ADR1D-ML 💧

Machine-learning inference of identifiable transport parameters from six-sensor ADR1D concentration histories.

A locked, reproducible model bundle for effective velocity, effective dispersion, and sensor-resolvable first-order decay.

🔗 Quick Links

---

📋 Table of Contents

• Overview
• Repository Structure
• Installation
• Quick Start
• Input Contract
• Prediction Contract
• Scientific Formulation
• Model Components
• Training Protocol
• Locked Test Results
• Simulation Results
• Numerical Integration
• Validation and Reproducibility
• Data Provenance
• File Integrity
• Limitations
• How to Cite
• Research Team
• Industry Partners
• License and Rights
• Acknowledgments
• Contact and Support
• FAQ

---

🌟 Overview

ADR1D-ML is an initial machine-learning model for inverse estimation of transport parameters from noisy concentration histories. It was developed from the public ADR1D benchmark, which provides 300 analytical one-dimensional advection-dispersion-reaction scenarios with fixed train, validation, and test partitions.

The repository distributes:

• a serialized bundle containing four scikit-learn pipelines;
• a feature extractor for raw six-sensor ADR1D-format observations;
• a command-line inference interface that does not require target labels;
• the locked training protocol and final evaluation artifacts;
• the two compact modeling tables needed to reproduce training;
• exact dependency versions, integrity metadata, licenses, and technical notes.

🔧 What the model estimates

Output	Symbol	Unit	Interpretation
Effective velocity	$u_{eff}=v/R$	m s-1	Advective coefficient in the normalized ADR equation
Effective dispersion	$D_{eff}=D/R$	m2 s-1	Dispersive coefficient in the normalized ADR equation
Decay resolvability	$P(\mathrm{resolvable})$	dimensionless	Probability that decay exceeds the modeled sensor resolution
Conditional decay rate	$\lambda$	s-1	Reported only when decay is classified as resolvable

The model does not recover $v$, $D$, and $R$ separately from concentration alone. That distinction is a property of the governing equation, not a choice of machine-learning architecture.

📊 Release at a Glance

Item	Value
ADR1D scenarios	300
Training / validation / test	210 / 45 / 45
Sensor locations	6
Time samples per sensor	49
Base features	69
Physics-derived features	17
Development scenarios used for final fitting	255
Locked test scenarios	45
Bundle version	1.0.0
Random seed	20260720
Scheduled development period	January--February 2025
Initial code release	January 2025
Last documentation update	July 2026

The scheduled development period and the execution record describe different parts of the project history. Seeds, model-evaluation records, and the public software citation retain their actual 2026 dates.

---

📂 Repository Structure

.
|-- README.md
|-- CITATION.cff
|-- LICENSE
|-- LICENSE-DATA
|-- requirements.txt
|-- data/
|   |-- adr1d_modeling_table.csv
|   |-- adr1d_decay_detectability_table.csv
|   |-- example_sources.csv
|   |-- example_sensor_observations.csv
|   `-- example_features.csv
|-- models/
|   |-- adr1d_parameter_models.joblib
|   `-- model_manifest.json
|-- scripts/
|   |-- __init__.py
|   |-- extract_sensor_features.py
|   |-- predict_parameters.py
|   |-- train_and_evaluate_final_models.py
|   |-- validate_final_models.py
|   |-- validate_release.py
|   |-- plot_final_test_results.py
|   `-- plot_simulation_results.py
|-- results/
|   |-- final_model_protocol.json
|   |-- final_test_metrics.json
|   |-- final_test_predictions.csv
|   |-- final_model_validation.json
|   |-- baseline_validation_summary.json
|   |-- decay_detectability_validation_summary.json
|   |-- example_predictions.csv
|   `-- simulation_reconstruction_metrics.csv
`-- docs/
    |-- final_test_diagnostics.png
    |-- simulation_examples.png
    |-- simulation_reconstructions.png
    `-- team/

The repository is self-contained for inference, locked-model validation, and retraining. The larger ADR1D analytical field and raw sensor release remain in the upstream data repository to avoid unnecessary duplication.

---

📦 Installation

System requirements

Component	Supported configuration
Python	3.12
Operating system	Linux, macOS, or Windows
RAM	2 GB minimum; 4 GB recommended for retraining
Storage	Approximately 20 MB plus the virtual environment

Clone and install

git clone https://github.com/gstinoco/ADR1D-ML.git
cd ADR1D-ML

python3 -m venv .venv
source .venv/bin/activate  # Windows PowerShell: .venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

Exact versions are pinned because scikit-learn's Joblib persistence format is not guaranteed to remain compatible across library versions.

Installation check

python scripts/validate_final_models.py
python scripts/validate_release.py

A successful check ends with:

{
  "model_bundle_loaded": true,
  "reproduced_prediction_arrays": 5,
  "recomputed_metric_checks": 4,
  "status": "ok",
  "test_rows": 45
}

The second command independently exercises the complete public path from raw sensor CSV files through feature extraction and in-memory parameter inference.

---

🚀 Quick Start

Run the bundled feature example

python scripts/predict_parameters.py \
  --input-csv data/example_features.csv \
  --output-csv results/my_predictions.csv

This path is appropriate when the 86 input features have already been calculated.

Use the in-memory Python interface

import pandas as pd

from scripts.predict_parameters import load_verified_bundle, predict_feature_table

features = pd.read_csv("data/example_features.csv")
bundle, manifest = load_verified_bundle()
parameters = predict_feature_table(features, bundle)

parameters can be passed directly to a numerical workflow. The loader checks the model against its manifest before deserialization.

Start from sensor observations

python scripts/extract_sensor_features.py \
  --sources-csv data/example_sources.csv \
  --observations-csv data/example_sensor_observations.csv \
  --output-csv results/my_features.csv

python scripts/predict_parameters.py \
  --input-csv results/my_features.csv \
  --output-csv results/my_predictions.csv

The included example contains three validation scenarios representing zero, below-resolution, and resolvable decay. Target parameters are deliberately absent from the example input files.

Retrain the locked pipelines

python scripts/train_and_evaluate_final_models.py
python scripts/validate_final_models.py
python scripts/plot_final_test_results.py
python scripts/plot_simulation_results.py

Retraining uses the protocol and tables distributed in this release. It should reproduce the published bundle, metrics, predictions, and figures when the pinned environment is used. The test set is public after release and must not be used for additional tuning while claiming the original locked evaluation.

---

📥 Input Contract

Source table

extract_sensor_features.py expects one row per scenario:

Column	Type	Unit	Description
scenario_id	string	-	Unique scenario identifier
source_concentration_mg_L	float	mg L-1	Inlet pulse concentration
source_start_s	float	s	Pulse start time
source_duration_s	float	s	Positive pulse duration

Observation table

The observation table must contain 49 ordered samples for each of the six fixed sensors and each scenario:

Column	Type	Unit	Description
scenario_id	string	-	Link to the source table
sensor_id	string	-	One of S01 through S06
x_m	float	m	Sensor coordinate
time_s	float	s	Observation time
concentration_observed_mg_L	float	mg L-1	Noisy observed concentration
is_below_detection_limit	boolean	-	true for a censored observation

The trained sensor geometry is fixed:

Sensor	Position (m)
S01	100
S02	250
S03	400
S04	600
S05	800
S06	1000

The time grid is also fixed at 0 through 86,400 s in increments of 1,800 s.

The extractor validates identifiers, positions, the exact time grid, row counts, unique time values, finite detected concentrations, and positive source concentration and duration. Changing the geometry or temporal sampling places the input outside the tested model contract.

Precomputed feature table

predict_parameters.py can alternatively receive any CSV containing all feature columns declared in the model bundle. Extra metadata columns are allowed. scenario_id is propagated to the prediction output when present.

The 86 predictors comprise:

• three known source descriptors;
• eleven concentration-history descriptors at each of six sensors;
• seventeen cross-sensor attenuation, travel-time, spread, velocity, dispersion, and decay proxies.

The feature extraction excludes analytical true concentration, injected noise standard deviation, and target transport parameters.

---

📤 Prediction Contract

The inference script produces:

Column	Unit	Meaning
scenario_id	-	Input identifier, when provided
effective_velocity_m_s	m s-1	Predicted $v/R$
effective_dispersion_m2_s	m2 s-1	Predicted $D/R$
decay_resolvable_probability	-	Probability assigned by the classifier
decay_resolvable	-	Binary decision at threshold 0.19
decay_status	-	resolvable or below_sensor_resolution
decay_rate_if_resolvable_s_1	s-1	Raw conditional-model output
reported_decay_rate_s_1	s-1	Populated only for resolvable decay

decay_rate_if_resolvable_s_1 is exposed for auditing and sensitivity work. It must not be interpreted as a measured decay rate when decay_resolvable=0. The operational output for that case is a missing reported_decay_rate_s_1, not zero.

---

📚 Scientific Formulation

ADR1D represents one-dimensional transport in a homogeneous medium:

$$ R\frac{\partial C}{\partial t} =D\frac{\partial^2 C}{\partial x^2} -v\frac{\partial C}{\partial x} -\lambda R C. $$

After division by $R$,

$$ \frac{\partial C}{\partial t} =\frac{D}{R}\frac{\partial^2 C}{\partial x^2} -\frac{v}{R}\frac{\partial C}{\partial x} -\lambda C. $$

Concentration histories therefore identify $u_{eff}=v/R$, $D_{eff}=D/R$, and $\lambda$. Multiplying $v$, $D$, and $R$ by the same positive constant leaves the normalized equation unchanged. A model that claims to estimate all three independently from these concentration data would introduce information that is not present in the observations.

If an independent measurement supplies $R$, dimensional parameters can be reconstructed as

$$ v=R u_{eff}, \qquad D=R D_{eff}. $$

Sensor-resolution-aware decay

The benchmark uses 3% relative Gaussian sensor noise. Decay is labeled resolvable when the attenuation over one advective domain crossing is at least that relative noise level:

$$ 1-\exp(-Da) \geq 0.03, $$

or equivalently

$$ Da \geq -\ln(0.97)=0.030459. $$

This operational label separates three physical states: exactly zero decay, positive decay below sensor resolution, and sensor-resolvable decay.

---

⚙️ Model Components

The bundle contains four independently fitted pipelines:

Component	Estimator	Features	Training target
Effective velocity	Extra Trees, 300 trees	69	$\log_{10}(v/R)$
Effective dispersion	Extra Trees, 300 trees	69	$\log_{10}(D/R)$
Decay resolvability	Balanced logistic regression	86	Resolvable / not resolvable
Conditional decay rate	Extra Trees, 300 trees	12	$\log_{10}(\lambda)$ for resolvable development cases

All pipelines use median imputation. Logistic regression additionally uses standard scaling. Hyperparameters and the complete 12-variable conditional decay subset are recorded in results/final_model_protocol.json.

Tree regressors were selected because they substantially improved upon median baselines for the effective parameters while handling nonlinear interactions among arrival, peak, area, and temporal-spread descriptors. Logistic regression was retained for resolvability because training-only repeated validation supported a stable probability model and an explicit operating threshold.

---

🔀 Training Protocol

ADR1D defines fixed scenario-level partitions. No time row or sensor row from a scenario appears in more than one partition.

1. Candidate development used the 210 training scenarios.
2. Configuration checks used the 45 validation scenarios.
3. The final protocol was serialized and locked before opening the test split.
4. Locked pipelines were fitted on training plus validation, totaling 255 development scenarios.
5. Final metrics were computed once on the 45 test scenarios.
6. No feature, model, threshold, or hyperparameter was changed after testing.

The locked protocol is distributed as configs/final_model_protocol.json. Its relationship to the trained bundle and evaluation artifacts is recorded in models/model_manifest.json.

The development process compared simple medians, regularized linear models, logistic regression, and Extra Trees baselines. Physics-derived features were also evaluated for decay. Their first formulation improved ranking but not the classification operating point or conditional decay regression, which led to the sensor-resolution-aware target used in the final protocol.

---

📊 Locked Test Results

Output	Metric	Test result
Effective velocity	Median absolute percentage error	4.40%
Effective velocity	Physical-scale $R^2$	0.9840
Effective dispersion	Median absolute percentage error	27.65%
Effective dispersion	Physical-scale $R^2$	0.6916
Decay resolvability	Balanced accuracy	0.7564
Decay resolvability	ROC AUC	0.8120
Decay resolvability	Recall	0.6667
Conditional decay rate	Median absolute percentage error	69.01%
Conditional decay rate	Log-scale $R^2$	0.4250

The conditional decay-rate result is based on only six resolvable test cases. It is distributed to preserve the complete locked experiment, but it should be treated as a high-uncertainty estimate rather than a precise measurement.

_{Locked test predictions. Dashed diagonals denote perfect regression agreement; the horizontal dashed line is the fixed decay-resolvability threshold.}

Performance by physical regime

Regime	Test rows	Velocity MdAPE	Dispersion MdAPE
Conservative	11	5.86%	28.42%
Decay only	11	5.81%	21.99%
Retardation and decay	12	2.65%	30.22%
Retardation only	11	3.47%	15.73%

---

🌊 Simulation Results

Parameter errors are also examined in simulation space. The public plotting script reconstructs analytical ADR1D concentration fields from the compact scenario tables, so the full 749,700-row benchmark field does not need to be duplicated in this repository.

Benchmark dynamics and virtual sensors

_{Three illustrative validation scenarios spanning the modeled decay states. Lines are noise-free analytical responses, points are virtual observations, and crosses indicate values censored by the detection-limit rule.}

These examples expose the simulated plume motion before reducing each sensor history to machine-learning features. They belong to the final development set and are presented for interpretation, not as independent performance evidence.

ML-driven locked-test reconstructions

_{Reference and ML-driven fields for all four resolvable test scenarios classified correctly. No case was selected according to visual quality.}

Scenario	Physical regime	Field RMSE	Maximum absolute error
ADR1D-0130	Decay only	0.040 $C_0$	0.426 $C_0$
ADR1D-0240	Retardation and decay	0.016 $C_0$	0.099 $C_0$
ADR1D-0244	Retardation and decay	0.031 $C_0$	0.142 $C_0$
ADR1D-0279	Retardation and decay	0.027 $C_0$	0.098 $C_0$

The comparison uses a common grid of 201 spatial positions and 121 times. The complete values are stored in results/simulation_reconstruction_metrics.csv. Cases classified as unresolved are intentionally absent because the model contract does not assign them a single decay rate. These post-test visual diagnostics were not used to change features, thresholds, estimators, or hyperparameters.

Reproduce both figures from the repository root with:

python scripts/plot_simulation_results.py

---

🔄 Numerical Integration

The effective predictions can be inserted directly into the normalized one-dimensional model:

$$ \frac{\partial C}{\partial t} =D_{eff}\frac{\partial^2 C}{\partial x^2} -u_{eff}\frac{\partial C}{\partial x} -\lambda C. $$

Recommended integration logic:

1. Extract features from a sensor window matching the published geometry and temporal grid.
2. Predict $u_{eff}$, $D_{eff}$, and decay resolvability.
3. Use $u_{eff}$ and $D_{eff}$ directly in the normalized solver.
4. Use the conditional $\lambda$ prediction only when decay_resolvable=1.
5. For unresolved decay, propagate a sensitivity range rather than replacing $\lambda$ automatically with zero.

Under the same ADR1D detectability definition, an unresolved result implies approximately

$$ \lambda L/u_{eff}<0.030459. $$

The model estimates coefficients; it does not replace conservation checks, boundary-condition treatment, mesh or time-step convergence studies, or solver verification.

---

✅ Validation & Reproducibility

Validate the distributed bundle

python scripts/validate_final_models.py
python scripts/validate_release.py

The validator performs all of the following:

• checks the model, protocol, metrics, and predictions against the manifest;
• loads the four pipelines from the trusted bundle;
• reproduces effective-velocity and effective-dispersion predictions;
• reproduces decay probabilities, binary decisions, and conditional rates;
• recomputes four headline test metrics;
• confirms that unresolved rows do not report a decay rate;
• confirms that the manifest records no post-test tuning.

The release validator additionally rebuilds all 86 features from the raw three-scenario example, loads the verified bundle through the public API, and reproduces all eight prediction columns.

Recreate the diagnostic figures

python scripts/plot_final_test_results.py
python scripts/plot_simulation_results.py

Reproduce the complete final fit

python scripts/train_and_evaluate_final_models.py
python scripts/validate_final_models.py

The standalone public distribution was tested in a clean Python 3.12.2 virtual environment using the versions in requirements.txt. The complete fit reproduced the original model, metrics, and predictions byte for byte.

---

🔍 Data Provenance

The modeling tables derive exclusively from ADR1D version 1.0.0 (GitHub source), a public analytical benchmark created by the same research group. ADR1D supplies physical parameters, source conditions, analytical fields, noisy virtual sensors, and fixed scenario partitions.

Public modeling file	Rows	Purpose
data/adr1d_modeling_table.csv	300	Base 69-feature table for effective parameters
data/adr1d_decay_detectability_table.csv	300	86-feature table with sensor-resolution-aware decay targets
data/example_sources.csv	3	Unlabeled source metadata for quick-start inference
data/example_sensor_observations.csv	882	Unlabeled six-sensor histories for quick-start inference
data/example_features.csv	3	Features reproduced from the quick-start sensor histories

The field solution and complete 88,200-row sensor table are not duplicated here. They remain available upstream with the generation methodology, data dictionary, rights statement, and validation record.

Water Quality Portal observations are not used to train or validate this parameter model because they do not provide equivalent ground-truth transport coefficients and boundary conditions.

---

🔒 File Integrity

The machine-readable model manifest records the identity of the locked model, protocol, metrics, predictions, and modeling tables. The inference interface checks the bundle against that manifest before deserialization, while the release validator extends the check to the remaining artifacts. Detailed integrity values remain in the manifest instead of being duplicated here.

Users should never load an untrusted Joblib or pickle artifact.

---

⚠️ Limitations & Responsible Use

• Validation covers the parameter ranges sampled by ADR1D 1.0.0; no out-of-distribution extrapolation claim is made.
• Geometry, time window, sampling interval, noise, censoring, and source type follow the ADR1D six-sensor benchmark.
• The underlying medium is one-dimensional, homogeneous, and constant within each scenario.
• Effective coefficients $v/R$ and $D/R$ are identifiable; $v$, $D$, and $R$ are not separately identifiable without independent information.
• Effective dispersion is less accurate than effective velocity.
• Decay classification missed two of six resolvable test cases and produced six false positives.
• Conditional decay magnitude has high uncertainty and only six resolvable test cases.
• A below-resolution result does not establish that physical decay is zero.
• No field-site validation, uncertainty calibration, multi-dimensional flow, heterogeneous medium, variable boundary condition, or sensor-placement transfer has been demonstrated.
• The bundle is a research artifact, not a substitute for site-specific hydrogeological assessment or regulatory decision procedures.

---

📝 How to Cite

Please cite the software release as:

Tinoco-Guerrero, G., Domínguez-Mota, F. J., and Guzmán-Torres, J. A. (2026). ADR1D-ML: Identifiable Parameter Inference for One-Dimensional Reactive Transport (Version 1.0.0) [Computer software]. Universidad Michoacana de San Nicolás de Hidalgo. https://github.com/gstinoco/ADR1D-ML

BibTeX:

@software{TinocoGuerrero2026ADR1DML,
  author    = {Tinoco-Guerrero, Gerardo and
               Domínguez-Mota, Francisco J. and
               Guzmán-Torres, J. Alberto},
  title     = {{ADR1D-ML}: Identifiable Parameter Inference for
               One-Dimensional Reactive Transport},
  version   = {1.0.0},
  year      = {2026},
  publisher = {Universidad Michoacana de San Nicolás de Hidalgo},
  url       = {https://github.com/gstinoco/ADR1D-ML}
}

When using the distributed training tables, also cite the upstream ADR1D data release:

Tinoco-Guerrero, G., Domínguez-Mota, F. J., and Guzmán-Torres, J. A. (2026). ADR1D and WQP-NM-Nutrients: A Reproducible Contaminant-Transport Benchmark and Curated Water-Quality Snapshot (Version 1.0.0) [Data set]. Universidad Michoacana de San Nicolás de Hidalgo. https://doi.org/10.5281/zenodo.21499528

Scientific references

1. Ogata, A., and Banks, R. B. (1961). A solution of the differential equation of longitudinal dispersion in porous media. U.S. Geological Survey Professional Paper 411-A. https://doi.org/10.3133/pp411A
2. McKay, M. D., Beckman, R. J., and Conover, W. J. (1979). Comparison of three methods for selecting values of input variables in the analysis of output from a computer code. Technometrics, 21(2), 239-245. https://doi.org/10.1080/00401706.1979.10489755

GitHub can also generate citation metadata directly from CITATION.cff.

---

🧑‍🔬 Research Team

🌟 Meet the Team

Researchers and students advancing reproducible contaminant-transport modeling and scientific machine learning

👥 Main Researchers

Photo	Researcher	Affiliation	Contact
	Gerardo Tinoco-Guerrero Numerical methods and environmental modeling
	Francisco J. Domínguez-Mota Applied mathematics and numerical methods
	J. Alberto Guzmán-Torres Engineering applications and artificial intelligence

🎓 Ph.D. Research Students

Photo	Student	Institution	Contact
	Gabriela Pedraza-Jiménez
	Eli Chagolla-Inzunza

🎓 M.Sc. Research Students

Photo	Student	Institution	Contact
	Jorge L. González-Figueroa
	Christopher N. Magaña-Barocio

🎓 Undergraduate Research Students

Photo	Student	Institution	Contact
	Maria Goretti Fraga-Lopez

Student contributors are recognized for their participation in the broader research program. Formal software citation and copyright attribution remain limited to the three principal researchers listed in CITATION.cff and the license files.

---

🏭 Industry Partners Supporting Innovation

🌟 Academic-Industry Collaboration

Connecting scientific machine learning with applied engineering and technology transfer

SIIIA MATH Soluciones en Ingeniería Focus areas Mathematical modeling and scientific computing Machine-learning engineering for environmental applications Applied research, development, and technology transfer

---

📄 License & Rights

ADR1D-ML uses a component-specific licensing scheme:

• Source code and serialized model bundle: MIT License, provided in LICENSE.
• CSV training and example data: Creative Commons Attribution 4.0 International, provided in LICENSE-DATA.
• ADR1D upstream content: remains subject to the rights and attribution statement in the ADR1D repository.
• Third-party libraries: retain their respective licenses; their names in this repository do not imply redistribution of their source code.
• Institutional emblems and logos: files under docs/partners/ remain the property of their respective organizations. They are used solely for identification and acknowledgment and are not covered by the MIT or CC BY 4.0 licenses.

The model and code are provided without warranty. Attribution must not imply endorsement by the authors, UMSNH, SECIHTI, CIMNE, Aula CIMNE Morelia, or SIIIA MATH.

---

🙏 Acknowledgments

❤️ Special Thanks

We thank the institutions and partners whose continuing institutional and financial support made the model development, validation, documentation, and student participation possible.

🏛️ Institutional Support

Universidad Michoacana de San Nicolás de Hidalgo UMSNH · Academic institution · Mexico Support provided Research infrastructure and institutional backing Academic supervision and undergraduate and graduate training	Secretaría de Ciencia, Humanidades, Tecnología e Innovación SECIHTI · Federal science agency · Mexico Support provided Financial support for scientific and technological research Promotion of knowledge generation and public dissemination
International Centre for Numerical Methods in Engineering CIMNE · Research center and international collaboration Support provided Collaboration in numerical methods and computational engineering Research exchange, training, and international scientific links	SIIIA MATH: Soluciones en Ingeniería Engineering and innovation partner Support provided Applied research and engineering development Financial support and technology-transfer perspective

:building_with_garden: Research Centers & Collaborations

Aula CIMNE Morelia Research collaboration and training space Collaboration highlights Numerical methods and computational engineering environment Research exchange and scientific-computing training

Coordinación de la Investigación Científica CIC-UMSNH · Institutional research coordination Collaboration highlights Institutional coordination for scientific research Support for model development, validation, and dissemination

💻 Technology Communities

📦 Framework	👥 Community	⭐ Contribution
	NumPy Community	Numerical array operations
	pandas Community	Tabular data processing
	scikit-learn Community	Model training and evaluation
	Matplotlib Community	Scientific visualization
	Joblib Community	Model serialization

---

📧 Contact & Support

Scientific questions, reproducibility reports, and model-integration support

Primary Research Contact Scientific coordination and model integration Gerardo Tinoco-Guerrero Universidad Michoacana de San Nicolás de Hidalgo Morelia, Michoacán, Mexico

Repository Support Questions, reproducibility reports, and integration problems Use GitHub Issues for reproducible software or documentation problems. Include the operating system, Python version, exact command, and traceback. Report any artifact-integrity failure produced by the release validator.

---

💬 FAQ

Why does the model predict v/R and D/R instead of v, D, and R?

The normalized concentration equation contains only the ratios $v/R$ and $D/R$. Concentration alone cannot distinguish parameter triples that preserve those ratios. An independent measurement of $R$ permits reconstruction of $v$ and $D$.

Does below sensor resolution mean lambda is zero?

No. It means the modeled decay effect is too small relative to the 3% sensor noise used by ADR1D. The numerical solver should use sensitivity analysis or additional observations rather than automatically setting $\lambda=0$.

Can I use a different sensor layout?

Not without retraining and new validation. Version 1.0.0 was evaluated only for six sensors at 100, 250, 400, 600, 800, and 1000 m with 49 temporal samples.

Can the model be used directly with Water Quality Portal records?

No. Those observations generally lack equivalent parameter labels, controlled source conditions, and complete boundary information. They were not used to train this release.

Why are exact package versions required?

The bundle uses Joblib serialization of scikit-learn estimators. Exact versions reduce persistence incompatibilities and support byte-for-byte reproduction of the locked release.

Is the conditional decay estimate reliable?

Its error remains high and only six resolvable cases occur in the test split. The output is useful for research and sensitivity studies, but it is not a precise field measurement.

---

Identifiable parameter inference for reproducible reactive-transport research

If this model supports your research, please cite it and consider giving the repository a star.