Simulation-Powered Computational Linguistics for Materials Foundation Models

SIMU-LINGUA will deliver an open, multimodal Scientific Foundation Model (SciFM) for materials that unifies text, structures, images and simulation data to accelerate discovery across design–simulate–validate loops. The SciFM is paired with a physics‑in‑the‑loop generative engine and an AI‑accelerated multi‑scale simulation pipeline (DFT → phase‑field → CFD/FEM). Together they form a closed loop: the model proposes candidates, fast surrogates verify plausibility and performance, and new data flows back through a FAIR knowledge graph. The KG provides ontologies, provenance and leakage‑safe splits, enabling transparent training, robust evaluation and reproducible releases. A natural‑language interface lets domain experts query the KG, launch virtual experiments and steer multi‑objective inverse design. We will progress from TRL‑1 to TRL‑4 in 48 months and validate generalisation on six use cases: lead‑free electroceramics; durable glasses; high‑entropy and recyclable alloys; analyte‑sensitive coatings; and a 64×128 device‑level olfactory array. Quantified targets include >50% reduction in time‑to‑discovery, 10^3× simulation speed‑ups with 80% physically valid, synthesizable generative proposals. Trust is built via UQ (ensembles/conformal), calibration (ECE ≤0.05), telemetry‑gated training and ELSI guardrails aligned with EU policy. All non‑restricted models, datasets and code will be released with model/data cards, DOIs and APIs; a practical hosting plan ensures community uptake beyond the project. SIMU‑LINGUA advances GenAI4EU with a sovereign, reusable blueprint for Scientific FMs, bridging data silos and physics‑grounded AI to deliver measurably faster, safer and more transparent materials innovation for Europe. Data and models follow FAIR/DOME practices with audited Data Gates and model cards for reproducibility. EuroHPC‑plus‑fallbacks de‑risk pre‑training; integration sprints, CI/CD and ICDs harmonise WPs.