Switzerland's Empa Releases an AI Agent That Calculates Alloys on Chat Request
Experimental materials scientists rarely have time to learn complex simulators, so Empa's lab built OptiMat Alloys, a chat agent for calculating high-entropy alloys. Users type requests in plain language, and the agent picks the software, runs ready-made workflows, and returns tables and charts for an expert to check. The numbers come from the tools, not the language model, and every calculation is saved to a database. That makes it harder for the agent to hallucinate.
AI-processed from TechXplore AI; edited by Hamidun News
Imagine you need to pick the composition of an alloy made from five metals in roughly equal proportions. The possible combinations are practically endless, and testing them in the lab by casting samples one after another is slow and expensive. Computer simulation can help, but not many people know how to use it. Researchers at Empa, the Swiss laboratory in Thun, decided to address this by putting an AI agent you can chat with in front of the simulators.
The problem with high-entropy alloys
Everyone knows the classic example of an alloy: take copper, add a bit of tin, and you get bronze. A base of one or two metals plus a few additives. High-entropy alloys work differently: they mix four or more metals in roughly equal amounts. These materials are interesting for their mechanical, thermal, and electromagnetic properties. They are considered for heat-resistant structures, coatings, catalysis, electronics, and magnetic components.
But the number of combinations is too large, and the old approach of making a few samples and seeing what happens quickly hits a ceiling. So researchers increasingly start with computer calculations and pick the most promising compositions for experiments. The catch is that the tools for these calculations demand serious expertise. According to Vladislav Turlo of Empa, experimentalists simply don't have time to learn these programs.
How OptiMat Alloys works
Turlo and his colleague Yan Hu built a platform called OptiMat Chat. Through it, a scientist downloads an "AI materials scientist" onto their own computer. The first such agent is called OptiMat Alloys. It understands requests in everyday language, selects the appropriate specialized tools, runs predefined calculation workflows, and formats the output as charts, tables, and text in the style of a scientific paper. The user does not need to program or configure complex parameters.
The interface resembles ChatGPT, but the language model here is only an intermediary. This is an actual agent: it can call external specialized programs and work with their inputs and outputs. The calculations run on the user's machine, not in the cloud.
For engineers building their own agents, the most interesting part is the approach to reliability. The authors describe three measures against hallucinations:
- structured tool calls instead of free-form "reasoning" by the model;
- predefined calculation workflows built on accepted best practices;
- storing records of each calculation.
Scientific results come from the connected programs, databases, and validated procedures, not from the LLM itself. In addition, all results go into a database so they can be traced, reused, and later compared or improved. According to the authors, similar AI tools usually lack this.
Limitations and what's next
The developers openly warn that articles should not be published based solely on the agent's output. It is a starting point for expert interpretation and experiment planning, not a finished scientific result. The agent may suggest that material A is better than material B for property C, but only humans can explain why and verify it.
The material includes no comparative measurements of accuracy or speed, so the gain cannot yet be quantified. There is a description of the approach, a link to the article in npj Computational Materials, and a preprint on arXiv.
Next, the team wants to add new agents to OptiMat Chat: for ceramics and metal oxides, as well as for amorphous materials. The authors argue that AI will take over routine calculations, freeing people for genuinely interesting scientific questions.
The practical takeaway for those building agents in narrow fields is that the working design is modest. The model understands the request and selects the tool, verified programs do the calculating, every step is logged, and the result always remains for human review. For science, medicine, and engineering, this design will likely prove more reliable than trying to make the LLM "know the answer" on its own.
- Nov 21, 2026Apple Is Building an AI Agent and Making It Harder for Outsiders to Get Access: Permission Now Takes 14 Steps, Not 4
- Nov 21, 2026Developer Skips Pinecone: Searching 72,450 Code Chunks in 50 ms
- Nov 21, 2026Gemini 4 Neon shrank the TS Rust binary from 47.44 to 7.97 MB in under 10 hours
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
Key Questions Answered
What are the key details on The problem with high-entropy alloys?
Everyone knows the classic example of an alloy: take copper, add a bit of tin, and you get bronze. A base of one or two metals plus a few additives. High-entropy alloys work differently: they mix four or more metals in roughly equal amounts. These materials are interesting for their mechanical, thermal, and electromagnetic properties. They are considered for heat-resistant structures, coatings, catalysis, electronics, and magnetic components.
How OptiMat Alloys works?
Turlo and his colleague Yan Hu built a platform called OptiMat Chat. Through it, a scientist downloads an "AI materials scientist" onto their own computer. The first such agent is called OptiMat Alloys. It understands requests in everyday language, selects the appropriate specialized tools, runs predefined calculation workflows, and formats the output as charts, tables, and text in the style of a scientific paper. The user does not need to program or configure complex parameters.
What are the key details on Limitations and what's next?
The developers openly warn that articles should not be published based solely on the agent's output. It is a starting point for expert interpretation and experiment planning, not a finished scientific result. The agent may suggest that material A is better than material B for property C, but only humans can explain why and verify it. The material includes no comparative measurements of accuracy or speed, so the gain cannot yet be quantified.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.