# NREL single-crystal silicon cell efficiency frontier

- **Series role:** primary physical-outcome frontier
- **Metric:** independently reported one-sun research-cell conversion efficiency
- **Technology:** `Crystalline Si Cells` / `Single crystal (non-concentrator)`
- **Coverage:** see `summary.json` and the central observation store
- **Upstream:** NREL research-cell efficiency workbook
- **Current status:** monitoring; descriptive physical frontier; no AI-causal estimate

## What this series measures

The series follows one stable category in NREL's Best Research-Cell Efficiency data rather than combining unlike photovoltaic technologies. It retains records whose workbook fields identify:

```text
Material class: Crystalline Si Cells
Cell type: Single crystal (non-concentrator)
Illumination: one sun or unstated
```

Within each measurement date, the highest eligible efficiency is retained. It becomes a frontier observation only when it exceeds the previous running maximum.

This produces a physical performance frontier. It is not a paper count, a count of photovoltaic experiments, or a measure of deployed solar power.

## Frozen inclusion rule

The primary series uses every record in the official NREL category above. It does **not** add a new testing-center or 1 cm² exclusion.

The decision audit is regenerated from the workbook. Its current eligible-row count, frontier-step count, area distribution, testing-center completeness, and sensitivity results are recorded in `summary.json`. The official category is the primary construct. Area, area type, accredited testing center, and uncertainty remain visible quality dimensions. The 1 cm² rule is retained as a sensitivity analysis.

## Current facts

The current frontier endpoints, total improvement, latest record metadata, date-audit counts, and sensitivity results are generated in `summary.json`. The checker ensures that the eligible category contains no date mismatch even when unrelated workbook rows carry audit flags.

## Date and efficiency rules

- Use the workbook's explicit `Measurement Date` serial when available.
- Retain disagreements with the separate `Month` and `Year` fields as audit flags.
- Prefer `Combined efficiency (%)`.
- If it is absent, fall back to `Revised/New Efficiency (%)`, then `Efficiency (%)`.

The current efficiency-basis distribution is recorded in `summary.json` and is checked against the raw workbook.

## Rebuild method

`fetch.py` supports three modes:

```sh
# Fetch NREL's current workbook, vendor it, and rebuild the panel
python3 series/nrel-research-cell-efficiency-frontier/fetch.py

# Rebuild every derived table without network access
python3 series/nrel-research-cell-efficiency-frontier/fetch.py --from-snapshot

# Fail only when the upstream workbook differs from the committed snapshot
python3 series/nrel-research-cell-efficiency-frontier/fetch.py --check-upstream
```

The rebuild creates:

- the committed raw workbook under `raw/`;
- `records.json`, containing all eligible category records;
- `summary.json`, containing the frozen definition, quality audit, sensitivities, and frontier steps;
- irregular frontier observations in `data/series-observations.json`;
- synchronized registry metadata in `data/series.json`.

`check.mjs` independently reconstructs the category and frontier from the raw workbook and compares it with every committed output and registry field.

## Interpretation

The series can answer:

> How did the independently reported single-crystal silicon research-cell efficiency frontier move over time?

It cannot answer:

> How much of that movement was caused by AI?

No AI exposure or treatment is assigned in the first release. The panel remains descriptive and its causal status remains unresolved. A later AI study would need research-process exposure, treated and comparison technologies or laboratories, and a credible counterfactual rather than a global “AI era” date.

## Limitations

- A record frontier ignores the full distribution of photovoltaic research outcomes.
- Efficiency changes reflect materials, device architecture, fabrication, metrology, investment, and accumulated scientific knowledge.
- NREL technology labels and chart-entry conventions can change across workbook versions.
- Older records often omit a populated testing-center field.
- Research-cell efficiency does not directly measure manufacturing cost, durability, module yield, or deployed climate impact.

## Methodological inspiration

The fixed category, committed primary-source snapshot, explicit quality sensitivities, and machine-checked public claims were inspired in part by METR's [“LLMs' Contribution to Discoveries”](https://metr.org/notes/2026-08-14-llm-contribution-to-discoveries/) and Thomas Cunningham's public [`ai-discovery-data`](https://github.com/tecunningham/ai-discovery-data) repository.

This tracker independently rebuilds the frontier from NREL's primary workbook and keeps the observed physical outcome separate from any AI attribution.