NIMS-OS: An automation software to implement a closed loop between artificial intelligence and robotic experiments in materials science

NIMS-OS is an open-source Python library and GUI application that enables fully automated, closed-loop materials research by integrating diverse AI algorithms with robotic experimental systems such as NAREE to autonomously discover new materials like electrolytes.

The Gist

Imagine you are trying to find the perfect recipe for a cake, but instead of a kitchen, you have a massive library containing 4,000 different ingredient combinations. A human chef could try a few, guess what might work, and try again, but it would take years to find the absolute best one.

Now imagine you have a super-smart robot chef and a genius critic who never sleep, never get tired, and never make mistakes. This is exactly what the work "NIMS-OS" describes, but for discovering new materials (such as better battery fluids) instead of cakes.

Here is the simple breakdown of how this system works:

The Problem: The "Needle in a Haystack"

In materials science, scientists want to find the best combination of chemicals to make things like batteries work better. There are so many possible combinations that it is impossible to test them all by hand. It is like trying to find a specific needle in a haystack the size of a mountain.

The Solution: NIMS-OS (The "Conductor")

The authors developed a software system called NIMS-OS (NIMS Orchestration System). Think of this software as a conductor in an orchestra.

• The Musicians: You have various "AI algorithms" (the smart critics) and "robotic arms" (the robot chefs).
• The Conductor: NIMS-OS tells the AI when to make a suggestion and tells the robot when to mix the chemicals. It connects the two so they work together in a perfect cycle without a human needing to press buttons.

How the Cycle Works (The "Three-Step Dance")

The system runs in a continuous cycle, like a game of "Hot and Cold":

1. The Guess (AI): The AI looks at a list of all possible chemical recipes (the "candidate file"). Based on what it has learned so far, it selects the most promising recipes to try next.
   • Analogy: The AI is like a detective saying, "Based on the clues, the culprit is probably in this neighborhood. Let's check these three houses first."
2. The Action (Robot): The software sends a signal to the robot. The robot automatically mixes the chemicals, puts them into a test cell, and performs the experiment.
   • Analogy: The robot is the detective's partner who actually goes to the houses and knocks on the doors.
3. The Result (Update): The robot finishes the test and sends the data back. The software updates the list and marks which recipes worked well and which did not. Then the AI looks at the new data and selects the next best guesses.
   • Analogy: The partner comes back and says, "House A was empty, but House B had a clue!" The detective uses this new information to choose the next house to check.

The Tools in the Toolbox

The work explains that NIMS-OS is flexible. You can swap out the "musicians" in the orchestra:

• Different AI Brains: You can use different types of AI logic. Some are good at finding the absolute best peak (Bayesian Optimization), some are good at exploring strange, unknown areas (Boundless Exploration), and some are good at mapping entire landscapes (Phase Diagrams).
• Different Robot Hands: You can connect different robots. The authors tested it with a specific system called NAREE, a robot designed to mix fluids and automatically test battery liquids.

The Real-World Test: The Battery Hunt

To prove it works, the team used NIMS-OS to search for a better electrolyte (the fluid in a battery) for lithium-metal batteries.

• The Setup: They had 16 different chemical additives to choose from. They wanted to find the best mixture of 5 additives.
• The Process:
  1. First, the robot tried 32 mixtures randomly, just to get some starting data (like trying a few random cakes).
  2. Then the AI took over. It analyzed the results and told the robot exactly which 32 mixtures to try next.
  3. This happened automatically for 10 hours straight, without a human touching anything.
• The Result: By the 7th test round, the system found a recipe that performed significantly better than the others. It discovered that a mixture with specific chemicals (such as VC and FEC) worked best, which aligns with what human experts already knew, but the robot found it much faster and without fatigue.

Why This Matters

The work argues that the biggest breakthrough is not just the robot or the AI alone, but the software that connects them.

• Standardization: Previously, every lab had to write its own custom code to connect its specific robot with its specific AI. NIMS-OS offers a universal "plug-and-play" system.
• No Human Errors: Since the cycle is closed, the robot does not get tired and the AI does not get distracted. It simply keeps optimizing until the job is done.

In short, NIMS-OS is a universal remote control that allows a computer brain to speak with a robot body so they can discover new materials independently, around the clock.

Technical Summary

1. Problem Statement

The integration of Artificial Intelligence (AI) and robotic experiments is crucial for accelerating automated materials exploration. Although significant progress has been made in both fields—AI algorithms (e.g., Bayesian Optimization) and robotic hardware—a major bottleneck remains: the lack of a generic control software system.

• Current Limitation: Existing control systems are typically developed on a "case-by-case" basis, where specific AI algorithms are tightly coupled with specific robotic hardware. This makes exchanging algorithms or hardware difficult without rewriting the entire control logic.
• The Need: There is a need for a modular, standardized orchestration system capable of establishing a closed loop between diverse AI models and various robotic experimental setups without human intervention, thereby enabling true autonomous materials discovery.

2. Methodology: NIMS-OS Architecture

The authors developed NIMS-OS (NIMS Orchestration System), a Python-based library designed to function as generic middleware. It decouples AI algorithms from robotic systems through a modular architecture.

Core Workflow

The system operates in a closed loop consisting of three main steps:

1. Selection (AI): An AI module selects promising experimental conditions from a predefined "candidate file" (the search space).
2. Execution (Robot): The system generates input files for the robotic system and triggers the experiment.
3. Analysis and Update: The system analyzes the experimental output, extracts objective function values, and updates the candidate file to reflect new data for the next iteration.

Main Components

• Candidate File: A CSV file defining the material search space (e.g., compositions, process parameters). It contains columns for descriptors (input) and objective function values (output), which are initially empty and filled during the course of experiments.
• AI Modules (Step 1): NIMS-OS implements standard interfaces for various algorithms:
  • PHYSBO: Bayesian Optimization (BO) using Gaussian Process Regression and Thompson Sampling for Multi-Objective Optimization.
  • BLOX: Boundless Objective-Free Exploration, which uses Stein Discrepancy to find "curious" materials and ensure uniform sampling in the objective function space.
  • PDC: Phase Diagram Construction via Active Learning (Uncertainty Sampling) to map phase boundaries with minimal experiments.
  • RE: Random Exploration, used for generating initial data or baselines.
• Robotic Modules (Steps 2 & 3):
  • STAN: A reference (virtual) module that simulates the robotic workflow (creating input files, sending start signals, waiting for completion, and reading output files). This enables software testing without physical hardware.
  • NAREE: A specific module for the NIMS Automated Robotic Electrochemical Experiments system, comprising a liquid dispenser, an electrochemical measurement unit, and a robotic arm for high-throughput electrolyte screening.
• Visualization: Integrated tools (nimsos.visualization) for displaying the optimization progress, objective function distributions, and phase diagrams in real-time.

3. Main Contributions

• Generic Orchestration System: NIMS-OS is the first software providing a standardized interface that allows any AI algorithm to control any robotic system simply by exchanging modules. This fosters interoperability and reduces development time for new automated laboratories.
• Standardization: The authors have established standard input/output formats for AI algorithms and robotic systems. This enables testing new algorithms immediately on existing hardware and vice versa.
• Dual Interface: The system is available as a Python library for developers and as a GUI application for easy operation by non-programmers.
• Scalability: The modular design allows for the easy addition of new AI strategies or robotic hardware, facilitating the expansion of the ecosystem.

4. Experimental Results

To demonstrate the system's effectiveness, the authors conducted an autonomous exploration of multi-component electrolytes for lithium-metal batteries using the NAREE system.

• Setup:
  • Search Space: Combinations of 5 additives selected from a pool of 16 compounds (4,368 possible combinations).
  • Goal: Maximization of discharge time (proxy for battery capacity).
  • Hardware: 32 parallel experiments performed per microtiter plate.
• Process:
  • Cycle 1: Random Exploration (RE) was used to generate initial data.
  • Cycles 2–12: Bayesian Optimization (PHYSBO) was used to iteratively select better candidates.
  • Duration: The system ran 10 hours fully autonomously without human intervention, performing 384 experiments over 12 cycles.
• Result:
  • The optimal electrolyte composition was identified in the 7th cycle.
  • Best Result: A composition with 100 mM LiPF6, 100 mM LiTFSI, 2 vol.% PC, 2 vol.% FEC, and 2 vol.% VC achieved a discharge time of 1439.09 seconds (close to the theoretical maximum of 1800 s).
  • Validation: The top-10 results consistently contained vinylene carbonate (VC) and/or fluoroethylene carbonate (FEC), aligning with established domain knowledge regarding their positive effects on lithium-metal electrodes.

5. Significance and Outlook

• Accelerating Discovery: NIMS-OS demonstrates that a closed-loop system can significantly shorten the time to discover high-performance materials by automating the cycle of hypothesis generation and testing.
• Technical Standards: By establishing a common framework, NIMS-OS lowers the entry barriers for creating automated laboratories, potentially leading to cost reductions and increased adoption of robotics in materials science.
• Future Directions: The authors plan to expand the library with additional AI algorithms and robotic modules. They also emphasize the need for improved data management capabilities to handle the large volumes of data generated by autonomous systems, positioning NIMS-OS as a key instrument for digital transformation (DX) in materials science.

In summary, NIMS-OS represents a significant step toward the paradigm of the "self-driving laboratory" and provides the necessary software infrastructure to seamlessly integrate AI decision-making and robotic execution for automated materials discovery.