Software / AI

Large Quantitative Models LQMs Pipelines - How to Train & Deploy LQM Models Effectively Pretrain chemistry foundation models on large open source materials datasets, then efficiently fine-tune with targeted, high-fidelity DFT data for battery-specific properties (voltage, stability, ionic transport, interfacial chemistry) to accelerate discovery-to-deployment.

ENDTOEND PIPELINE Pretraining Datasets – Large, open multi-domain materials datasets Materials Project, OC20/OC25, AQCat25

Self-Supervised Foundation Pretraining – GNNs/ Transformers architectures (e.g. MACE, GNoME learn universal embeddings

Next-gen Cathodes

Chemistry-specific Fine-Tuning – Targeted, high-fidelity DFT datasets and narrow chemistry classes (e.g. cathode, solidstate electrolytes)

New SSEs

Validation & Guardrails – DFT/MD benchmarks, uncertainty quantification, physics & chemistry sanity checks

Layered Structures

Source: GNoME 2023, OMat24 2024, Open Catalyst Project, JACS Au 2024, AQCat25 2025, Argonne/IBM 2025

Deployment & optimization – API serving, CI/CD, GPU scaling, Bayesian optimization integration & closedloop workflows

Key Metrics

Time-to-candidate ↓ 90% vs. brute-force DFT

RMSE 0.1 eV for formation energies

5× higher validated material hit-rate

20× compute cost reduction

Extracted from the PDF; charts and images may read imperfectly.