DreamLufei/invDesMobility-alignn-mobility-ranker

InvDesMobility ALIGNN Mobility Acquisition Ranker

This repository contains the ALIGNN mobility acquisition/ranking models used to prioritize generated 2D candidate structures before first-principles validation.

Paper and Repositories

• Paper: InvDesMobility: a reliability-gated first-principles feedback framework for closed-loop materials discovery
• arXiv: https://arxiv.org/abs/2606.16133
• Project page: https://dreamlufei.github.io/invDesMobility/
• GitHub: https://github.com/DreamLufei/invDesMobility
• Loop repository: https://github.com/DreamLufei/invdesmobility_loop
• Mobility workflow: https://github.com/DreamLufei/2d-mobility
• Zenodo: https://doi.org/10.5281/zenodo.20475023

Files

• baseline/: seed mobility ranker trained before closed-loop feedback.
• rounds/round_XX/: feedback-updated rankers for closed-loop rounds 01, 02, 03, 04, 06, 07, 08, and 09.

Each directory contains:

• best_model.pt: checkpoint used by the ranking/screening scripts.
• config.json: required ALIGNN architecture and graph configuration.
• Train_results.json, Val_results.json, Test_results.json: retained evaluation outputs.
• history_train.json, history_val.json: compact training-history records.
• ids_train_val_test.json: split identifiers.

Files such as current_model.pt, last_model.pt, W&B logs, cached graphs, and temporary data ranges are intentionally omitted from this minimal reproducibility release.

Intended Use

These models are acquisition/ranking models. They estimate a mobility-related score for generated candidates so that promising structures can be selected for DFT-scale validation. The model output is not a final trusted mobility label.

Training Data

The baseline model was trained on the seed mobility dataset. Closed-loop models were retrained with trusted feedback records extracted from completed 2d-mobility first-principles validation runs. The corresponding dataset manifests and feedback CSV files are packaged in DreamLufei/invDesMobility-data.

Training Parameters

The included config.json files contain the exact ALIGNNAtomWise model settings. For the round-09 model, key settings include:

• ALIGNNAtomWise with 4 ALIGNN layers and 4 GCN layers.
• Hidden dimension 256, embedding dimension 64.
• k-nearest graph construction, cutoff 8.0, max neighbors 12.
• AdamW optimizer, learning rate 5e-4, batch size 8.
• MSE loss, one-cycle scheduler, early stopping patience 30.

Evaluation

Round-09 retained evaluation artifacts:

• Train_results.json: 32 prediction records.
• Val_results.json: 4 prediction records.
• Test_results.json: 32 prediction records.

The JSON files are included for auditability rather than summarized away.

Limitations

The ranker is trained on a small, feedback-biased set of trusted DFT mobility records and is intended only for candidate prioritization. It may extrapolate poorly outside the distribution of the seed and feedback structures. Final claims require deterministic VASP-based mobility calculations from 2d-mobility.