| English | 中文 |
Scientific question: how well do environmental and human-activity proxy features predict station-level long-term mean ozone, and how stable are the resulting predictive associations across feature groups, model choices and validation designs?
The stable project ID is ozone_environmental_drivers. The directory number 0001 is for navigation; new results use the stable ID. Here, “drivers” names candidate explanatory features, not identified causal effects.
| Item | Definition |
|---|---|
| Public input | AQ-Bench v1 station metrics and variable dictionary; fixed version referenced in project.json |
| Response | o3_average_values, in ppb; station-level aggregated ozone for 2010–2014 |
| Research stage | 03_atmospheric_transport |
| General workflow | ml_regression, version 1.0.0 |
| Registered tool | ml_regression_workflow |
| Operational skill | Regression skill |
| Scientific/source reference | Ozone reference case and public input |
The study plan defines hypotheses, experiments and interpretation. review.md is a pending review template. The existing small configuration demonstrates the computation path; it is not a pre-approved full study design.
| Experiment | Scientific purpose | Execution status |
|---|---|---|
| E01 — Baseline comparison | Compare allowed predictive models against the fixed mean baseline on a common cohort | First paper prediction run and independent metric checks complete; reference comparison complete; human scientific review pending |
| E02 — Feature-group comparison | Assess how environmental and human-activity proxy groups change predictive performance and interpretation | First nine- versus three-predictor comparison complete; reference comparison complete; human scientific review pending |
| E03 — Robustness and interpretation | Assess dependence on development seeds, validation design and correlated features | Planned; alternative spatial splits and additional explanation methods need adaptation |
| E04 — Original implementation comparison | Audit provenance, common cohorts, predictions and metrics | compare_existing complete; historical training provenance incomplete and human review pending |
The first paper prediction and evaluation experiment is complete, using external assets and a frozen prompt and protocol. It retains 4077 development and 1009 test stations. Both models underwent ten-fold selection, refitting on all development stations and final test assessment. Computation and independent checks are complete; original models/results have been compared read-only through the model_reproduction workflow. Historical training provenance retains gaps; robustness experiments and human scientific review remain pending.
Study-specific project scripts implement ten-fold orchestration using core launch mechanisms, the regression metric helper and mlr3 guidance. This mode is not implemented by generic ml_regression_workflow. See the study guide for its entry point and actual status; earlier generic smoke evidence retains its original scope.
For generic adapter tasks, read the study plan, prepare experiment-specific task settings in the private configuration root, then call the existing tool using --project-id ozone_environmental_drivers --stage 03_atmospheric_transport. Follow the separate launcher instructions for the paper experiment. Use a fresh run for each experiment configuration. The project manifest references the workflow; it does not execute it automatically. See project operating rules and workflow execution.
Study outputs stay within external assets/results/ozone_environmental_drivers/; choose experiment directories and relative artifact names while retaining the experiment/run mapping. General DAG and estimand design lives in methods and workflows. This project remains predictive; adding a causal skill branch does not automatically run observational causal experiments.
references/ holds this question’s legacy source audit, public-input description and reproduction adaptations. These are project references, not completed experiment results or universal defaults; reusable scientific design remains in methods.
Inputs have matching repository and external copies. project.json records both paths for explicit selection; project loading remains manual, and result locations are unchanged.