Sword Labs
Physics-informed machine learning for quantum materials, superconductors, and biomolecular discovery.
Sword Labs is an independent research operation. We build machine-learning systems that learn corrections to first-principles physics rather than learning physics from data. The architectural commitment — physics as residual input, with the exceptional Lie algebras G2 ⊂ F4 ⊂ E6 ⊂ E7 ⊂ E8 constraining the functional form of the physics heads — lets a single backbone generalize across chemistry domains.
The lab’s working thesis is that the parameter-to-data ratio of modern deep learning only closes for chemistry when the parameters are not trying to learn physics from scratch. We give the model the physics; it learns the experimental correction map.
Drug–protein binding
Meridian, our production binding model, trains on 6.2 million molecules across 30,000 protein targets. The V33 generation reaches val MAE 0.807, val Pearson 0.741, and out-of-distribution Pearson 0.462 on the curated drug panel.
Materials
The V34 materials line fuses the V33b2 molecular representation with a SchNet–rcut15 atomic foundation encoder. On the QMOF benchmark, the production stage reaches hse06 R² = 0.78. The molecular representation outperforms RDKit on downstream materials tasks — same encoder, no fine-tuning.
Superconductors
The v5 multi-family lambda-factor model predicts spin-fluctuation coupling and Migdal–Eliashberg Tc across cuprate and pnictide families. Leave-one-out MAE: 6.1 K (pnictide), 11.4 K (BSCCO), 15.6 K (YBCO), 23.2 K (Tl-cuprate). Three actinide-substituted candidate compositions are predicted at or above the liquid-nitrogen boundary, with phonon DFT validation in progress.
The shared backbone, Juniper, combines G2 rotation features, a per-position singlet head, an exceptional-algebra gate, sector fusion across Cartan / Short / Long routings, and a one-loop vacuum coupling of 249/248 from the F4 embedding. Physics decomposition heads (thermo, kinetic, electrostatic, desolvation) produce deterministic outputs from the joint representation; the network learns the residual between physics and experimental ground truth, weighted by per-dimension learnable trust gates.
We commit to numeric predictions before running experiments and publish negative results. A recent four-week development period recorded 33 runs killed or marked no-ship after failing pre-committed acceptance criteria. Falsification is the central method, not a failure mode.
Sword Labs is currently seeking research partnerships with synthesis and characterization groups working in cuprate and topological superconductors, and methods funding for the materials and binding lines. A complete model inventory, supplementary results, and IP disclosure are available on request.