continuum,
                            "cont_unshifted": baseline, "raw_flux": original,