bliskai-graded-disk-design

BliskAI model bundle

This private model bundle contains the frozen artifacts used by the reproducible ternary graded-disk workflow.

Components

• diffusion/ema_epoch_001.pth: EMA weights for the three-channel DDPM trained on constrained ternary alloy-fraction fields.
• material/models/best.pth: pixelwise ternary composition-to-property MLP.
• performance/models/best.pth: FiLM-conditioned property-map and operating- point surrogate for the pointwise shakedown SD boundary.
• model_manifest.json: repository locations and SHA-256 checksums.

The diffusion checkpoint is stored in the private Hugging Face model repository because it exceeds the preferred Git object size. The material and performance checkpoints are versioned with the code repository.

Intended use

The bundle is intended only for the geometry, property definitions, scalers, and operating-point convention documented in the associated private GitHub repository. It is not a general materials model and does not establish experimental qualification.

Reproducibility

Use bliskai models download and bliskai models verify. The latter checks every artifact against the frozen SHA-256 manifest before a generation run.

This work was funded by DARPA grant HR001123S0029-METALS-FP-015.