Active learning path-dependent properties using a cloud-based materials acceleration platform

Dan Guevarra, Michael J. Statt, Kostiantyn Popovich, Brian A. Rohr, John M. Gregoire, Kevin Tran, Santosh K. Suram, Joel A. Haber, Willie NeiswangerDigital 2025 · DOI

Abstract: We demonstrate a framework for real-time active learning in state-changing experimental systems, integrating path-dependent acquisition functions with a fully autonomous, cloud-controlled experimental platform to accelerate materials discovery.

Introduction

Autonomous materials science is rapidly evolving, yet it often struggles with "path-dependent" properties—characteristics that change based on a material's history of exposure to different environments. In traditional active learning, an AI selects the next experiment based on previous results, assuming each trial is independent. However, in real-world scenarios like battery degradation or corrosion, applying a high voltage might provide information but also irreversibly dissolve the sample, precluding further measurements.

This paper introduces a cloud-based Materials Acceleration Platform (MAP) designed to navigate these experimental sequences where early decisions can alter or destroy a sample. By optimizing entire experimental trajectories rather than isolated points, the system balances the desire for information with the risk of sample failure.

Methods and Architecture

The researchers developed several key innovations to bridge the gap between AI planning and physical experimentation:

Iterated Expected Information Gain (iEIG)

The core algorithmic contribution is the iEIG acquisition function. Unlike standard functions that evaluate a single point, iEIG assesses an entire trajectory of potential observations. It calculates the total expected reduction in uncertainty over a sequence, weighted by the probability that the material remains stable at each step. This approach naturally favors "safer" starting points, as sequences starting with high-risk, high-stress steps yield lower iEIG scores due to the high probability of sample destruction.

Cloud-Based MAP and Hardware

The system utilizes a "stateful broker" design hosted on Amazon Web Services (AWS). This architecture allows the AI manager (the "brain") to remain in the cloud while controlling remote hardware (the "hands") via the HELAO (Hierarchical Experimental Laboratory Automation and Orchestration) framework. This decoupling enables distributed MAPs, where a single cloud-based AI could manage experiments across multiple global laboratories.

Schematic of the operando spectro-electrochemical scanning droplet cell used for real-time stability monitoring.

To measure stability in real-time, the team employed a custom scanning droplet cell (shown above). This setup performs operando spectro-electrochemistry, monitoring changes in light absorbance as electrochemical potential is varied to detect material degradation immediately.

Experimental Campaign and Results

The platform was tested on the (Co–Ni–Sb)O$_z$ metal oxide system. The autonomous system investigated eight material libraries across two pH levels over 11 days of continuous operation, requiring human intervention only for physical sample plate swaps.

  • Data Generation: The campaign produced 9,755 chronoamperometry measurements across 887 unique composition-pH combinations.
  • Model Performance: The Mean Absolute Error (MAE) of the model's predictions decreased consistently. The system successfully adapted as the design space expanded from binary to complex ternary oxides.
  • Stability Mapping: The campaign identified "optically stable potential windows," creating a comprehensive map of where these materials can operate without degradation. Notably, experimental results often diverged from theoretical Pourbaix energetics, highlighting the necessity of autonomous experimental platforms for capturing complex, kinetic-driven behaviors.

Conclusion and Future Outlook

This work demonstrates the feasibility of using AI "science managers" to plan multi-step sequences in path-dependent systems. While currently focused on electrochemistry, the iEIG framework is theoretically applicable to any process where history matters, such as additive manufacturing or complex chemical synthesis.

Future improvements may include distinguishing between reversible and irreversible material changes and incorporating multi-fidelity data (combining simulations with experiments) to further accelerate the learning process.

Summary updated

Abstract

We demonstrate a framework for real-time active learning in state-changing experimental systems, integrating path-dependent acquisition functions with a fully autonomous, cloud-controlled experimental platform to accelerate materials discovery.
BibTeX
@inproceedings{guevarra2025active,
  author = {Dan Guevarra and Michael J. Statt and Kostiantyn Popovich and Brian A. Rohr and John M. Gregoire and Kevin Tran and Santosh K. Suram and Joel A. Haber and Willie Neiswanger},
  title = {{Active learning path-dependent properties using a cloud-based materials acceleration platform}},
  booktitle = {Digital},
  year = {2025},
  doi = {10.1039/d5dd00325c},
}
Figure