Access Official Technical PDF

Share Summary:LinkedInEmail

Materials informatics applies data-driven methods to the search for materials with desired properties. Its relevance to industrial automation lies in how it connects computation, experiments, measurement systems and decision-making: rather than treating each test as an isolated task, a connected workflow can use results to inform what should be tested next. Research presented at the 2019 JST International Symposium on Materials Informatics illustrates both the promise of this approach and the engineering challenges involved.

Why materials discovery is a complex search problem

Finding a material for a particular application can require exploring many candidate compositions alongside processing conditions. The number of possibilities grows rapidly as variables are added, making exhaustive experimental testing impractical. The challenge is not simply collecting more data; it is choosing useful experiments, interpreting their results and updating the search based on what has been learned.

Symposium presentations addressed this challenge across areas including catalysts, polymers, semiconductors, battery electrolytes and solar-cell materials. The work ranged from machine-learning models that estimate material properties to computational methods for exploring atomic structures. Together, these examples show how data and models can help narrow a search, while experiments remain important for checking whether predicted candidates behave as expected.

Materials informatics and the automation feedback loop

A central idea is to link prediction and experimentation in a repeated cycle. A model uses available data to propose candidates or estimate properties. Experiments or simulations then evaluate selected candidates, and the resulting data can update the model and shape the next round of investigation. When this cycle is automated, instruments can become part of a broader system for planning, executing and analyzing experiments.

This is more than automating instrument movement or sample handling. The useful connection is between the equipment, the data it produces and the logic used to decide what happens next. In an autonomous materials workflow, a system may combine:

  • Experimental or computational data describing candidate materials and measured or calculated properties.
  • Models that predict outcomes, rank candidates or guide the next search step.
  • Automated equipment that carries out selected experiments or measurements.
  • Analysis and feedback that incorporates results into subsequent decisions.

The symposium included work on AI-driven high-throughput materials science and on combining high-throughput experimentation with machine learning. These approaches point to a practical constraint: predictions can be generated faster than experiments can necessarily be performed. Connecting models to automated testing can help address that mismatch, but the experimental workflow and data pipeline must be designed to work together.

Models need useful data—and careful interpretation

Machine learning depends on the information available to train and evaluate it. Materials datasets may be limited, inconsistent or expensive to extend. One presentation on polymer design, for example, described using transfer learning: a model trained on a larger body of data was adapted to a smaller dataset for a target property. Other work used literature-derived measurements and digitized chemical structures to screen candidates. These examples underscore the importance of data preparation and of understanding what a model can reasonably infer from its inputs.

Model choice also affects how results can be used. A presentation on explainable machine learning highlighted the value of models that can provide interpretable information to materials scientists, alongside prediction accuracy. In industrial settings, this distinction matters when technical teams need to assess why a candidate was prioritized, whether the data support the recommendation and how to plan a validation test.

Models are not substitutes for evidence. A machine-learning prediction reflects the data and assumptions used to build it, and a model may perform poorly when applied beyond that context. Research on heterogeneous catalysts explicitly noted the need to balance exploration—collecting new knowledge—with exploitation—using existing knowledge to improve outcomes. That balance is also a useful consideration when deciding how an automated test program allocates experiments.

From research concepts to industrial workflows

For automation engineers and system integrators, the most transferable lesson is architectural: effective data-driven discovery requires a coordinated workflow, not an isolated algorithm. Teams considering this approach can start by defining how experimental results will be recorded, linked to material and process variables, analyzed and made available to later decisions. They can then identify where high-throughput testing or automation can support repeated, consistent measurements.

Several research examples show how different methods can serve different parts of that workflow:

  • Virtual screening and Bayesian optimization can help prioritize candidate structures or experiments when the search space is large.
  • High-throughput experimentation can generate batches of comparable measurements for model development and evaluation.
  • Physics-based computation and machine learning can complement one another in estimating properties and exploring material structures.
  • Explainable models can help specialists inspect the factors behind predictions and extract actionable information.

These are research approaches, not turnkey guarantees of faster development. Their practical value depends on the quality and relevance of the data, the suitability of the model and the ability of the experimental system to test model-generated proposals.

Building a disciplined path to smarter discovery

Materials informatics offers a way to organize complex discovery work around data, models and iterative testing. For industrial teams, the opportunity is to connect laboratory automation and measurement with a disciplined decision loop: capture dependable data, use models to guide the next investigation, and validate recommendations through appropriate experiments. The symposium’s examples show why this requires collaboration among materials specialists, data scientists and automation experts.

The practical implication is clear: automation can do more than execute predefined procedures when it is integrated with analysis and feedback. But useful autonomy depends on well-managed data, transparent decision processes and validation. Those foundations can help organizations make materials research workflows more systematic and better equipped to explore complex possibilities.

Access Official Technical PDF

About The Author