NIMO: A Software Platform for Closed-Loop Materials Exploration with Diverse AI Algorithms

This paper introduces NIMO, an open-source software platform that bridges diverse AI algorithms and heterogeneous robotic hardware through a modular CSV-based architecture and unified Python interface, enabling seamless closed-loop materials discovery across various experimental domains while also providing a no-code tool for non-programmers.

The Gist

Imagine you are trying to find the perfect recipe for a cake, but instead of baking one cake at a time, you have a team of robots that can bake hundreds simultaneously. The problem is, you have a million possible ingredient combinations, and you don't know which ones will actually work or if they will even fit in your oven.

This is the challenge scientists face when discovering new materials. Enter NIMO, a software platform described in this paper as the "universal translator" and "smart manager" for these high-tech, self-driving laboratories.

Here is how NIMO works, broken down into simple concepts:

1. The "Post-It Note" System (The Core Innovation)

Usually, connecting a smart AI brain to a clunky robot arm is like trying to plug a modern USB-C cable into a 1990s VCR. They speak different languages, and the wiring is a nightmare.

NIMO solves this with a surprisingly simple trick: CSV files (which are just plain text spreadsheets, like an Excel sheet).

• How it works: The AI writes its next suggestion on a digital "Post-It note" (a file called proposals.csv). The robot reads that note, does the experiment, and writes the results back on a different note (updating candidates.csv).
• Why it's great: It doesn't matter if the robot speaks Python, Visual Basic, or even if a human is reading the notes. As long as the robot can read and write a simple spreadsheet, it can talk to NIMO. It's like a universal adapter that lets any device plug into the AI.

2. The "Menu" Instead of a "Blank Canvas"

In many AI systems, the AI is given a blank canvas and told, "Draw anything between 0 and 100." Sometimes, the AI draws a picture that is physically impossible (like asking for 150% of an ingredient).

NIMO uses a Candidate Pool approach.

• The Analogy: Think of a restaurant menu. The chef (the user) writes down every single dish that is actually possible to make with the current ingredients and equipment. This list is the candidates.csv.
• The AI's Job: The AI doesn't invent new dishes; it simply scans the menu and picks the next best dish to order. Because the AI is only choosing from the pre-approved menu, it can never suggest an impossible experiment. This prevents the robots from wasting time trying to do things that can't be done.

3. The "Swiss Army Knife" of Algorithms

NIMO comes pre-loaded with 12 different AI strategies (algorithms), each designed for a different goal. You don't need to be a coder to switch between them; you just pick the tool you need.

• The "Hunt for the Best" (PHYSBO): If you want the single best material (like the strongest metal), this tool hunts for the peak.
• The "Safe Zone" (PTR): If you need a material that falls within a specific range (like a battery that lasts between 10 and 12 hours, not just the longest possible), this tool finds the safe zone.
• The "Explorer" (BLOX): If you want to map out a whole new territory and find weird, unique materials, this tool spreads out to cover as much ground as possible.
• The "Text Reader" (LLMEP): If your goal is described in words rather than numbers (like "find a crystal that looks like this"), this tool uses advanced language models to understand the description.

4. Real-World "Self-Driving Labs"

The paper shows NIMO working in six different "labs," proving it can handle anything from chemistry to biology:

• NAREE: Robots mixing battery liquids to find the perfect electrolyte.
• CHEMSPEED: Automating chemical reactions to make new medicines faster.
• COMBAT: Creating thin films for next-generation electronics.
• ROPES: Optimizing the manufacturing process for fuel cells.
• Coffee Ring SDL: Studying how droplets dry (the "coffee ring" effect) to understand fluid physics.
• BioDot: Connecting to an old, legacy machine that uses outdated software, proving NIMO can upgrade old labs without throwing away expensive equipment.

5. The "No-Code" Dashboard (NIMO Desktop)

Not every scientist knows how to write computer code. To fix this, the creators built NIMO Desktop, a simple app with buttons and menus.

• How it works: A scientist can click "Start," pick an algorithm from a list, load their data, and watch the loop run. They don't need to type a single line of code. It's like driving a car with an automatic transmission instead of having to manually shift gears.

Summary

NIMO is a bridge. It connects the "brain" (diverse AI algorithms) to the "hands" (robots and lab equipment) using a simple, universal language (spreadsheets). It ensures the AI only asks for things that are possible, offers a toolbox for every type of discovery goal, and makes autonomous science accessible to everyone, not just computer programmers.

Technical Summary

Technical Summary: NIMO – A Software Platform for Closed-Loop Materials Exploration

Problem Statement

The discovery of advanced materials is currently hindered by the slow pace of conventional trial-and-error experimentation, which struggles to meet societal and industrial demands. While Self-Driving Laboratories (SDLs)—systems where AI proposes experiments and robotics execute them—offer a solution, a critical bottleneck remains: the difficulty of seamlessly integrating diverse AI algorithms with heterogeneous robotic hardware across different laboratories. Existing orchestration software often focuses on workflow management but relies on external, specialized packages for AI, creating integration friction. Furthermore, many AI algorithms are tailored to specific goals (e.g., finding a single optimum vs. mapping a diverse property space), requiring distinct implementations that are not easily interchangeable within a single system.

Methodology

The authors present NIMO (NIMS Orchestration System), an open-source Python platform designed to decouple AI decision-making from experimental execution through three core architectural paradigms:

1. Modular AI–Robot Decoupling via CSV Exchange:
   Instead of relying on complex in-memory API calls or language-specific bindings, NIMO communicates with robotic systems exclusively through standard CSV files. The workflow follows a three-step iterative cycle:

   • Selection: NIMO reads a candidates.csv file (the search space), selects promising conditions using an AI algorithm, and writes them to a proposals.csv file.
   • Evaluation: A robotic system or human operator reads the proposals, executes experiments, and records results.
   • Update: The results are written back to the candidates.csv file, closing the loop.
     This file-based interface ensures NIMO is agnostic to the underlying hardware control language (Python, Visual Basic, LabVIEW, etc.) and allows NIMO to function as a universal decision engine for external orchestration frameworks (e.g., IvoryOS).

2. Discrete Candidate-Pool Architecture:
   Unlike traditional Bayesian Optimization (BO) methods that define continuous search spaces and require post-hoc feasibility checks, NIMO operates on a pre-enumerated discrete candidate pool. Users define all physically realizable experimental conditions in a single candidates.csv file.

   • Advantages: This approach inherently prevents the AI from proposing infeasible experiments. It allows for the direct embedding of domain knowledge (e.g., constraints where parameters must sum to 100% or specific resolution requirements) without complex algorithmic constraints. It also natively supports mixed-variable types (continuous, discrete, categorical) without requiring specialized encoding.

3. Unified Algorithm Interface:
   NIMO embeds a library of twelve distinct AI algorithms directly within its framework, accessible via a single nimo.selection function. The specific backend is determined by a string argument (e.g., method="PHYSBO"), while input/output arguments remain universal.

Key Contributions

1. The NIMO Algorithm Portfolio

NIMO integrates twelve algorithms categorized by their exploration goals, eliminating the need for users to source external packages:

• Optimization:
  • PHYSBO: Discrete-space Bayesian Optimization (single/multi-objective) using Gaussian processes.
  • BOMP: Bayesian Optimization with Material Process, enforcing shared process parameters within batches.
  • NTS: Nested Thompson Sampling for threshold-based selection with diversity preservation.
  • COMBI: Composition-spread exploration mimicking combinatorial sputtering physics.
• Target-Range & Objective-Free:
  • PTR: Probability of Target Range for steering properties into specific windows.
  • BLOX: Boundless Objective-free eXploration for mapping diverse property spaces using Stein discrepancy.
• Categorical & Ranking:
  • PDC: Phase Diagram Construction using semi-supervised classification for categorical labels.
  • RSVM: Ranking Support Vector Machine for ordinal optimization and transfer learning.
• Initial Sampling (No Training Data):
  • RE: Random Exploration.
  • DOE: Design of Experiments (Greedy, Distance, D-optimal, LHS).
  • ES: Exhaustive Search.
• LLM-Based:
  • LLMEP: LLM-based Experimental Planner for natural-language-guided selection and text-valued objectives.

2. Six Validated SDL Implementations

The authors demonstrated NIMO's versatility across six distinct domains, all utilizing the same core logic and file formats:

• NAREE: Autonomous electrolyte discovery for lithium-metal anodes (4,368 combinations) using a microplate-based electrochemical platform.
• CHEMSPEED: Optimization of Suzuki-Miyaura coupling reactions via a Chemspeed synthesis platform coupled with SFC.
• COMBAT: Autonomous thin-film exploration of quinary alloys for spintronics using combinatorial sputtering and the COMBI algorithm.
• ROPES: Process informatics for fuel-cell catalyst layer manufacturing, optimizing coating and drying parameters.
• Coffee Ring SDL: Mapping the phase boundary of the coffee-ring effect; notably demonstrated seamless interoperability with the IvoryOS orchestration framework.
• BioDot: Integration with a legacy Visual Basic-controlled liquid dispenser, proving NIMO's ability to interface with non-Python, legacy hardware without modification.

3. NIMO Desktop

To democratize access for non-programmers, the authors introduced NIMO Desktop, a cross-platform (Windows/macOS) GUI. It replicates the closed-loop workflow via two tabs: "Selection" (choosing algorithms and generating proposals) and "Update" (logging experimental results). It shares the exact data structures with the Python API, ensuring compatibility.

Results

The paper validates NIMO through successful deployment in the six SDLs mentioned above. Key outcomes include:

• Interoperability: NIMO successfully controlled diverse hardware ranging from modern Python-based robots to legacy Visual Basic systems and web-based orchestration frameworks (IvoryOS).
• Efficiency: In the NAREE electrolyte discovery, NIMO efficiently navigated a massive combinatorial space to identify optimal formulations. In the CHEMSPEED synthesis, it rapidly converged on optimal reaction conditions for Suzuki-Miyaura coupling.
• Flexibility: The system successfully handled diverse objectives, from maximizing a single metric (discharge time, yield) to mapping phase diagrams and optimizing multi-parameter manufacturing processes.
• Accessibility: The NIMO Desktop application enabled human-in-the-loop exploration without requiring coding skills, while the file-based interface allowed integration with existing laboratory infrastructure.

Significance and Future Outlook

The paper posits that NIMO addresses the "integration bottleneck" in autonomous science by providing a plug-and-play foundation that separates AI logic from hardware control. Its significance lies in:

• Democratization: Making advanced AI-driven exploration accessible to non-programmers via the Desktop app and to legacy labs via the CSV interface.
• Modularity: Allowing a single system to be repurposed for different materials exploration tasks simply by switching the algorithm argument, rather than rewriting integration code.
• Scalability: The candidate-pool design facilitates the vision of globally connected SDLs, where different laboratories can collaborate on a shared candidate pool using their specific equipment and expertise, communicating only through standard data files.

The authors conclude that the future of NIMO involves expanding the algorithm portfolio (e.g., multi-fidelity BO, LLM-augmented optimizers), enhancing user interfaces (including Model Context Protocol support), and fostering a network of interconnected autonomous laboratories. They emphasize that advancing autonomous discovery requires "AI for Science"—algorithms that natively embody the practical constraints and workflows of materials exploration.