Searching for Better Molecular Qubits with a Differentiable Physics Workflow
This case study examines molecular qubits, spin-bearing molecules that could become useful building blocks for quantum sensing and other quantum technologies. Using published vanadyl measurements, the study builds a reproducible workflow to rank candidates, test where the model improves, and show which claims the evidence can actually support. It reports a meaningful gain from differentiable physics and a deterministic 125-candidate search that surfaces promising leads for follow-up. Just as importantly, it draws a clear line between a strong screening result and a verified molecular-qubit discovery.
We took a molecular-qubit discovery workflow to full scale: ~150 real, sourced literature measurements, a vanadium electron-spin-property corpus, surrogate models trained end-to-end through public Attryx nodes, and a differentiable physics optimization that assembles a semiempirical ligand-field Hamiltonian, diagonalizes it, and back-propagates through the eigensolve. A differentiable physics block nearly quadrupled held-out rank correlation (Spearman 0.13 -> 0.51), and a deterministic 125-candidate search was run under six physical-meaningfulness gates - with every claim scoped to what the public data can defend.
- Spearman 0.13 -> 0.51
- 125-candidate gated search
- Differentiable physics
What the study set out to do
Molecular qubits are single molecules whose electron spin can act as a quantum bit. This study focuses on vanadyl complexes and asks which ligand environments and geometries best support coherence, controllable spin behavior, and future molecular-qubit or quantum-sensing applications.
The workflow was designed to stay honest about claim boundaries: ingest real public measurements, train differentiable surrogates over physically meaningful descriptors, search a bounded candidate space under explicit gates, and only promote claims the data can support.
What the public data allowed
The sourced corpus produced 149 ingested rows but only 55 relaxation-model rows with usable geometry and T1/T2 labels. The anisotropic g/A readout path was blocked because the public corpus carried only one row with all four labels, so the artifact marks that route as blocked instead of pretending the model was trained.
What the workflow found
On held-out ranking, a differentiable physics block nearly quadrupled Spearman rank correlation from 0.13 to 0.51 and cleared the 0.5 gate for exploratory structure-computed relaxation ranking.
The fixed-topology search evaluated 125 candidates under six physical-meaningfulness gates and narrowed the set to five candidates for follow-up, while still drawing a firm line between a strong screening result and a verified molecular-qubit discovery.