Stop computing what you are not going to synthesise.
Ion-Flow ranks your candidate structures before the NEB/DFT runs, and hands you a shortlist together with its written limits. You then decide where the expensive compute goes — not before.
Output: CSV + Markdown + limits report + data attribution. Cost stated before spending.
(15-material benchmark)
rejected
(5-seed mean, n=260)
(CHGNet, RTX 3080, reference recipe)
The problem is not the compute. It is what you compute.
A DFT NEB campaign costs hours of compute per material. Run on a candidate list built by similarity or by intuition, it spends most of its budget on structures that never had a chance.
Everything costs the same
Every candidate costs the same compute time, whether it is promising or an obviously poor conductor. The ranking arrives after the money is gone.
You choose the spending
An ordered shortlist, a cost estimate before launch, and a report that says what the ranking is worth — and what it is not. Expensive compute only runs on what you decide.
What you receive
A dossier, not a dashboard. It is written to be read by an engineer and discussed in a meeting.
- Ranked shortlist — CSV, Markdown and JSON, with estimated barrier, provenance and calculator backend.
- LIMITS.md — what the numbers are, the known false positives present in your shortlist, and what the deliverable does not authorise.
- Cost plan approved before spending — number of structures, GPU hours, price at an explicit rate.
- Data attribution — source and licence (CC BY 4.0 for Materials Project) attached to the exports.
- Verified reproducibility — reference barriers are recomputed on the execution machine; if they drift, the run fails instead of shipping.
Actual excerpt from LIMITS.md
## What the numbers are - Activation energies are **chgnet-v0.4.2-cuda proxies**, not DFT values. - Each material gets the barrier of **one hop**, not a macroscopic conductivity. - **No experimental data is used.** There is no measured conductivity sigma. ## Known false positives present in this shortlist - **Li3PO4** — poor Li conductor in the literature; this funnel surfaces it because barrier + percolation do not capture the prefactor/path chemistry. ## What this deliverable does NOT authorise - Presenting any row as a **discovered material**. - Publishing a **conductivity value** as a literature result. - Starting a **synthesis programme** without DFT validation and experimental confirmation.
Generated automatically from the run. It is part of the deliverable, not an appendix.
Evidence before promises
The pipeline is first evaluated on known electrolytes: if it cannot reproduce the literature, it cannot claim to rank the unknown.
| Material | Expected | Computed barrier | Verdict |
|---|---|---|---|
| Li6PS5Cl argyrodite | superionic | 0.103 eV | surfaced |
| Li10GeP2S12 LGPS | superionic | 0.065 eV | surfaced |
| Li7La3Zr2O12 LLZO | superionic | 0.147 eV | surfaced |
| Li3InCl6 | superionic | 0.058 eV | surfaced |
| Li3N | superionic | 0.138 eV | surfaced |
| LiNbO3 | poor conductor | 0.949 eV | rejected |
| LiTaO3 | poor conductor | 1.096 eV | rejected |
| Li2O | poor conductor | 0.692 eV | rejected |
The shortlist pre-ranker is a crystal-graph model plus descriptors, evaluated 5-fold across 5 seeds with out-of-fold predictions: Spearman 0.655 ± 0.021 (n=260 labels), and 0.693 for the ensemble mean with a 95% CI of [0.61, 0.76]. The intervals are published because at this size a single number proves nothing.
The limits, written before you ask
A pre-screening deliverable that does not say what it cannot do is a sales document, not a technical one. Here is what the pipeline does not do.
Assumed limits
- Proxies, not DFT. Barriers are computed with a machine-learned potential (CHGNet). The potential's error propagates into the ranking.
- One hop per material. The figure is the barrier of one hop, not a macroscopic conductivity.
- No experimental data. No conductivity σ, no synthesis validation.
- One documented false positive. Li3PO4, a poor conductor in the literature, would be surfaced. It is named in the deliverable.
- n=260 labels. The uncertainty is published, not hidden.
What I will never say
- "Discovery of new materials" — that is not what the pipeline does.
- "Validated predictions" — no experimental validation exists.
- "DFT results" — these are MLP proxies, and every deliverable says so.
- A conductivity figure that is not in the report.
The validation pilot
You do not pay to believe me. You pay to verify.
I pre-screen
You give me your candidate list. I return a ranked shortlist, its compute cost and its limits. Nothing is spent on your side yet.
You test
You pick 10 materials from the shortlist and 10 from the "rejected" pile, and run your DFT on them. 20 calculations, not 200.
We decide
If the ranking holds, we move to the full run. If it does not, you have spent 20 calculations and you know exactly where you stand.
From €2,500
Shortlist of 10 to 30 candidates • cost plan before spending • LIMITS.md • control protocol • data attribution. Duration: 3 to 5 days.
Three pieces of information
The number of candidates • the cost of one DFT NEB in your lab • whether the structures are confidential (if so: contracted GPU, no third-party hardware).
Let's talk about your list.
Send me the number of candidates and the order of magnitude of the cost of one calculation in your lab: I reply with a plan, a cost and a date — or I tell you it is not worth it at that size. It has happened, and it is more useful than a quote.
What you get first
An estimate for your batch (number of structures, GPU hours, cost), and a frank opinion on whether it is worth doing. No commitment before the pilot.
The Materials Project data used is licensed CC BY 4.0; exports carry their attribution.