ADR1D-EWS 🚨

Causal machine-learning early warning for downstream contaminant arrival in the ADR1D benchmark.

A frozen one-hour classifier, complete decision evidence, and reproducible interfaces for training, inference, validation, and visualization.

🔗 Quick Links

📋 Table of Contents

Overview
Repository Structure
Installation
Quick Start
Early-Warning Task
Data and Feature Contract
Model Development
Validated Results
Visual Results
Inference Interface
Validation and Reproducibility
Data Provenance
Limitations
How to Cite
Research Team
Industry Partners
License and Rights
Acknowledgments
Contact and Support
FAQ

🌟 Overview

ADR1D-EWS predicts whether contaminant concentration at a protected downstream location will reach a defined threshold within the next hour. It was developed from the open ADR1D benchmark, which provides 300 reactive-transport scenarios observed every 30 minutes at six virtual sensors.

The warning system protects sensor S06 at 1000 m. It receives only measurements from upstream sensors S01-S05 available through the current decision time. The model never uses future observations, S06 measurements, source parameters, transport parameters, scenario identifiers, labels, or partition assignments as predictors.

The release separates three questions that are often reported as though they were interchangeable:

Detection: has an upstream sensor already measured a concentration above its detection limit?
Prediction: will the protected location reach 0.01 mg/L within one hour?
Early warning: does the prediction become positive before arrival at S06, and how much valid warning time does it provide?

🔧 Key Capabilities

⌛ Physical warning horizon: a primary one-hour task with 30-minute and two-hour sensitivity labels retained in the decision table.
🔒 Causal predictors: 52 features computed from upstream histories through each decision time only.
🔀 Scenario-independent partitions: complete scenarios remain in train, validation, or test; temporal windows never cross those boundaries.
⚖️ Comparable references: two interpretable alert rules and three supervised classifiers use the same evaluation units.
❄️ Frozen deployment artifact: the selected estimator and threshold were serialized before the locked test evaluation.
🔬 Independent inference check: the saved artifact reproduces all 200 test alerts and rejects missing or non-finite features.
📊 Operational metrics: pointwise discrimination is reported together with event detection, warning time, and false-alert episodes.

📊 Release at a Glance

Item	Value
ADR1D scenarios	300
Train / validation / test scenarios	210 / 45 / 45
Decision rows	1,346
Causal input features	52
Sampling interval	30 min
Primary warning horizon	1 h
Protected sensor and threshold	S06 at 1000 m; 0.01 mg/L
Selected estimator	Histogram gradient boosting
Frozen probability threshold	0.368151
Test events detected	45 of 45
Median valid test warning time	1.0 h
Scheduled development period	August-September 2025
Initial code release	August 2025
Validation and documentation update	July 2026
Package version	1.0.0

The scheduled period records the project plan. The current execution, validation, and public-release preparation were completed in July 2026 and retain their actual dates in the machine-readable evidence.

📂 Repository Structure

.
|-- README.md
|-- CITATION.cff
|-- LICENSE
|-- LICENSE-DATA
|-- requirements.txt
|-- data/
|   |-- synthetic_adr1d_sensor_observations.csv
|   `-- adr1d_early_warning_decision_points.csv
|-- models/
|   `-- adr1d_early_warning_model.joblib
|-- scripts/
|   |-- build_decision_table.py
|   |-- train_early_warning_models.py
|   |-- predict_early_warning.py
|   |-- validate_inference_system.py
|   `-- plot_early_warning_results.py
|-- results/
|   |-- decision_table_validation.json
|   |-- model_selection_validation.json
|   |-- test_evaluation.json
|   |-- test_predictions.csv
|   |-- inference_system_validation.json
|   |-- figure_manifest.json
|   `-- figures/
|       |-- model_comparison.png
|       |-- selected_model_diagnostics.png
|       `-- alert_timelines.png
`-- docs/
    |-- team/
    `-- partners/

The repository uses one authoritative README. Schemas, protocols, results, and limitations are documented here, while exact validation values remain in JSON and CSV files suitable for automated inspection.

📦 Installation

System Requirements

Component	Supported configuration
Python	3.12
Operating system	Linux, macOS, or Windows
RAM	2 GB for inference; 4 GB recommended for reconstruction and plotting
Storage	Approximately 400 MB including the virtual environment
Accelerator	Not required; all documented workflows run on CPU

The canonical environment used Python 3.12.2 on Apple arm64. The model relies on scikit-learn serialization and should be loaded with the pinned package versions whenever exact reproduction is required.

Pinned Dependencies

Package	Version	Role
NumPy	2.0.2	Numeric arrays and validation
pandas	2.2.3	Sensor, feature, and prediction tables
SciPy	1.15.2	Numerical dependency used by scikit-learn
scikit-learn	1.7.0	Classifiers, metrics, and model inference
Joblib	1.4.2	Serialized model bundle
Matplotlib	3.10.0	Diagnostic figures

Clone and Install

git clone https://github.com/gstinoco/ADR1D-EWS.git
cd ADR1D-EWS
python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install --upgrade pip
python3 -m pip install -r requirements.txt

On Windows PowerShell, activate the environment with:

.venv\Scripts\Activate.ps1

Installation Check

python3 scripts/validate_inference_system.py

Expected final message:

Validation status: ok (200 rows)

🚀 Quick Start

Generate Alerts with the Frozen Model

python3 scripts/predict_early_warning.py \
    --input-csv data/adr1d_early_warning_decision_points.csv \
    --split test \
    --output-csv test_alerts.csv

The output contains the preserved identifiers, estimator name, model version, probability of arrival within one hour, frozen decision threshold, and binary alert.

Rebuild the Causal Decision Table

python3 scripts/build_decision_table.py

This command reads the bundled six-sensor histories, rebuilds all pre-arrival decision rows, checks labels and partitions, and perturbs future observations to confirm that prior features remain unchanged.

Reproduce Model Development

python3 scripts/train_early_warning_models.py
python3 scripts/validate_inference_system.py
python3 scripts/plot_early_warning_results.py

The training command recreates the published comparison under the locked protocol. It must not be used to choose a different model or threshold from the already reported test results.

📐 Early-Warning Task

Protected Event

For scenario (s), let (C_{s,6}(t)) denote the true concentration at S06. The event time is the first sampled time at which:

$$t_s^{\mathrm{event}} = \min\left\{t:C_{s,6}(t)\geq 0.01\ \mathrm{mg/L}\right\}.$$

At an eligible decision time (t<t_s^{\mathrm{event}}), the primary label is:

$$y_s(t) = \begin{cases} 1, & 0<t_s^{\mathrm{event}}-t\leq 1\ \mathrm{h},\\\ 0, & t_s^{\mathrm{event}}-t>1\ \mathrm{h}. \end{cases}$$

This definition asks whether arrival is imminent. It does not classify an unknown chemical species and does not treat hydrologic context as a contamination label.

Information Available at Decision Time

Only noisy, detection-limit-aware observations from S01-S05 at times (\tau\leq t) enter the feature vector:

$$\mathbf{x}_s(t) = \Phi\left( \left\{\widetilde{C}_{s,j}(\tau): j\in\{1,\ldots,5\},\ \tau\leq t\right\} \right).$$

S06 is used only to determine event timing and labels during benchmark construction. The model input contains no target-sensor concentration, time-to-event value, future sample, physical parameter, or split identifier.

Independent Evaluation Unit

The scenario is the independent unit. All decision times from one scenario remain in a single partition:

Partition	Scenarios	Decision rows	Role
Train	210	903	Fit supervised estimators
Validation	45	243	Select estimator and threshold
Test	45	200	One locked final evaluation

Each row receives the reciprocal of its scenario's decision count so that every scenario contributes total weight one, regardless of its travel time.

📥 Data & Feature Contract

Source Sensor Table

data/synthetic_adr1d_sensor_observations.csv contains 88,200 observations: 300 scenarios, six sensors, and 49 half-hour samples per sensor.

Column	Meaning	Unit
`observation_id`	Stable observation identifier	none
`scenario_id`	Independent ADR1D scenario	none
`split`	Frozen train, validation, or test assignment	none
`sensor_id`	Sensor S01-S06	none
`x_m`	Sensor position	m
`time_s`	Elapsed time	s
`concentration_true_mg_L`	Analytical concentration	mg/L
`noise_std_mg_L`	Observation-noise standard deviation	mg/L
`concentration_observed_mg_L`	Non-negative noisy observation	mg/L
`detection_limit_mg_L`	Reporting limit	mg/L
`is_below_detection_limit`	Censoring indicator	Boolean

Decision Table

data/adr1d_early_warning_decision_points.csv contains one row for every eligible scenario-time decision before physical arrival at S06.

Column family	Purpose	Used by model?
Identifiers and split	Traceability and independent partitioning	No
Decision and event times	Evaluation and warning-time calculation	No
Horizon labels	Supervised targets at 0.5, 1, and 2 h	No
Scenario weight	Equal total contribution per scenario	Fit only
`evaluation_*`	Observable lead-time audit	No
`feature_*`	Causal upstream predictors	Yes

The 52 Predictors

Group	Construction	Count
Per-sensor history	Current value, 30-min lag and change, 1-h mean and maximum, current detection, 1-h detection fraction, and prior detection	40
Network state	Mean and maximum concentration, current and historical detection counts, farthest detected sensor, concentration centroid, and largest positive change	8
Spatial gradients	Current concentration gradient between adjacent upstream sensors	4

All concentration features use mg/L; positions use m; gradients use mg/L/m. Missing feature columns, non-finite values, duplicated sensor-time rows, unexpected geometry, invalid split assignments, and asynchronous sample grids are rejected with explicit errors.

⚙️ Model Development

Fixed Operational Baselines

Persistent upstream detection: alert from the first upstream detection onward.
Current S05 detection: alert only while the nearest upstream sensor is currently above its detection limit.

The first rule favors sensitivity and tends to remain active too early. The second favors specificity but often detects arrival too late.

Supervised Candidates

logistic regression with standardization and deterministic liblinear optimization;
random forest with 400 trees and a single-thread deterministic execution policy;
histogram gradient boosting with 250 boosting iterations, learning rate 0.05, at most 15 leaf nodes, minimum leaf size 10, and L2 regularization 1.0.

All candidates use random seed 20250801, the same 52 predictors, the same scenario weights, and the same train and validation scenarios. The comparison is deliberately limited to three model families rather than a broad hyperparameter search.

Selection and Threshold Protocol

For each supervised candidate, the validation threshold maximizes equal-scenario balanced accuracy. Ties are resolved by F2, event detection, false-alert episode rate, and the higher threshold. Model families are ranked by validation balanced accuracy, F2, average precision, event detection, and false-alert episode rate.

Histogram gradient boosting achieved the largest validation objective and was frozen at threshold 0.3681513624958759. The fitted estimator was trained only on the training partition. The threshold was selected only on validation. The test partition was opened after both choices were serialized.

📊 Validated Results

Validation Comparison

Metrics below assign equal total weight to every validation scenario.

Method	Balanced accuracy	Sensitivity	Specificity	Event detection	False-alert episodes/h
Persistent upstream	0.750	0.874	0.627	1.000	0.235
Current S05	0.637	0.299	0.974	0.644	0.078
Logistic regression	0.837	0.792	0.882	1.000	0.183
Random forest	0.839	0.807	0.870	0.978	0.196
Histogram gradient boosting	0.849	0.796	0.901	0.978	0.170

Locked Test Evaluation

Metric	Result
Equal-scenario balanced accuracy	0.818
Equal-scenario sensitivity	0.738
Equal-scenario specificity	0.898
Equal-scenario precision	0.897
Equal-scenario F1 / F2	0.810 / 0.765
ROC-AUC / average precision	0.848 / 0.891
Unweighted decision accuracy	0.840
Events detected	45 of 45
Missed events	0
Median valid warning time	1.0 h
Valid warning-time range	0.5-1.0 h
False-alert points	14
False-alert episodes	13
Negative exposure	55 scenario-hours
False-alert episodes per scenario-hour	0.236
Scenarios with at least one false alert	12 of 45

Event detection and pointwise sensitivity answer different questions. An event may have up to two eligible decisions within the one-hour window; detecting at least one produces a detected event even if another eligible decision is negative. For this reason, 45 of 45 detected events is compatible with pointwise sensitivity of 0.738.

The selected model improves specificity and precision relative to the persistent upstream rule while retaining event-level detection on this test partition. False alerts remain visible: 12 test scenarios contain at least one false-alert episode.

Exact validation candidates, thresholds, test baselines, confusion components, probability metrics, and warning-time summaries are stored in results/model_selection_validation.json and results/test_evaluation.json.

🎨 Visual Results

Method Comparison

Validation and locked-test comparison of ADR1D-EWS baselines and classifiers

The panels separate balanced accuracy, sensitivity, specificity, event detection, and false-alert episode rates. The test comparison is an audit after selection and was not used to rerank models.

Selected-Model Diagnostics

ROC, precision-recall, confusion, and probability diagnostics for the selected ADR1D-EWS model

Probability distributions overlap near the frozen threshold, which explains the remaining false positives and false negatives at individual decision times. The confusion matrix therefore accompanies, rather than replaces, the event-level analysis.

Representative Alert Timelines

Representative ADR1D-EWS sensor histories, warning windows, model probabilities, and alerts

The timelines show upstream measurements, physical arrival at S06, the valid one-hour warning window, probability threshold crossings, correct alerts, and false alerts. Scenario identifiers used in the figure are recorded in results/figure_manifest.json.

💻 Inference Interface

Command-Line Use

python3 scripts/predict_early_warning.py \
    --model models/adr1d_early_warning_model.joblib \
    --input-csv my_feature_rows.csv \
    --output-csv my_alerts.csv

The input may include metadata, but it must contain all 52 frozen feature_* columns. Their order in the CSV is not significant because the model bundle stores and applies its own feature order.

The output contract is:

Column	Meaning
Preserved metadata	Any available decision, scenario, split, and timing fields
`estimator_name`	`hist_gradient_boosting`
`model_version`	`1.0.0`
`alert_probability`	Predicted probability of threshold arrival within one hour
`alert_threshold`	Frozen threshold `0.3681513624958759`
`alert`	Binary thresholded alert

Python Use

from pathlib import Path

import pandas as pd

from scripts.predict_early_warning import load_model_payload, predict_alerts


payload = load_model_payload(Path("models/adr1d_early_warning_model.joblib"))
feature_rows = pd.read_csv("my_feature_rows.csv")
alerts = predict_alerts(feature_rows, payload)

Joblib artifacts can execute code during loading. Use the bundled model or another artifact obtained from a trusted source.

✅ Validation & Reproducibility

Level 1: Frozen Inference Validation

python3 scripts/validate_inference_system.py

The validator:

reloads the serialized estimator;
confirms the model version, feature count, and frozen threshold;
reproduces all 200 locked test decisions;
requires exact agreement for all 200 binary alerts;
checks probability agreement within 1e-15;
checks one-row inference;
verifies rejection of a missing feature;
verifies rejection of a non-finite feature;
confirms that the model and threshold were not changed after testing.

The current status is ok. Reloaded probabilities differ from the reference by at most 1.11e-16.

Level 2: Rebuild and Validate Features

python3 scripts/build_decision_table.py

The builder verifies 300 scenario-level split assignments, 1,346 pre-arrival decisions, 52 finite predictor columns, three horizon labels, and equal-scenario weights. It also modifies future upstream measurements in one scenario from each partition and requires all prior features to remain exactly unchanged.

Level 3: Recreate the Published Experiment

python3 scripts/train_early_warning_models.py
python3 scripts/validate_inference_system.py
python3 scripts/plot_early_warning_results.py

The model-development script uses one thread for the random forest comparator and 15-significant-digit CSV serialization. Consecutive runs in the canonical environment reproduce the model choice, threshold, metrics, alerts, and published prediction table.

🔍 Data Provenance

Resource	Role	Location
ADR1D 1.0.0	Canonical scenarios and six-sensor histories	GitHub, Zenodo
ADR1D sensor observations	Bundled source for decision-table reconstruction	`data/synthetic_adr1d_sensor_observations.csv`
ADR1D-EWS decision table	Derived causal predictors, labels, and evaluation metadata	`data/adr1d_early_warning_decision_points.csv`
ADR1D-EWS model	Frozen histogram-gradient-boosting estimator and threshold	`models/adr1d_early_warning_model.joblib`

Water Quality Portal data were reviewed as observational context during task definition but were not used to train, select, or test ADR1D-EWS. Their sampling intervals, incomplete sample-time coverage, and absence of explicit contamination-onset labels do not support the present high-frequency supervised warning experiment. This repository therefore does not redistribute WQP records.

The release contains no publisher material, private measurements, external credentials, or institutional-access content.

⚠️ Limitations & Responsible Use

The evidence comes from one-dimensional synthetic ADR1D scenarios rather than field monitoring campaigns.
Every current scenario contains a contaminant pulse. The test partition has pre-event negative decisions but no independent source-free campaign.
Event detection of 45 of 45 does not establish field sensitivity, transfer across contaminants, or performance under unknown sensor failures.
The target is arrival at S06 above 0.01 mg/L; it is not chemical-species identification, source localization, or health-risk classification.
The 30-minute sampling interval limits warning-time resolution to 0.5 h.
The model assumes the documented geometry, units, censoring rule, and five upstream sensors. A changed network requires new feature construction and validation.
Probabilities are discriminative scores, not calibrated operational risk estimates. No probability-calibration claim is made.
Changing the threshold changes sensitivity, specificity, and false-alert burden. A deployment threshold must be chosen with independent validation and explicit response costs.
The benchmark does not represent sensor drift, missing streams, maintenance outages, adversarial corruption, or changing field chemistry.
ADR1D-EWS should support scientific evaluation and prototype integration; it should not be the sole basis for environmental, regulatory, or public-health action.

📝 How to Cite

Preferred Software Citation

Gerardo Tinoco-Guerrero, Francisco J. Domínguez-Mota, and J. Alberto Guzmán-Torres (2026). ADR1D-EWS: Causal Machine-Learning Early Warning for Reactive Contaminant Transport. Version 1.0.0. Universidad Michoacana de San Nicolás de Hidalgo. https://github.com/gstinoco/ADR1D-EWS

Machine-readable metadata are available in CITATION.cff. GitHub can display them through its Cite this repository interface. An archival DOI can be added after the first public release is deposited.

Analyses should also cite the upstream ADR1D benchmark:

Gerardo Tinoco-Guerrero, Francisco J. Domínguez-Mota, and J. Alberto Guzmán-Torres (2026). ADR1D and WQP-NM-Nutrients: A Reproducible Contaminant-Transport Benchmark and Curated Water-Quality Snapshot. https://doi.org/10.5281/zenodo.21499528

Scientific References

Friedman, J. H. (2001). Greedy function approximation: A gradient boosting machine. The Annals of Statistics, 29(5), 1189-1232. https://doi.org/10.1214/aos/1013203451
Breiman, L. (2001). Random forests. Machine Learning, 45, 5-32. https://doi.org/10.1023/A:1010933404324
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

🧑‍🔬 Research Team

🌟 Meet the Team

Researchers and students advancing reproducible scientific machine learning and numerical modeling for environmental transport

👥 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 acknowledged 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 environmental modeling, scientific machine learning, and applied engineering

SIIIA MATH
_{Soluciones en Ingeniería}

Focus areas

Mathematical modeling and scientific computing
Early-warning systems for environmental transport applications
Applied research, development, and technology transfer

📄 License & Rights

ADR1D-EWS uses component-specific licensing:

Source code and serialized model: MIT License in LICENSE.
Data tables, reported results, and figures: Creative Commons Attribution 4.0 International in LICENSE-DATA.
Upstream ADR1D content: remains subject to the attribution and licensing terms of the archived ADR1D release.
Third-party software: remains subject to its own licenses.
Institutional logos and personal photographs: excluded from the MIT and CC BY grants and retained by their respective rights holders.

Copyright attribution identifies Gerardo Tinoco-Guerrero, Francisco J. Domínguez-Mota, and J. Alberto Guzmán-Torres as the principal investigators. The software and model are provided without warranty. Attribution must not imply endorsement by the authors or supporting institutions.

🙏 Acknowledgments

❤️ Special Thanks

We thank the institutions and partners whose continuing institutional and financial support made early-warning model development, validation, documentation, dissemination, 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 early-warning development, validation, and dissemination

💻 Technology Communities

📦 Framework	👥 Community	⭐ Contribution
	Python Community	Reproducible scientific workflows
	NumPy Community	Numerical arrays and validation
	pandas Community	Sensor and decision tables
	scikit-learn Community	Classification, metrics, and inference
	Matplotlib Community	Scientific visualization
	Joblib Community	Trusted local model serialization

📧 Contact & Support

Scientific questions, reproducibility reports, and early-warning integration support

Primary Research Contact
_{Scientific coordination and early-warning 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 whether python3 scripts/validate_inference_system.py succeeds.

💬 FAQ

What does ADR1D-EWS predict?

It predicts whether true concentration at ADR1D sensor S06 will first reach 0.01 mg/L within the next hour, using only upstream S01-S05 histories available through the decision time.

Why are all 45 test events detected when pointwise sensitivity is 0.738?

One event can have two eligible half-hour decisions inside the one-hour window. Event detection requires at least one positive alert; pointwise sensitivity evaluates every eligible decision separately.

Is ADR1D-EWS ready for autonomous field deployment?

No. The current evidence is synthetic, one-dimensional, and does not include independent source-free campaigns, sensor outages, drift, or changing field chemistry. Field use requires new data, calibration, operational thresholds, and independent validation.

Were Water Quality Portal records used to train the classifier?

No. They informed the initial data-suitability review but were excluded from training, selection, and testing because the available records do not provide regular high-frequency histories or explicit contamination-onset labels for this task.

Can the probability threshold be changed?

Technically yes, but the published test metrics apply only to 0.3681513624958759. A different threshold must be selected using independent validation evidence and the intended cost of missed events and false alerts.

Can any Joblib model be loaded safely?

No. Joblib and pickle-based artifacts may execute code while loading. Use the bundled file or another model obtained from a trusted source.

Reproducible early warning for environmental transport models

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

Name		Name	Last commit message	Last commit date
data		data
docs		docs
models		models
results		results
scripts		scripts
.gitignore		.gitignore
CITATION.cff		CITATION.cff
LICENSE		LICENSE
LICENSE-DATA		LICENSE-DATA
README.md		README.md
requirements.txt		requirements.txt

Latest commit

History

Folders and files

Repository files navigation