Skip to content
This repository was archived by the owner on Sep 18, 2025. It is now read-only.
This repository was archived by the owner on Sep 18, 2025. It is now read-only.

Atom type prediction returns all zeros in conditional CSP with both null and simplex manifolds #19

Description

@hallofuture

Thank you for developing and maintaining FlowMM and I'm encountering a critical bug where the models predict atomic number 0 for all atoms during conditional crystal structure prediction, regardless of whether I use null_params or simplex_params models.

Steps to Reproduce:

Train conditional models:
python scripts_model/run.py data=perov model=null_params or
python scripts_model/run.py data=perov model=simplex_params

Create input formulas file (my_formulas.txt):
[[1, 1, 8], [6, 8, 8], [26, 8, 8, 8]]

Run prediction:
python scripts_model/evaluate.py predict [checkpoint_path] my_formulas.txt --subdir csp_results --pred_id 1

Consolidate and examine results:
python scripts_model/evaluate.py consolidate [checkpoint_path] --subdir csp_result

The model should predict proper atomic types matching the input formulas:

[1, 1, 8] → H, H, O
[6, 8, 8] → C, O, O
[26, 8, 8, 8] → Fe, O, O, O

But actually all predicted atomic types are 0, showing as 'X0' when converted to element symbols. The model correctly receives the input formulas (verified in input_data_batch) but outputs zeros.

Image Image

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions