Accretion CoreWORKING NAME
One system that takes a target property and returns lab-ready candidate materials — with predicted values, confidence, and synthesis routes.
This is a mockup of the interface we're building toward, not a live query — Accretion Core has no measured results yet. The build log below tracks what's actually in progress.
Generate, predict, screen, verify — then loop.
We build on published technique, not proprietary magic.
Three ideas, each independently validated by other labs, compose into the platform above.
Generative modelling, run backward
Instead of testing materials one at a time, diffusion-based generative models take a target property as input and propose an atomic arrangement designed to hit it — inverse design rather than trial and error.
Graph neural networks as surrogates
Atoms become nodes, bonds become edges. Trained on simulation data, these networks estimate how a structure behaves under heat, pressure, or an electric field in milliseconds instead of hours.
DFT-in-the-loop active learning
Every high-throughput prediction is checked against Density Functional Theory for the candidates that matter, and the result feeds back into training — so the model keeps sharpening on its own blind spots.
Materials Discovery Platform
A natural-language property query is translated into constraints over composition and structure space. Generative models propose novel candidates; screening models filter them against stability and your stated targets. Results feed back into training, so each screen improves the next.
Property Prediction Engine
Graph neural network surrogates trained on DFT simulation data estimate electronic, thermal, and mechanical properties in milliseconds — replacing simulations that take hours per structure.
Synthesis Route Generation
For each shortlisted structure, models trained on published synthesis literature propose precursor sets, reaction conditions, and processing windows — so candidates arrive lab-ready, not just theoretical.
Scientific Data & APIs
The datasets, benchmarks, and models behind the platform are exposed through documented endpoints, so your computational team can run screening inside its own workflows.
Datasets, benchmarks, and inference APIs.
Curated simulation and experimental datasets, published benchmark results, and low-latency inference endpoints that integrate with your existing R&D pipeline.
# the request shape we're building toward — this endpoint doesn't exist yet curl https://api.accretionlabs.net/v1/predict \ -H "Authorization: Bearer $KEY" \ -d '{ "structure": "Li7La3Zr2O12", "properties": ["ionic_conductivity", "band_gap", "e_hull"] }' # → 501 Not Implemented { "status": "in_development" }
Designed to sit inside an existing lab, not replace it.
Your data stays yours
Screening runs against your target properties without your proprietary formulations leaving your environment unless you choose to share them.
Team workspaces
Shared runs, comments, and shortlists per project — so a screen a chemist starts is a screen the whole team can pick up.
Audit trail by default
Every prediction is versioned against the model and dataset that produced it, so a result from six months ago is still explainable today.
API-first
Everything in the interface is also an endpoint — batch screening slots into a pipeline your computational team already runs.