# Microsoft expands MatterSim with a multi-task materials model > MatterSim’s May update combines a new multi-task model, simulation improvements and experimental follow-up on a predicted thermal conductor. Canonical URL: https://www.devobs.io/news/news-mattersim-mt-materials-update/ By: Jonah Reed Published: 2026-09-06T11:58:54.630Z Updated: 2026-09-06T11:58:54.630Z Event date: 2026-05-12 Section: AI Microsoft Research announced a MatterSim update on May 12 that combines a new multi-task model, improvements to its existing simulation software and experimental follow-up on a materials prediction. The [research post](https://www.microsoft.com/en-us/research/blog/advancing-ai-for-materials-with-mattersim-experimental-synthesis-faster-simulation-and-multi-task-models/) links computational screening with subsequent laboratory work. ## Simulation and experiment appear together The update introduces MatterSim-MT for predicting several material properties, alongside energy, forces and stress. Microsoft also described faster MatterSim-v1 inference and integration with LAMMPS. In a collaboration with university researchers, a screened tantalum-phosphorus material was synthesized and its thermal conductivity measured. These are related parts of one announcement. The experimental result concerns a particular candidate and samples; it does not establish that all model predictions have been validated or that the material is ready for commercial deployment. ## Preserve the chain of evidence Our suggested workflow would keep each candidate’s predicted properties, simulation settings and later measurements together. A screening score should remain distinguishable from a physical observation. That distinction becomes especially important when a result moves from a research notebook into a planning document. Before replacing a simulation component, select representative reference cases and compare the properties that matter to the intended study. Faster inference is useful only if the resulting accuracy remains acceptable for those decisions. ## Define the next experiment A practical evaluation should identify which uncertainty the model helps reduce. It might narrow a candidate list or suggest conditions worth testing, rather than provide a final materials-selection answer. Record unsuccessful follow-up experiments as carefully as positive ones. They help reveal where a model’s useful domain ends and prevent later users from seeing only the candidates that worked. The announcement is most informative as an example of a computational-to-experimental workflow with traceable stages. ## Source references -