[Paper Review] Computationally accelerated experimental materials characterization -- drawing inspiration from high-throughput simulation workflows

Markus Stricker, Lars Banko+6 more|arXiv (Cornell University)|Dec 9, 2022

Machine Learning in Materials Science4 citations

TL;DR

This paper proposes integrating experimental materials characterization into the pyiron framework—originally designed for high-throughput simulations—by embedding Gaussian Process Regression (GPR) to autonomously guide measurements on composition spread materials libraries (CSMLs). The approach accelerates characterization by predicting unmeasured points with uncertainty, reducing experimental effort by ~10× while ensuring data provenance, reproducibility, and reusability through unified workflow management.

ABSTRACT

Computational materials science increasingly benefits from data management, automation, and algorithm-based decision-making for the simulation of material properties and behavior. Experimental materials science also changes rapidly by incorporation of `machine learning' in materials discovery campaigns. The obvious benefits which include automation, reproducibility, data provenance, and reusability of managed data, however, is not widely available in the experimental domain. We present an implementation of a Active Learning loop with a direct interface to an experimental measurement device in pyiron, a framework designed for high-throughput simulations, as demonstrator how to combine experimental and simulated data in one framework. Apart from the acceleration provided by the active learning approach, additional acceleration of the experimental characterization is achieved by using prior knowledge from density functional theory simulations as well as composition-property predictions from literature mining using correlations in word embeddings. With data from all domains in the same framework, a heretofore untapped and much-needed potential for the acceleration of materials characterization and materials discovery campaigns becomes available.

Motivation & Objective

• Address the lack of automation, data provenance, and reusability in experimental materials characterization workflows.
• Bridge the gap between high-throughput simulation and experimental characterization by unifying both within a single workflow framework.
• Enable active learning strategies—specifically Gaussian Process Regression—directly on experimental measurement devices to optimize sampling.
• Demonstrate that combining simulated and experimental data in one system unlocks new potential for accelerating materials discovery.
• Establish a foundation for autonomous, agent-based materials discovery that dynamically chooses between simulation, synthesis, or real-material characterization.

Proposed method

• Extended the pyiron framework with a new custom job type to manage experimental workflows, including input configuration for devices, samples, and measurement parameters.
• Implemented a Python-based dummy experimental device that emulates real measurement APIs, returning resistance values based on composition for testing.
• Integrated the GPy library to handle Gaussian Process Regression for active learning, enabling real-time model updates and uncertainty-aware prediction of unmeasured points.
• Designed a workflow where the GPR model selects the next optimal measurement point on a CSML to minimize uncertainty and maximize information gain.
• Ensured full data provenance and reproducibility by leveraging pyiron’s hierarchical data management and job tracking system.
• Enabled interoperability between simulation and experimental data by storing all results in a unified, accessible, and version-controlled database.

Experimental results

Research questions

• RQ1 Can active learning via Gaussian Process Regression be effectively applied to real experimental measurement devices in materials characterization?
• RQ2 To what extent can integrating experimental workflows into a simulation-focused framework like pyiron improve automation, reproducibility, and data reusability?
• RQ3 How much can the number of required measurements be reduced when using uncertainty-aware, model-guided sampling in CSML characterization?
• RQ4 What are the technical and infrastructural prerequisites for enabling autonomous, multi-agent discovery cycles combining simulation, synthesis, and real-material testing?
• RQ5 Can a unified software layer that manages both simulated and experimental data enable faster, more reliable materials discovery?

Key findings

• The integration of Gaussian Process Regression into experimental workflows via pyiron reduced the number of required measurements by approximately one order of magnitude.
• The system achieved accurate prediction of unmeasured resistance values on composition spread materials libraries with associated uncertainty estimates, enabling informed sampling decisions.
• All experimental and modeling steps were automatically documented, ensuring full reproducibility and data provenance through pyiron’s native workflow tracking.
• The framework demonstrated that experimental data can be managed with the same rigor as simulation data, enabling reuse and strategic initialization of models using prior knowledge.
• The prototype successfully demonstrated the feasibility of a unified platform for autonomous materials discovery that dynamically selects between simulation, synthesis, or experimental characterization.
• The approach paves the way for a federated system of intelligent agents that use both simulated and real experimental data to guide future measurements with minimal human intervention.