> ## Documentation Index
> Fetch the complete documentation index at: https://docs.atomscale.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Choose a Reference Growth

> Compare growths by RHEED behavior and similarity, then adapt the investigation to your process

Which growth should you use as a reference for the next campaign? Use an analysis notebook to compare candidates, make the selection criteria explicit, and check how closely other growths resemble them.

This example follows a thin-film growth campaign with RHEED recordings and process logs. We compare two ways of scoring the RHEED data, find that they favor different growths, and examine similarity to each candidate before choosing what to investigate next.

## Start with the decision and the available data

The campaign has five growths, labeled **Growth A** through **Growth E**, recorded under comparable conditions. Each has RHEED recordings and a process log identifying preparation, active deposition, pauses, and cooldown. The logs let us distinguish deposition from other parts of each run.

Open **Analysis**, select **New Notebook**, and enter a request such as:

> Help me choose a reference from the growths in this campaign. First list the available growths and identify a comparable cohort. Use the RHEED data and process logs to compare sustained oscillations with pattern stability. Score only active deposition, report missing measurements separately, and show how closely other growths resemble the leading candidates. Explain the assumptions before recommending a reference.

Click **Create Notebook**. Review the selected growths before continuing: material, process conditions, and measurement settings should support the comparison. See [Creating a Notebook](https://docs.atomscale.ai/platform/reference/notebooks.md) for the full creation instructions.

## Check which part of each run is being scored

For this example, the process log identifies **active deposition intervals**. The analysis uses those intervals to select the measurements that enter the scores.

| Recorded process phase | Included in the scores? |
| - | - |
| Preparation | No |
| Active deposition | Yes |
| Pause or cooldown | No |

Inspect the process timeline alongside the selected intervals. A whole-run average would also include preparation, pauses, or transitions, which answer a different question from how the film behaves during deposition.

If an oscillation cannot be fitted reliably, the example leaves its score missing. It does not assign a zero that would make an unmeasured growth look like a poor performer.

## Compare two definitions of a strong candidate

The example computes two scores:

* **Oscillation fit:** measures how well a fitted oscillation describes the specular-intensity signal after removing slow drift. The example uses the fit's R² over sustained oscillatory intervals; higher values indicate a closer fit.
* **Pattern stability:** summarizes variation in streak width and consistency between successive RHEED windows, scaled from 0 to 1 within the campaign. Higher values indicate steadier patterns under this chosen definition.

These are criteria chosen for this investigation. Both are shown on a 0–1 scale, but they measure different behavior. Neither is a direct measurement of the film's final performance.

![Two illustrative ranked bar charts. Growth A leads on oscillation fit at 0.94. Growth B leads on pattern stability at 0.91. Growth C ranks second under both criteria.](https://mintcdn.com/atomscale/ctQ7vpXU24t0DEF2/images/guides/reference-growth-rankings.svg?fit=max\&auto=format\&n=ctQ7vpXU24t0DEF2\&q=85\&s=4344ecfab7fe354df2ae7950273b4506)

![Two illustrative ranked bar charts. Growth A leads on oscillation fit at 0.94. Growth B leads on pattern stability at 0.91. Growth C ranks second under both criteria.](https://mintcdn.com/atomscale/ctQ7vpXU24t0DEF2/images/guides/reference-growth-rankings-dark.svg?fit=max\&auto=format\&n=ctQ7vpXU24t0DEF2\&q=85\&s=8ac69f4bfa713501bf6ac5400560bdc9)

Illustrative rankings using fictional values. Compare growths within each panel; the scores measure different behavior.

**The two methods select different growths.** Growth A leads on oscillation fit, while Growth B leads on pattern stability. Growth C ranks second under both criteria, making it another candidate worth inspecting.

Use that disagreement to refine the question:

> Show the deposition traces and representative RHEED frames for these candidates. Which features explain their rankings? Would the choice change if we prioritized sustained oscillations or pattern stability?

The notebook now helps you examine a specific tradeoff instead of treating “best growth” as a single, predefined property.

## Check how closely other growths resemble each candidate

Next, compare each winner with other growths using RHEED fingerprint similarity. Compare equivalent process phases and measurement views so that differences in timing or geometry do not dominate the result.

| Candidate | Selected by | Closest other growth's similarity |
| - | - | -: |
| Growth A | Oscillation fit | 0.71 |
| Growth B | Pattern stability | 0.83 |

In this fictional example, Growth B has the closer nearest match. That is a reason to inspect the matching growth, not enough evidence to conclude that this candidate is generally more reproducible.

Ask for the full comparison:

> Show the similarity distribution for each candidate, which measurement views contribute, and how much data is available for each match. Confirm that the comparison uses equivalent deposition intervals. How many comparable growths support each candidate?

Review the distribution before choosing a cutoff for close matches. A cutoff chosen to explore this campaign is not a validated process acceptance limit. Similar RHEED patterns also do not establish equal electrical, structural, or spatial properties.

## Record a decision the next engineer can assess

The example illustrates how different criteria can favor different candidates. Choosing a reference still requires a criterion tied to your engineering objective.

For a campaign prioritizing pattern stability, you could record this provisional decision:

> Take Growth B forward for characterization review because it leads on pattern stability. Keep Growth A as the oscillation benchmark. Compare the film properties we care about before adopting either as the campaign reference; the RHEED rankings and nearest matches alone do not settle that choice.

If sustained oscillations are your main objective, reverse that priority. Ask the assistant to preserve the reasoning with the result:

> Summarize the candidate I selected, the criterion behind that choice, the supporting measurements, and the unresolved checks. Keep the observed rankings separate from hypotheses about film performance.

[Save the notebook](https://docs.atomscale.ai/platform/reference/notebooks/saving-and-access.md) so colleagues can inspect the code and evidence. Record the selected reference and its limitations on the relevant sample or project. If you repeat this comparison, save a [reusable template](https://docs.atomscale.ai/platform/reference/notebooks/templates.md) after reviewing its assumptions.

## Adapt the comparison to your process

Keep the decision structure and change the inputs that give it meaning:

| In this example | In your investigation |
| - | - |
| Process logs define active deposition | Use recorded recipe steps, plasma state, precursor exposure, or another justified process boundary. |
| RHEED features define candidate scores | Choose measurements tied to your objective, such as thickness per cycle, a diffraction peak width, or a measured film property. |
| Matching measurement views support comparison | Account for measurement geometry, instrument settings, and which views are available. |
| RHEED similarity identifies related runs | Examine the relevant distribution across comparable growths, then check whether the outcome you care about agrees. |

For example, an ALD comparison could rank runs by thickness per cycle during equivalent recipe steps, then compare how consistently other runs achieve that value. The useful reference is the one supported by your process objective and measurements.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.