MolBasis
MF-FQSL · the physical-property kernel basis

Physical properties,
as closed-form equations.

Drop a σ-profile in, get a property out — melting point, solubility, hydration free energy, logP, Hansen parameters. No trained model, no black box: every kernel is a closed-form ridge equation over quantum-chemical descriptors, on a ¼-integer exponent grid, validated on scaffold-blind held-out chemistry.

8
A.IMPROVE kernels
shipped, scaffold-blind validated
4
B.UNDERFIT kernels
honestly labelled, not hidden
6,876
σ-profiles in the atlas
wB97M-V / def2-SVP / C-PCM (ε=80, Bondi cavity, Klamt-purge)
¼
integer exponent grid
finer grids add <0.002 r

Closed-form, not trained

ŷ = β₀ + Σ βⱼ·z(fⱼ^pⱼ), solved by a single Cholesky solve. No SGD, no learnable weights, no hyperparameters. We say 'fit' and 'solve', never 'train'.

Quantum inputs only

The 9-dim feature vector comes from one DFT SCF per molecule (wB97M-V/def2-SVP/C-PCM). No empirical group contributions, no fitted correction tables.

Honest by construction

Every kernel carries its 25% scaffold-blind lift and a plain verdict — A.IMPROVE or B.UNDERFIT. We show the corrected numbers, including the ones that went down.

Featured kernels

All 12 kernels →
the credibility moat

We publish the corrections that hurt.

Our melting-point kernel was published at r = 0.826. A later audit found a hydrogen-bond classification bug inflating it — aromatic π-clouds miscounted as donors. The honest, corrected value is r = 0.655. We changed it everywhere. Finding and reporting your own bugs is the whole point of a lab.

Read all the findings →

Have your own property data?

Upload a set of mfsig σ-profiles with measured values and we can fit a kernel against it — same rigor gate, same honest verdict.

Upload your mfsig dataset