🔬 materials science

MatOS: An Agent Operating System for Materials Research

This paper introduces MatOS, an agent operating system for materials research that utilizes a layered scientific asset graph, stateful adaptive workflows, and evidence-driven asset evolution to autonomously manage complex, long-running scientific tasks, demonstrating superior performance over existing agents and successfully executing diverse multi-stage materials discovery studies.

Original authors: Dingyang Lyu, Jing Gao, Zhonghao Chang, Jingjie Huang, Xiangran Cheng, Xin He, Yongda Huang, Zefeng Li, Jiahui Liu, Jian Wen, Feng Yu, Bonan Zhu, Yongheng Li, Hongwei Du, Baole Wei, Ziheng Lu

Published 2026-10-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dingyang Lyu, Jing Gao, Zhonghao Chang, Jingjie Huang, Xiangran Cheng, Xin He, Yongda Huang, Zefeng Li, Jiahui Liu, Jian Wen, Feng Yu, Bonan Zhu, Yongheng Li, Hongwei Du, Baole Wei, Ziheng Lu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The search for new materials—substances that can store more energy, conduct electricity better, or withstand extreme heat—has long been a slow, fragmented process. It relies on a vast, scattered collection of data, computer simulations, and laboratory experiments. A single study often requires a researcher to spend days or weeks moving between reading scientific papers, running complex calculations on supercomputers, and interpreting the results to decide what to try next. While artificial intelligence has recently begun to help automate these steps, teaching a computer to navigate this chaotic landscape without getting lost or making irreversible mistakes has remained a difficult challenge. The core problem is not just that the computer needs to know how to use a tool, but that it must understand how to keep a long, multi-stage project on track, learn from its own errors, and adapt its methods as new information arrives, all without human intervention at every turn.

To solve this, a team of researchers has built a new system called MatOS, which functions as an operating system specifically designed for materials research. Just as a computer operating system manages files and programs so that applications can run smoothly, MatOS manages the tools, data, and methods that scientists use. Instead of treating every new task as a fresh start, the system organizes research capabilities into a structured library of reusable "assets." These assets include everything from raw data and scientific knowledge to software tools and specific skills. When a researcher asks the system to find a new material, the system does not just guess; it assembles the necessary pieces from this library, creates a clear, step-by-step plan, and executes it while constantly checking its own work. If the plan needs to change because of a new result or a user's correction, the system can pause, adjust the path, and resume without losing its place, ensuring that long-running studies remain reliable and reproducible.

The system was tested on a series of complex research challenges to see if it could outperform existing artificial intelligence tools. In a benchmark designed to evaluate how well an agent can conduct end-to-end research, the system improved upon the previous best-performing agent by more than 11 percent while using the same underlying brain. It also proved more successful than other leading coding assistants when asked to use specific materials science software or solve multi-step scientific problems. The researchers found that the system's ability to keep a record of every step and check its work against scientific rules allowed it to avoid common errors, such as misinterpreting data or failing to connect the results of one calculation to the next. This reliability allowed it to handle tasks that were previously too risky for automated agents, such as running simulations that take days to complete or navigating complex workflows that require constant adjustment.

In eight specific research cases, the system demonstrated its ability to tackle real-world scientific problems. In one major effort, it helped narrow down a pool of more than 30 million candidate crystal structures to just 97 promising candidates for solid-electrolyte batteries, a critical component for safer and more efficient energy storage. In another case, it generated thousands of variations of a specific type of salt and identified 89 new versions that showed higher electrical conductivity than their original forms. The system also helped scientists understand why certain computer models for predicting atomic behavior were making errors, linking those mistakes to specific local environments around atoms. Furthermore, it assisted in developing a new method for selecting training data that reduced errors in potential energy models by 3.1 percent, and it created a new model for predicting the melting temperatures of alloys that was 13.6 percent more accurate than the best existing methods.

The researchers also used the system to reproduce a classic fluid dynamics simulation, confirming that it could handle continuous flow problems and even spot inconsistencies in the data used for the benchmark. Throughout these tasks, the system did not just follow a rigid script; it evolved. After each task, it reviewed its own performance and user feedback to refine the assets it used. If a tool behaved unexpectedly or a method proved inefficient, the system updated its internal library to prevent the same issue from happening again. This process of learning and adaptation meant that the system became more capable with every study it completed, rather than requiring a complete rebuild for each new project.

The success of this system suggests that the future of materials discovery may lie in organizing scientific knowledge not as a static collection of facts, but as a dynamic, interconnected web of reusable capabilities. By treating data, tools, and methods as assets that can be combined, checked, and improved over time, the system allows artificial intelligence to carry out long, complex studies with a level of reliability that was previously out of reach. The researchers showed that with the right structure, machines can not only perform calculations but also manage the entire scientific process, from the initial screening of millions of possibilities to the development of new methods for understanding the physical world. This approach offers a path forward for accelerating the discovery of materials needed for clean energy, advanced manufacturing, and national security, turning the chaotic process of scientific research into a streamlined, self-improving workflow.

Technical Summary: MatOS – An Agent Operating System for Materials Research

1. Problem Statement

Materials research is a data-intensive, multi-stage process involving literature analysis, computational simulation, and experimental synthesis. These workflows often span days or weeks, requiring agents to iterate between disparate sources of data, knowledge, and tools. While Large Language Model (LLM) agents have demonstrated the ability to retrieve data and operate simulation codes, three critical limitations hinder their deployment in reliable, long-horizon materials research:

Scalability of Capabilities: New databases, simulation codes, and machine learning potentials emerge continuously. Existing systems often require rebuilding pipelines or adding ad-hoc tool functions to incorporate new capabilities, rather than integrating them seamlessly.
Reliability in Long-Running Tasks: Single studies involve complex dependencies where errors in early stages propagate. Existing recovery mechanisms (e.g., job restarts) often operate on single jobs or saved states without maintaining the validity of the broader research plan when an agent revises its strategy mid-task.
Experience Accumulation: Much of the tacit knowledge required for successful research (e.g., appropriate simulation timescales, failure diagnosis) is rarely codified. Current systems struggle to retain, validate, and reuse experience from past tasks to improve future performance without altering foundation model parameters.

2. Methodology: The MatOS Architecture

MatOS addresses these challenges by functioning as an operating system for research agents, managing capabilities, task execution, and accumulated experience. Its architecture is built on three core components:

2.1 Layered Scientific Asset Graph

MatOS organizes research capabilities as scientific assets, the basic unit of the system. There are six asset types:

Data (D): Records of materials systems, measurements, and simulation outputs.
Knowledge (K): Generalizations, physical laws, and method applicability conditions.
Tools (T): Callable capabilities, including software interfaces and instrument controls.
Skills (S): Methods specifying how knowledge and tools are selected and used.
Roles (R): Agent configurations defined by responsibility prompts and linked assets.
Workflows (W): Reusable process descriptions organizing tools, roles, and sub-workflows.

These assets are structured in a layered scientific asset graph ( $G = (V, E, \tau)$ ) with four semantic layers: Data (base), Knowledge/Tools, Skills, and Roles/Workflows (organizational). Assets are connected via typed relations (dependency, mount, call, containment, evidence, verification). This structure allows new capabilities to be added by registering assets rather than rebuilding the agent, enabling roles to compose necessary data, skills, and tools dynamically from their subgraph.

2.2 Stateful Adaptive Workflow

To ensure reliability in long-running tasks, MatOS executes research tasks as an explicit execution graph rather than relying solely on LLM context.

Node Types: The workflow decomposes tasks into nine node types, including Decision nodes (Agent/Human), Execution nodes (Exec/Backend), Check nodes (Guard), and Control nodes (Subworkflow, Map, Loop, Dynamic).
State Management: The runtime maintains a persistent state ( $S_t$ ) including graph snapshots, node records, and event logs, independent of the model's context. This allows tasks to be paused, resumed, and traced even after interruptions or long external computations.
Checked Updates: When an agent or researcher proposes a change to the workflow (e.g., adding a validation step), the update undergoes three checks: structural (graph validity), policy (configuration permissions), and compilation (runtime validity). Only admitted updates are compiled into the next graph snapshot ( $W^{(k+1)}$ ).

2.3 Evidence-Driven Asset Evolution

MatOS enables the system to evolve by revising assets based on execution traces and user feedback, without modifying foundation model weights. The evolution process follows six stages:

Evidence Critique: Analyzes execution traces and feedback to determine if a problem is plausible and the experience reusable.
Target Location: Identifies the specific assets (e.g., a skill, tool, or workflow) responsible for the issue.
Focused Editing: Generates candidate updates to the identified assets.
Dual Acceptance: Candidates must pass structural checks (ensuring no new errors are introduced) and semantic review (ensuring the revision addresses the evidence and preserves applicability).
Adoption: Accepted revisions are integrated into the asset graph.
Validation: Subsequent tasks and benchmarks verify that the original problem does not recur.

3. Key Contributions

Scientific Asset Framework: Introduced the concept of the "scientific asset" to organize 3,135 datasets, 65 knowledge assets, 194 tools, 386 skills, 51 roles, and 104 workflows into a unified, composable graph.
Stateful Adaptive Runtime: Designed a workflow engine that executes directed acyclic graphs and state machines with persistent state, scientific checks, and validated plan updates.
Evidence-Driven Evolution: Implemented a mechanism for revising assets through critique and dual acceptance, allowing the system to accumulate and reuse experience across tasks.
Comprehensive Evaluation: Validated the system across three benchmarks and eight diverse research cases, ranging from large-scale screening to method development.

4. Experimental Results

4.1 Benchmark Performance

ResearchClawBench: MatOS improved the score of the previously best-performing agent (MIRA) by 11.15% using the same base model (GPT-5.4). It also extended the cost–score Pareto frontier, achieving higher scores at comparable costs.
MatTools: Using a common base model (DeepSeek-V4-Pro), MatOS achieved a 97.96% success rate on complete problems, outperforming Codex CLI (95.92%) and Claude Code (89.80%).
SciAgentGym: MatOS achieved a 59.5% case success rate, exceeding Codex CLI and Claude Code (both 57.1%) by a relative gain of 4.17%.

4.2 Research Case Studies

MatOS was applied to eight research cases, demonstrating its ability to handle complex, multi-stage studies:

Interatomic Potentials: Evaluated 11 pretrained potentials on 7,599 configurations, linking force errors to local atomic environments. Developed RarityGuard-MP, a training-data selection strategy that reduced interatomic-potential errors by 3.1% relative to the DIRECT baseline.
Solid Electrolytes: Narrowed 30.3 million candidate structures to 97 solid-electrolyte candidates through a hierarchical screening workflow involving composition, stability, and molecular dynamics filters.
Halide Variants: Generated 2,211 variants from 23 low-conductivity structures, identifying 89 variants with higher computed conductivity than their parents.
Liquid Electrolytes: Compared solvation and transport in three liquid electrolytes, calculating densities, viscosities, and Onsager conductivities consistent with experimental trends.
Alloys:
- Conducted a composition search for alloys melting near 300°C, identifying candidates within a narrow interval.
- Developed DuoCompNet, a melting-temperature prediction model that achieved a test MAE of 3.18°C, 13.6% lower than the best baseline (Roost).
CFDBench: Reproduced a lid-driven cavity benchmark in continuum fluid dynamics, validating the system's ability to handle non-materials-specific continuum flow problems.

5. Significance and Claims

The paper claims that MatOS demonstrates that organizing research capabilities as reusable, evolving assets enables agents to reliably execute long, multistage materials studies. Key significance points include:

Scalability: By treating new databases, codes, and methods as assets, the system can expand its capabilities without architectural reconstruction.
Reliability: The explicit execution graph with scientific checks ensures that long-running tasks remain traceable and reproducible, even when plans are revised mid-stream.
Continuous Improvement: The evidence-driven evolution mechanism allows the system to learn from past failures and successes, turning tacit research experience into checked, reusable capabilities.
Integration: The same asset and workflow runtime can extend to instruments and experimental protocols, potentially linking computational and experimental evidence within a single study.

The authors conclude that MatOS represents a shift from static agent pipelines to a dynamic, operating-system-like environment where research capabilities accumulate with use, supporting everything from large-scale screening to the development of new scientific methods.