Microsoft updates Graphormer redox weights for flow battery design
Image generated by the Inferama Editorial Team
01

Artifact replacement and manifest publication

Microsoft updated the public Graphormer repository on Hugging Face, replacing the weights for redox potential prediction models. The change, identified by commit 8394616, replaces the oxidation and reduction files with a v2 version and adds a new manifest file documenting the technical specifications of the artifacts.

This update does not introduce a new language model or autonomous agent, but rather updates the molecular property predictors used in aqueous organic redox flow battery (AORFB) research. The replacement involves a change in content identifiers (SHA-256) and an increase in file size, suggesting retraining or an improvement in the base model's capacity.

The replacement of the previous weights with the v2 version signifies a distinct update in the model artifacts, as evidenced by the change in content identifiers. This modification ensures that the repository reflects the latest iteration of the molecular property predictors, allowing users to access the updated parameters directly from the public source without ambiguity regarding the version in use.

Artifact comparison before and after the update

Details of the changes in checkpoint files according to the commit record.

PropertyFilePrevious size (bytes)New size (bytes)New SHA-256 hash
Reductiongraphormer_reduction.pt1378101291931467178893ba5b4de6dd541f59a070e517119f9828073f227bbe5082e25ea6d8e3bc8a
Oxidationgraphormer_oxidation.pt1378101291931495332961e45c906a177c398aa61b9c3b2d28259e88e1f88aae787e77e8b72f21c8f5
02

Technical specifications of the v2 architecture

The new manifest.json defines the architecture of the checkpoints as graphormer_base. This configuration includes a 12-layer encoder, an embedding dimension of 768, and 32 attention heads. Serialization is performed as a PyTorch zip archive containing both the configuration and the model.

Each artifact contains 210 entries in the state dictionary. The distinction between the oxidation and reduction models is fundamental, as they predict distinct electrochemical properties and are not interchangeable. The manifest allows users to verify the integrity of downloaded files using the provided hashes.

The uniformity in the state dictionary entries across both artifacts confirms that the oxidation and reduction models share an identical structural backbone. This consistency simplifies the implementation of loading pipelines, as the same deserialization logic applies to both files, ensuring that the technical specifications remain aligned regardless of the specific electrochemical property being predicted.

Inference manifest structure · json

Fragment of manifest.json detailing the properties of the reduction and oxidation artifacts.

{
  "format": "orfb-design Graphormer inference checkpoints",
  "format_version": 1,
  "artifacts": [
    {
      "property": "reduction",
      "filename": "graphormer_reduction.pt",
      "architecture": "graphormer_base",
      "encoder_layers": 12,
      "encoder_embed_dim": 768,
      "encoder_attention_heads": 32
    },
    {
      "property": "oxidation",
      "filename": "graphormer_oxidation.pt",
      "architecture": "graphormer_base",
      "encoder_layers": 12,
      "encoder_embed_dim": 768,
      "encoder_attention_heads": 32
    }
  ]
}
03

Scientific context: the role of CLIO

These checkpoints are part of the ecosystem developed for the CLIO (Cognitive Loop via In-Situ Optimization) project, described in the paper "Closed-Loop Molecular Design with Calibrated Deference." CLIO is an agent that combines a belief-state graph with a recursive plan-then-act loop for molecular design.

In the negolyte design campaign for flow batteries, CLIO used these predictors to propose candidates. Although the system achieved a 130 mV improvement in redox potential, experimental characterization revealed poor electrochemical reversibility. CLIO identified the cause as potassium ion pairing with the phosphonate and prescribed a sulfonate replacement, closing the design-make-test-redesign loop.

The integration of these predictors within the CLIO framework demonstrates how computational estimates are combined with experimental feedback to refine molecular design. The agent's recursive planning loop allows it to interpret negative results, such as the observed reversibility issues, and adjust the design strategy accordingly, highlighting the importance of the closed-loop process in achieving successful outcomes.

04

Access, license, and infrastructure requirements

The weights are available under the MIT license in the `orfb-design` directory of the repository. There is no commercial API or per-query fees; usage requires own infrastructure to load the PyTorch files and run inference.

Teams wishing to use these models must verify the SHA-256 hashes against the manifest to ensure download integrity. The availability of advanced Graphormer versions is limited to Azure Quantum Elements, but these specific redox checkpoints are open in the public repository.

The requirement for local infrastructure means that research teams must manage their own computational resources to run inference on the PyTorch files. This setup places the responsibility for performance optimization and resource allocation directly on the users, who must ensure that their environment is correctly configured to handle the specific serialization format of the checkpoints.

Integrity verification process

Recommended steps to validate downloaded artifacts before use in research.

  1. 01Download the graphormer_reduction.pt and graphormer_oxidation.pt files from commit 8394616.
  2. 02Calculate the SHA-256 hash of each file locally.
  3. 03Compare the calculated hashes with the values declared in manifest.json.
  4. 04Verify that the file sizes match the specified values (193146717 and 193149533 bytes).
05

Implications for flow battery research

The publication of these v2 weights allows other researchers to reproduce and extend the negolyte design work. Transparency in architecture and hashes facilitates reproducibility and auditing of results obtained with these predictors.

It is important to distinguish between the property predictor and the full CLIO agent. The checkpoints do not encapsulate the agent's "calibrated deference" capability, but are prediction tools that the agent uses. Results of the full system should not be attributed exclusively to the accuracy of these isolated models.

The open access to the v2 weights enables a broader community to validate the findings of the original negolyte design campaign, supporting the scientific principle of reproducibility. Independent groups can test the predictors against their own datasets, while the clear separation between the prediction tools and the full agent framework ensures accurate attribution of performance outcomes to the specific components utilized.

06

Keep exploring

Sources consulted

03

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.

Submit a correction