A provably-correct, no-AI, GPU-accelerated molecular DESIGN and TRIAGE sampler over a fixed scaffold plus attachment-site combinatorial chemical space (materials/polymers by default, a small-molecule aromatic regime too).
A provably-correct, no-AI, GPU-accelerated molecular DESIGN and TRIAGE sampler over a fixed scaffold plus attachment-site combinatorial chemical space (materials/polymers by default, a small-molecule aromatic regime too). It PROPOSES candidates with a symmetric single-site Metropolis-Hastings walker that provably targets the Gibbs measure exp(-beta V) - the exact detailed-balance residual of the kernel is measured and reported as the trust verdict - and VERIFIES every candidate exactly against cited methods: molecular weight from standard atomic masses, logP by Hansch-Leo 1979, TPSA by Ertl 2000, pKa by Hammett where applicable, and HBD, HBA, rotatable bonds and ring count by their standard graph definitions, with Lipinski 2001 and Veber 2002 as derived flags. The whole index space is enumerated and scored on the GPU for throughput, and the log-space partition function is re-reduced on a 16-limb GeoNum GPU tree. There is NO learned model, no training, and no tuning of the energy to the target. The GTI verdict is on the COMPUTATION, never on whether a molecule is good, novel, or synthesizable, and no claim is made that the geometric substrate is chemically special (mesh57/HVP benchmarked as information-equivalent to Weisfeiler-Leman, a tested-negative result). This is the pre-cluster design step that runs BEFORE expensive HPC docking and FEP, not a substitute for them.
Open it inside GDBS to save runs to Sandbox, attach results to a Worklog case, or share through a Gate client portal - all in the browser, nothing uploaded to anyone’s cloud.