Would AI agents for materials discovery need industry standards before deployment?

Why do AI agents need standards? Because their predictions are only as good as the data they learn from.

AI agents for materials discovery are trained on data from experiments and simulations. But that data is often messy: it's scattered across different labs, recorded in different formats, and sometimes not published at all [2]. One review notes that materials research faces a 'near-infinite chemical design space' and that conventional experimental data remain 'fragmented, dark, or inconsistent due to heterogeneous laboratory standards' [2]. This means an AI trained on one lab's data might not work for another lab's materials, because the underlying data isn't comparable.

A concrete example comes from a 2026 study that built an AI workflow to extract data from scientific papers. They found that directly using commercial multimodal models to read figures and tables was inaccurate, so they developed a custom workflow (DIVE) that improved extraction accuracy by 10–15% over commercial models and over 30% over open-source models [1]. This shows that even state-of-the-art AI struggles with unstructured data—and that without standardized data formats, AI agents will keep hitting this bottleneck.

What kind of standards are needed? Data formats, experimental protocols, and validation benchmarks.

The papers point to three areas where standards are critical. First, data formats: one review calls for 'standardized, multi-source data infrastructures that facilitate the seamless integration of heterogeneous datasets' [2]. This means agreeing on how to record and share data so that AI can actually use it. Second, experimental protocols: because labs use different methods, results can't be compared unless protocols are standardized [2]. Third, validation: AI predictions need to be checked against real experiments, and that requires consistent benchmarks.

The push for autonomous labs—where AI proposes candidates, robots run experiments, and the results feed back into the model—makes standards even more urgent. One review describes how self-driving labs integrate machine learning, automation, and robotics to run experiments selected by AI [4]. For these systems to work reliably across different labs, they need common data formats and interfaces. Another paper notes that combinatorial synthesis (automated, parallel synthesis) is a natural fit for AI-driven discovery, but its success depends on the quality and comparability of the data produced [5].

Are standards always necessary? Not for every use case—but they become critical as AI agents scale up.

For a single lab using AI to screen a specific class of materials, internal consistency might be enough. For example, a 2025 study used an AI search engine to discover 50 new altermagnetic materials, and the predictions were confirmed by first-principles calculations [3]. That worked because the AI was trained on a specific dataset and validated within that context. But the same approach might not transfer to other materials or labs without standardized data.

The bigger risk is when AI agents are used to make decisions that affect real-world products. One review emphasizes that challenges remain in 'model generalizability, standardized data formats, experimental validation, and energy efficiency' [6]. It also calls for open-access datasets that include negative experiments (failed attempts) and ethical frameworks to ensure responsible deployment [6]. So while standards might not be needed for every exploratory project, they are essential for building trust and ensuring that AI-driven discoveries can be reliably reproduced and scaled.

About These Sources

This answer is built on 6 peer-reviewed studies — published from 2023 to 2026, 4 from 2024 or later, 4 in Q1 journals, collectively cited 584 times — selected as the most relevant from 7 studies that passed quality screening, drawn from 66 papers retrieved from a database of over 500 million.

Sources used in this answer

Related research questions