AI now predicts how a new material is made.
On 3 August 2026 Lawrence Berkeley National Laboratory presented a model that adds diffusion kinetics to thermodynamic reaction prediction and simulates a solid-state synthesis in minutes. What becomes computable is the recipe — the part of materials work that companies normally keep as know-how.
Insights bi îngilîzî tên weşandin.
On 3 August 2026 Lawrence Berkeley National Laboratory presented a computational model that predicts how an inorganic solid actually forms when its ingredients are mixed and heated. The underlying study — Vir Karan, Max C. Gallant, Yuxing Fei, Gerbrand Ceder and Kristin A. Persson, “Ion correlations explain kinetic selectivity in diffusion-limited solid-state synthesis reactions” — appeared online in Nature Materials on 28 April 2026 (volume 25, pages 1385–1392), following a preprint circulated in January 2025.
The problem it addresses is narrow and, for anyone who manufactures with materials, familiar: screening now proposes candidate compounds far faster than laboratories can make them. Berkeley Lab puts the consequence plainly: finding the right recipe can take “weeks to even years of trial-and-error experimentation”. The compound is identified long before the process to produce it is. What is new here is which part of the reaction a model predicts.
Thermodynamics says which reactions are favourable; diffusion decides which ones occur
Earlier pathway-design methods navigated the thermodynamic landscape using first-principles data, comparing reactions pair by pair. Thermodynamics, in Kristin Persson’s formulation, describes “how much two solids ‘want’ to react with each other”. That is sufficient when one product is clearly favoured, and not when several competing phases have similar formation energies — precisely where synthesis goes wrong, and where the deciding factor is how fast ions can reach the reaction front.
The new framework couples machine-learning-derived transport properties, calculated through the disordered “liquid-like” layer of product that forms between two reacting solids, to a thermodynamic cellular reaction model. It was tested on the barium–titanium–oxide system, chosen because an unusual number of ternary phases sit on or very close to the convex hull of stability; barium titanate is a ferroelectric. Across varying BaO:TiO₂ ratios, the model predicted which phases form as a function of time and temperature. The authors report that cross-ion transport coefficients — how the movement of one ion species is correlated with another — are decisive for this selectivity.
The model takes three inputs: the starting materials, their ratios and the temperature ramp-up. In minutes it returns the full reaction pathway, including intermediate compounds, the final products and the impurities formed along the way. Checked against decades of published experiments, in the range typical of this chemistry — roughly 1,000 to 1,300 °C — it reproduced the sequence of events, impurity formation included. Berkeley Lab describes it as the first predictive model to account for how atoms travel through materials during solid-state reactions.
The bottleneck stopped being discovery some time ago
The scale of the mismatch is documented. On 29 November 2023 Google DeepMind announced GNoME, which predicted 2.2 million new crystal structures, of which 380,000 were judged the most stable; the same announcement recorded 736 of them as independently created by external laboratories. On the same day Nature published the A-Lab, an autonomous laboratory built by an overlapping Berkeley team, which combined computational screening, literature-derived synthesis heuristics and robotics, and reported 41 compounds from 58 targets over seventeen days of continuous operation. An Author Correction of 19 January 2026 revised that record: after concerns about the identification of the structures by diffraction, the authors re-analysed the patterns and confirmed the correct conclusion in 36 of the 40 reported successes, four being inconclusive, and clarified that the compounds were new to the prediction platform, not necessarily new to science.
Set the orders of magnitude side by side: candidates in the hundreds of thousands, confirmed syntheses in the hundreds. Automation raised the throughput of the laboratory; it did not explain the failures. The 2026 model is aimed at that residue — not at making more attempts, but at making fewer wrong ones. The United States Materials Genome Initiative defines its objective as discovering, manufacturing and deploying advanced materials “twice as fast and at a fraction of the cost”.
What is becoming computable is process knowledge, and that is a different kind of asset
For a manufacturer, the operative word in the paper is not “materials” but “pathway”. The composition of a finished ceramic component can be determined by analysis. The route that produced it — precursor grades and particle sizes, ratios, ramp rates, dwell times, atmosphere — usually cannot, which is why firms keep it as know-how rather than patenting it. A model that reconstructs the sequence of intermediates from precursors and a temperature programme brings part of that asset within reach of anyone with access to the computation.
Two qualifications belong with that, both from the authors. The demonstration covers one oxide system, in powder form, by solid-state reaction; it says nothing yet about alloys, polymers or solution routes. And the general-purpose tool does not exist: the stated next step is to train the model on other classes of solid-state materials, and eventually to build a foundation model from large kinetics datasets.
So the near-term consequence is not that recipes become public, but that the cost of reconstructing someone else’s falls, and that the period over which process secrecy holds as a defence is likely to shorten.
In Europe the step after synthesis is administrative, and it has not become faster
A shorter development cycle does not shorten the route to market. In the European Economic Area, a company that manufactures or imports a substance in quantities of one tonne or more per year must register it under REACH, and the European Commission’s guidance for businesses puts it without qualification: “Non-registered substances must not be marketed or used.” For a Swiss firm selling into the EU, this is the binding constraint on a new material, and the one part of the timeline computation does not compress.
The computational side, by contrast, has a Swiss address. Alps, the system operated by the Swiss National Supercomputing Centre and inaugurated in September 2024, is geo-distributed across Lugano, Lausanne, Villigen and Bologna and is credited with 434.9 petaflops of sustained peak performance, with 2,688 NVIDIA Grace-Hopper nodes. The methods layer has a Swiss lineage too: NCCR MARVEL, funded by the Swiss National Science Foundation, ran for twelve years on the computational design and discovery of materials and held its closing event in Lausanne on 9 July 2026, leaving open-source codes and open-access datasets.
What a Swiss company can put in motion this quarter
- Start capturing the process data you already generate. Time–temperature profiles, precursor lots and particle sizes, atmosphere, and the diffraction result for each firing. Kinetic models are trained on exactly this kind of record, and what stays in a logbook rather than a database cannot be used later.
- Buy a pre-screen before the next laboratory campaign. The Innosuisse innovation cheque covers 100% of a Swiss research partner’s costs up to CHF 15,000, for firms with fewer than 250 full-time equivalents holding a Swiss UID that have not drawn one in the previous two years. The decision takes four to six weeks and the study must be completed within six months. The research partner has to be identified before applying.
- Size the larger step honestly. In an Innosuisse innovation project with an implementation partner, the company carries 40 to 60% of total project costs and pays at least 5% of that total in cash to the research partner, with a decision in six to eight weeks. The Innovation Council meets quarterly; the remaining 2026 dates for Energy and Environment are 11 September, 30 October and 27 November.
- Re-examine how the recipe is protected. If the commercial defence of a product rests on an undocumented firing schedule, establish how long that schedule would survive being modelled, and whether a patent, a supply agreement or a customer qualification would hold longer.
- Put registration on the development timeline. One tonne per year into the EEA is the trigger; the paperwork does not accelerate with the chemistry.
Two things would change the picture, and both are observable. The first is the extension of the method beyond barium–titanium oxides to other classes of solid-state materials, which the team has announced as its next step, without a date. The second is independent replication by groups outside the Berkeley orbit, which is what turns a demonstration into a tool. The sober reading is the useful one: a step that used to be found by experiment can now, in one chemistry, be calculated in advance, and the firms placed to benefit first are those that have kept a record of how they make things.