Components from open-source chemistry
FormulaIQ Layer 1 is fed by a multi-source open materials ETL — not a single vendor dump. Components stay CAS-keyed into inverse formulation.
Formula · open data
Open chemistry sources
Technology stack
Every layer chemists need — from open components to the plant floor.
Formula components from open-source chemistry data, ML/DL models we train, LLM evaluation, regulation RAG, workflow simulation & digital twins, process control commissioned from plant data, and Chemical Engineering Agents — a growing fleet for chemistry.
prediction = first-principles baseline + ML residual ± uncertainty
Open components9 sources
ML / DL + LLMtrain · eval
Twin · agentsHITL
Full chemistry stackinspectable
The full technology map
One inspectable stack. Each layer is purpose-built for chemistry — not a generic LLM wrapper.
- 01
Formula components · open-source data
Component-level properties and structures from multiple open chemistry databases — QM9, Materials Project, OQMD, AFLOW, NOMAD, PoLyInfo, Polymer Genome, PubChem, ZINC — normalised onto a CAS-keyed backbone for FormulaIQ Layer 1.
- 02
ML / DL models
In-house machine learning and deep learning: molecular descriptors & GNNs, mixture models, physics-informed residuals, inverse search (Bayesian / multi-objective). Trained by Lavoisier — not rented property APIs.
- 03
LLM evaluation
Chemistry LLMs are evaluated on formulation, process, and regulation tasks before they orchestrate agents — grounded answers, cited sources, and failure modes measured — not vibes.
- 04
RAG for regulations
Local LLM + RAG over REACH, RoHS, market rules, and customer policy docs. Compliance answers cite clauses; inference can stay private.
- 05
Workflow · simulation · digital twin
Studio Pro workflow tools: full-metric process simulation, thermodynamic packages (Ideal, NRTL, Peng-Robinson, UNIQUAC), Monte Carlo risk, Bayesian optimization, and digital twin profiles ready for scale-up.
- 06
Process modelling & control
Models of a live process, built from its own operating data on top of the physics that governs it — with a verdict on whether the data was ever enough to support the model. From one model we commission the controller the loop needs, classical or model predictive, into the control system already running it.
- 07
Chemical Engineering Agents
Agents answer across formulation, process, and plant context under chemistry constraints, incomplete data, and quantified uncertainty — citing the run or the clause behind each answer. Advisory, with humans in the loop.
- 08
Growing agent fleet
A fleet of chemistry AI agents — plant advisory today; formulation, regulation, simulation, and scale-up roles expanding on one backbone.
From open components to models you can trust.
FormulaIQ starts with real components drawn from open-source chemistry corpora, then runs ML/DL forward models and inverse search so recipes stay physically plausible. LLM evaluation gates what agents are allowed to say; regulation RAG constrains what they are allowed to recommend.
See how FormulaIQ uses this stack.
Explore FormulaIQHybrid prediction
First-principles baseline plus ML residual — confidence and out-of-domain flags on every answer.
Physics + ML control
Confidence80
Mixture level+ High
Uncertainty
Data flywheel
Validated experiments
25→200
Simulation, digital twins, process control — and agents.
Studio Pro runs the workflow: simulate the process, extract a digital twin profile, then commission the controller against the real plant. Agents read the same record and advise — plant advisory today, a growing fleet ahead.
Formula confidence
+7.8%80.5%
Twin stability
+4.5%94%
Model readiness
72
Eval score
LLM evaluation72%
Chemical Engineering Agents
- Plant advisory agent
- Formulation advisor agent
- Regulation agent
- Simulation agent
- Scale-up agent
Diligence detail by layer
Open components, ML/DL + LLM evaluation, digital twin workflows, and chemistry agents — the stack investors and PhDs both ask about.
Formula · open data
9
Open chemistry sources
Questions
- Where do FormulaIQ components come from?
- Component-level properties and structures are built from multiple open chemistry databases — including QM9, Materials Project, OQMD, AFLOW, NOMAD, PoLyInfo, Polymer Genome, PubChem, and ZINC — normalised onto a CAS-keyed backbone. Customer formulation data never trains shared models without a written agreement.
- How do ML/DL models, LLM evaluation, and RAG fit together?
- Lavoisier trains ML/DL property and mixture models in-house. Chemistry LLMs are evaluated on formulation, process, and compliance tasks before they orchestrate agents. Regulation answers use a local RAG pipeline over REACH, market rules, and customer policy — with cited clauses.
- What about digital twins, RoboBridge, and Chemical Engineering Agents?
- Studio Pro runs workflow simulation and digital twin profiles with real thermodynamic packages. RoboBridge closes the lab loop with Hardware-in-the-Loop validation. ControlIQ models the plant from its own data and commissions the controller into the system already running it. Chemical Engineering Agents advise across all of it — evaluated LLMs, citations, and human approval for anything that acts.
Walk the full stack on your chemistry.
Open components, ML/DL, LLM evaluation, RAG, digital twins, process control, and chemistry agents — in one technical session.