The model can show which compounds are likely to appear, including temporary compounds that form during the reaction and unwanted impurities in the final material. According to Berkeley Lab, it can produce these predictions in minutes.
The research could eventually help laboratories reduce some of the trial and error involved in developing materials for electronics, batteries, sensors, and other technologies.
Berkeley Lab announced the research on August 3. The underlying peer-reviewed study was published in Nature Materials on April 28, 2026.
Kristin Persson, one of the study’s authors, said: “Our new model enables material scientists and industry stakeholders to make promising new materials dramatically faster — and with higher purity and yield.”
“The model can help accelerate the advancement of solid materials to cost-effective manufacturing and commercialization. It closes the gap between material discovery and new technologies that benefit society,” Persson added.
Finding a material is only part of the problem
Computer models can help scientists identify materials that may have useful properties. But finding a promising material does not automatically explain how to make it.
Many solid materials are produced by mixing powders and heating them. Small changes in the ingredients, their amounts, the temperature, or the heating time can produce different results.
A mixture may form the intended material. It may also form other compounds or unwanted impurities.
Researchers can therefore spend weeks, months, or even years testing different conditions before finding a reliable way to make a material.
Why existing predictions can fall short
Many earlier models focused mainly on thermodynamics. In simple terms, thermodynamics helps scientists predict which compounds should be the most stable.
However, the most stable compound is not always the one that forms during a real experiment.
Atoms move slowly inside solid materials. A reaction may appear favorable on paper, but it can still fail if the necessary atoms cannot move to the right place quickly enough.
The Berkeley Lab model considers both parts of the process. It looks at which compounds are likely to be stable and uses machine learning to estimate how atoms move while the materials react.
What the model actually does
Researchers enter information about the proposed experiment. This includes the starting materials, their amounts, the reaction atmosphere, and the way the temperature will increase over time.
The model then predicts how the mixture is likely to change during heating. It can estimate which compounds will form first, which ones may appear temporarily, and which products or impurities may remain at the end.
The system does not automatically invent a complete factory-ready recipe. Instead, it examines a proposed set of ingredients and heating conditions and predicts what is likely to happen.
Researchers tested one material system
The team tested the model on materials containing barium, titanium, and oxygen. Compounds within this family are used in electronic applications.
The researchers chose this system because several different compounds can form under similar conditions. The final result depends partly on how easily barium and titanium atoms move during the reaction.
The study compared the model’s predictions with four previously published experiments. In those comparisons, the model closely matched the compounds observed by researchers and how those compounds changed as heating continued.
It also predicted intermediate compounds and secondary phases that could affect the purity of the final material.
Why the research could matter to businesses
If the model can be adapted to other materials, it could help researchers choose more promising experiments before using laboratory equipment, raw materials, and energy.
This could be useful for companies developing specialized materials for batteries, electronic components, sensors, medical devices, and other products.
Better predictions could also help researchers identify conditions that are more likely to avoid unwanted compounds.
However, the study did not test the model on an industrial production line. It also did not measure commercial cost savings, manufacturing speed, or improvements in product yield.
Those are possible future benefits, not results demonstrated by the current research.
The model remains limited
The machine-learning part of the system was trained specifically on the barium-titanium-oxygen material family.
The researchers have not yet shown that it can accurately predict reactions across many different types of materials. The paper also notes that the current method is not suitable for every kind of solid-state reaction.
The team plans to collect data from more material systems. Its longer-term goal is to develop a broader model that could work across several classes of solid materials.
That wider model has not yet been created.
For now, the research is a proof of concept. It shows that combining information about material stability with predictions of atomic movement can produce more realistic simulations of how certain solid materials form.
The approach may eventually help close an important gap in materials research: identifying a promising material is becoming easier, but finding a reliable way to make it can still take considerable time.
