AI for materials R&D · validation · pooling · security
AI designs materials faster than labs can make them.Synzcuor predicts which ones can actually be made.
The answer is in failed syntheses, and laboratories never publish them. Synzcuor’s validator learns from them where they live, pooled across laboratories without any data leaving its owner.
Pre-seed · pre-product · team forming · public-data results only
6 of 6
size-matched validation runs in which the model outperforms both baselines on newly reported compounds
34–40%
reduction in prediction error from pooling, relative to the best single laboratory
3.2×
training-time overhead of secure aggregation over a network, with no measurable loss of accuracy
All results are from public benchmark data with simulated laboratories. No proprietary data has been used. Each result is reported with its method and limitations.
Full results and limitations →The gap
The field is optimising the wrong step
Generative models keep improving on simulation benchmarks. Simulations can say a crystal should exist; only a laboratory can say it can be made, and synthesis is where the weeks are lost.
380,000
stable crystals predicted by Google DeepMind's GNoME (2023)
736
of them made independently by laboratories at the time of publication
~4 in 10
designs from the best-known generative models that pass simulation checks for stability and novelty, which never test whether a laboratory can make them
What we build
A validator, fed by private pooling
Validation
Predicts which candidate materials can be synthesised, learning from the laboratory records nobody publishes: failed experiments.
StatusBeats both baselines in 6 of 6 size-matched runs on public data; no failure data yet
Pooling
How the validator gets its data. Every laboratory's attempts, failures included, train one shared model that no single laboratory could build alone.
Status34–40% lower error than the best single laboratory, on public benchmarks
Security
Why laboratories can contribute. Data never leaves its owner; updates are masked, so no party, Synzcuor included, sees an individual contribution.
StatusDemonstrated over a network; limitations documented
The mechanism
Your data stays on your machines
Each participant trains on its own hardware, behind its own firewall. Only model updates leave, never the data. Each update is masked with random values that cancel when all updates are summed, so the aggregating server sees only the total, never an individual contribution.
The data never moves
Raw measurements are never transmitted, in encrypted form or otherwise. They remain on infrastructure controlled by the organisation that produced them.
Updates are masked before they leave
Model updates can reveal information about the underlying data, so each is masked through secure aggregation before it leaves. This conceals individual updates from the server. It does not prevent the trained model itself from revealing information; our threat model states where that boundary lies.
Each participant receives a private model
Each model is fine-tuned on the participant's own data, with the aim of outperforming anything it could train alone. In simulation this holds for two of three members; the strongest specialist is not yet better off, and closing that gap is ongoing work.
What you keep
Proposed ownership: your model is yours; the shared component is held neutrally
Most materials models divide into a shared component, which learns general patterns from all participants’ data, and a final component fitted to each participant’s property and instruments. We propose drawing the ownership line at that boundary. These terms are proposals; no party has signed them.
Yours, outright
- —Your raw data, which never leaves your control
- —Your fine-tuned model and all of its outputs
- —All discoveries made with it; we claim no downstream rights
- —A permanent copy for offline use
- —An escrow claim on the shared component if Synzcuor fails or is acquired by a competitor
Held by us, as custodian
- —The shared component, trained across all participants
- —The training pipeline and the harmonisation of data across laboratories
- —The evaluation of whether pooling benefited each member
The shared component must be held by a neutral party. No member of a pool of competitors can hold it, because the others would be entrusting a rival with an asset built partly from their own data.
Where it starts
Batteries first. Then anything made by solid-state synthesis.
The validator on public data
Trained on published syntheses and tested on compounds first reported after its training cutoff. It runs today as research code.
A first laboratory's own records
Installed on the laboratory's machines and trained on its own attempts, failures included. Initial market: solid-state battery electrolytes.
Pooled across laboratories
Every laboratory's failures improve everyone's validator, through secure aggregation. The same problem extends to semiconductor materials, catalysts, alloys, magnets and ceramics.
Why this is a company
The best-known precedent was a fixed-term research project
Continuity
MELLODDY, in which ten pharmaceutical companies trained shared models without sharing data, ran as a three-year research consortium. A shared model retains its value only if it continues to train on new data, which a fixed-term project cannot sustain.
Neutrality
Laboratories will not hand their failures to a company racing them to the same discoveries. Synzcuor does not run its own discovery programme; it serves the laboratories that do.
The mechanism is open
We intend to publish the method and code so participants can verify that we cannot access their data. The code is not yet public. The value lies in the trained model, which cannot be reconstructed from the code.
Current status
Most of this is not yet built
What exists: a co-authored preprint (in revision after peer review), a validator that beats both baselines on newly reported compounds in every size-matched run, and private pooling that runs over a network. All on public data; no proprietary data has been used, and the validator has no failure data yet. The next milestone is a first laboratory contributing its attempt records. If real records do not improve the validator, we will publish that result and stop.
Working with synthesis laboratories and battery R&D teams
We are seeking a first laboratory to run the validator on its own attempt records, failures included, in a free pilot on its own machines. Well-argued criticism of the approach is equally welcome.