DeepChemical
DIVISION FILE 727-DC
Chemical
Predictive models for formulated matter. Graph neural networks, physics-informed regressors and spectral encoders, trained in-house on 4.1 million proprietary formulation records. Stability, efficacy, cure, emission and cost are predicted before a batch is mixed.
language model
wrapper.
We hold the weights. Architectures are selected per property and trained from initialisation on the client's own laboratory and plant history — not prompt engineeringxxxxxxxxxxxxxxxxxx, not retrieval over a vendor's model. Validation runs against held-out campaigns. Delivery is a versioned artefact that executes on the client's own premises.
Formulated products fail on properties, not on molecules. A gypsum board fails on set time and flexural strength. A serum fails on emulsion stability at 40 °C. A filter medium fails on pressure drop after 2,000 h of loading. We model the property, with the recipe, the process window and the raw-material lot as inputs.
Representation is the work. Ingredients enter as molecular graphs with learned descriptors; mixtures as attention-pooled sets under composition constraints; process history as sequence. Spectra are encoded by a 1-D convolutional trunk and fused late, so a model can be interrogated when it is wrong.
Every prediction carries an interval. Ensembles and evidential heads report uncertainty, and the pipeline refuses to extrapolate outside the hull of its training corpus. It says so, and proposes the smallest set of experiments that would close the gap.
The loop closes in the plant. Bayesian optimisation proposes the next batch, the laboratory runs it, the result returns to the corpus overnight, the model is re-fitted and re-versioned. Median campaign: 41 experiments where 300 were budgeted.
Sectors under contract span cosmetics and personal care, gypsum and building materials, agriculture and crop inputs, air filtration media, and industrial coatings — five of seven annexes released. The corpus is not transferable between clients; each is trained, stored and served in isolation, and the weights remain the client's on termination.
to agriculture to air.
RETAINED 24 MONTHS
Skeletal structures are drawn from the released subset of the reference corpus. Sector tag indicates the property family the structure participates in; two specimens are withheld in full, as their identity is recoverable from the substitution pattern alone.
not the buzzword.
Full benchmarks, per-sector case records, architecture notes and reference calls are released under mutual NDA. A property you cannot predict, a campaign that costs too much, a specification you keep missing.
Materials released under this file are confidential and may not be reproduced in whole or in part.
MILAN · NEW YORK (COMING SOON) · DC@727.SERVICES