Lisa+model+chemal+and+gegg+sets+175+link Work -

: A US Department of Energy innovation hub focused on artificial photosynthesis and converting sunlight into solar fuels. Relevant Publications

| Feature | Description | |---------|-------------| | | Transformer‑based encoder‑decoder with cross‑modal attention layers. | | Parameters | Approximately 1.5 billion trainable weights (base model) with optional fine‑tuned variants up to 6 B. | | Training Data | 1.2 TB of paired text‑image data plus a curated corpus of scientific papers (chemistry, materials science). | | Modalities | Text, static images (up to 1024 × 1024 px), and limited video‑frame input (single‑frame inference). | | Safety | Built‑in toxic‑content filter and a “chemistry‑aware” guardrail that flags potentially hazardous synthesis instructions. | lisa+model+chemal+and+gegg+sets+175+link

| Feature | Description | |---------|-------------| | | Operate directly on molecular graphs, preserving permutation invariance. | | Algebraic Embedding | Encode orbital symmetries and conservation laws as constraints, reducing overfitting. | | Active Learning Loop | CHEM‑AL queries LISA for high‑uncertainty configurations, computes reference QM data, and retrains the model on‑the‑fly. | | Transferability | Trained models on GEGG Set 1 (organic molecules) can be adapted to GEGG Set 4 (metal–organic frameworks) with minimal data. | : A US Department of Energy innovation hub