SciLM

Multi-domain scientific foundation model, swarm-of-primitives architecture

Overview

A physics-constrained foundation model spanning gravitational-wave, condensed matter, and materials domains. Replaces mixture-of-experts routing with six composable mathematical primitives — conserved-current, response/Green's function, extremal/variational, spectral/eigenstructure, group-equivariant, and renormalization — that assemble into domain experts. Phase-1 falsification target: GW quasinormal-mode pole recovery and cascade damage scaling. Built in collaboration with Michael Lawler (Cornell/Binghamton).

Architecture

• Swarm-of-primitives architecture replaces mixture-of-experts routing entirely
• Six composable mathematical primitives: conserved-current, response/Green's function, extremal/variational, spectral/eigenstructure, group-conditioned equivariant, and scaling/renormalization
• Primitives compose into domain-specific expert models rather than being selected between at inference time
• Domains targeted: gravitational-wave, condensed matter, materials science
• Full Claude Code build pack assembled (CLAUDE.md, pyproject.toml, seven task files) for the Lawler collaboration
• Built in collaboration with Michael Lawler (Cornell/Binghamton), progressing from earlier Li-electrolyte conductivity modeling (v3 DKRR, R²=0.974) as a proof of concept

Milestones

1. 01Phase-1 falsification target: gravitational-wave quasinormal-mode pole recovery
2. 02Phase-1 falsification target: cascade damage scaling
3. 03Execute the seven-task Claude Code build pack with Lawler
4. 04Validate swarm-of-primitives against a mixture-of-experts baseline on the GW domain
5. 05Extend to a second domain (condensed matter or materials) once GW phase-1 passes

Open questions

• Whether swarm-of-primitives generalizes across all three target domains cleanly, or needs domain-specific primitive weighting
• How falsification results feed back into primitive selection and composition
• Compute budget for multi-domain training given the collaboration's resources