news.ycombinator.com• Jun 3, 2026• 1 min read
SpectraLoRA: GNN -> RamanArcadia Science just dropped SpectraLoRA: they take the pretrained 22.5M-param DetaNet equivariant GNN and lightly post-train it with SO(3)-equivariant LoRA (AdaLoRA on invariant layers + ELoRA on tensor products) to predict Hessians, vibrational frequencies, and full Raman spectra straight from 3D atomic coords. Trained on ~2M molecules (QM9, SPICE, etc.), it already hits 60% fingerprint coverage at DFT-level accuracy zero-shot on out-of-distribution drug-like stuff. Then they close the sim-to-real gap with a two-stage alignment trick: a 1D U-Net (FiLM-conditioned on Morgan fingerprints) + Natural Evolution Strategies optimizing a non-differentiable F1 score. Jumps experimental fingerprint F1@15 from 0.426 → 0.532 on their RamanBiolib set. Code + Colab demo are open-source: https://github.com/Arcadia-Science/spectralora-molecular-alignment Paper: https://doi.org/10.57844/arcadia-kn60-kxsk This feels like the kind of lightweight adaptation + clever alignment hack that could actually ma
