Access Official Technical PDF

Share Summary:LinkedInEmail

AI materials discovery can help engineers and researchers navigate the enormous range of possible material compositions and properties. For industrial organizations, the opportunity is not simply to predict a material characteristic. It is to connect credible data and computational models with experiments, processing conditions and the requirements of a manufactured product. A U.S. Department of Energy workshop on the topic identified several practical priorities for making that connection: improve data quality and access, integrate models across scales, strengthen public-private collaboration, and develop expertise across disciplines.

Why AI materials discovery depends on usable data

Machine learning (ML) models learn patterns from data. In materials research, those data may come from experiments, simulations, published studies or manufacturing processes. If datasets use inconsistent formats or omit information about how measurements were collected, comparing and combining them becomes difficult. A model trained on such information may be harder to interpret or apply to a new problem.

For industrial teams, this creates a foundational requirement: material-property values need context. Metadata about experimental conditions, equipment settings and measurement uncertainty helps users assess whether results are comparable and reproducible. Without it, a reported value may not provide enough information to repeat a test or judge whether it applies to a specific process.

The workshop participants recommended applying the FAIR principles: data should be findable, accessible, interoperable and reusable. In practice, that means more than storing files in a shared location. Data must be described consistently, searchable, and available in formats that analytical tools can use. Standardized interfaces and formats can also help teams work across different instruments and data sources.

Make research data more complete and reusable

Published results often emphasize successful experiments, while information about failed attempts may remain unavailable. That can leave other researchers unaware of routes that did not work, potentially leading them to repeat unproductive tests. Sharing negative results, where possible, can give a fuller account of the search process and help teams make better-informed decisions about subsequent experiments.

There are real barriers to sharing. Industrial data may be proprietary, and organizations need ways to collaborate without compromising intellectual property or data integrity. Workshop participants called for legal frameworks and collaboration mechanisms that make responsible data sharing easier. They also noted that sustainable governance and business models are needed to maintain accessible databases over time.

Legacy information presents another challenge. Useful results may exist in papers, tables, charts or figures that are difficult for software to analyze directly. Better data extraction and visualization tools could make these sources more usable, while clear reporting standards can improve the quality of newly generated data.

Use uncertainty to guide experiments

AI is not limited to analyzing a fixed dataset after experiments are complete. In adaptive experimental design, a model can help select what to test next, and the resulting measurements can be fed back into the model. This creates an iterative loop intended to focus experimental effort on informative or promising candidates.

The workshop described an example involving nickel-titanium alloys. Researchers began with a set of characterized samples, used an adaptive approach to identify additional compositions to investigate, and incorporated new results through successive feedback loops. The broader lesson is that an effective search strategy needs both exploration of areas with limited information and exploitation of promising candidates. A clearly defined target property is essential to deciding what the model should learn and which experiment is useful next.

Uncertainty is central to this process. Measurement and model uncertainty can help researchers judge the reliability of predictions, compare datasets and decide where additional testing is warranted. The workshop identified uncertainty quantification and validation of theory-based models as areas requiring further attention. AI outputs should therefore be treated as decision support, not as a substitute for checking whether predictions are scientifically credible and experimentally reproducible.

Connect models to processing and product requirements

A material’s composition alone does not determine whether it will work in an application. Processing conditions, part geometry, manufacturing constraints and required performance all matter. This is especially apparent in additive manufacturing, where material development needs to be considered alongside the process used to produce a part.

The workshop highlighted the need to integrate models across length and time scales and connect material design with processing and manufacturing. Such integration is challenging when products involve composites, compounds, interfaces or multiple materials. A useful workflow must represent not only an individual material, but also how materials behave as part of a larger system.

For engineering and operations teams, this points to a practical approach: define the application and performance requirements early, include manufacturing constraints in the design problem, and bring experimental evidence into the modeling cycle. AI may help prioritize candidates, but physical models and testing remain important for evaluating whether a proposed material and process are viable.

Build the partnerships and skills to apply AI

Materials informatics brings together materials science, chemistry, computer science, mathematics, experimentation and manufacturing knowledge. No single discipline supplies all the expertise required to build reliable datasets, develop useful models and interpret results in an industrial context.

Public-private partnerships can help connect company needs with university research and national laboratory capabilities, including computing resources and specialized experimental facilities. Effective collaboration also depends on involving business and operations stakeholders. The workshop discussion noted that close interaction between technical researchers and business units can support understanding and acceptance of ML-based tools.

Education is another part of the infrastructure. Materials specialists may need opportunities to learn ML concepts and tools, while data scientists need enough domain knowledge to work responsibly with physical measurements and models. Cross-disciplinary training can help teams identify appropriate problems, recognize limitations and put analytical results to practical use.

Practical implications for industrial teams

AI materials discovery offers a way to make materials research more data-driven and to guide experiments more deliberately. Its value depends on the quality and context of the data, transparent treatment of uncertainty, and the ability to connect predictions with physics, processing and product requirements. For manufacturers and system integrators, the central implication is to treat AI as part of an integrated R&D workflow—not as a standalone prediction engine. Clear objectives, reusable data, validated models and collaboration across technical and business functions provide a stronger basis for translating materials insights into manufacturing decisions.

Access Official Technical PDF

About The Author