Materials Informatics is moving toward a stage in which artificial intelligence is used not only to analyze data or suggest new compositions, but also to determine which questions should be asked and to design the research cycle itself. Published on September 9, 2026, the article discusses this shift through the role of large language models (LLMs) and intelligent agents in the materials development cycle.
The central idea is not that artificial intelligence will replace the researcher, but that the way roles are distributed between the human and the system will change. Instead of conducting one experiment and then analyzing its result, it is possible to build an iterative loop in which the algorithm selects the next experiment based on previous data, while keeping verification and intervention decisions within the process design.
From Optimizing Results to Selecting the Next Experiment
The article explains that the use of Bayesian optimization in materials engineering now makes it possible to select subsequent experiments sequentially. The method relies on previous results and on uncertainty in the model, then selects the point expected to be most useful to the research. However, this approach alone does not guarantee a complete interpretation of the results or the discovery of all influential factors; its primary aim is to reach the best value within a limited number of experiments.
The discussion indicates that the success of Bayesian optimization depends on how variables and constraints are defined. If some conditions are not included or are considered outside the scope of the research, the range of artificial intelligence recommendations may change without this being clear to the user. Defining the objective itself may also be a research problem, particularly when the relationship between a material's composition and its performance is not sufficiently known.
What Do Large Language Models Add?
Large language models can contribute to question design by reading previous records, organizing information, extracting patterns, or formulating initial hypotheses. However, the article cautions against equating this capability with the automatic production of reliable knowledge. The inference that models construct from the available information may differ from the outcome the researcher wants to reach, and their outputs may reflect the limitations of the data and the way the question is formulated.
For this reason, simply inserting a language model into the workflow is not enough. What is required is a mechanism that connects the question to the data, the experiment, and the verification criterion, then returns the result to the research loop to update the question or select the next experiment. In this form, the model becomes part of a reviewable discovery process rather than a sole source of scientific judgment.
The Human Inside or Above the Loop
The article discusses the concept of human-on-the-loop as a model in which the human monitors the system and has the authority to intervene when necessary, rather than participating manually in every step as in the human-in-the-loop model. This idea is connected to what the article calls, in the context of software development, “loop engineering”: determining where artificial intelligence fits within the cycle, what the human monitors, and when the process is repeated or stopped.
This view expands the scope of design from the model's capabilities to the design of the entire system. The issue is not only whether artificial intelligence can propose an answer, but also who reviews it, how the quality of the result is measured, what action follows it, and when the decision returns to the human. The article states that discussion of the human role in automation intensified during 2025, while calls to consider loop engineering when developing artificial intelligence agents emerged in 2026.
Why Does This News Matter?
The practical impact of this proposition is that materials engineering projects will not be measured only by model accuracy or by the number of experiments they can conduct. They will also require a clear definition of objectives, the scope of variables, exclusion conditions, and points for human review. As the system's ability to move from one result to a new experiment increases, the traceability and verifiability of decisions become more important.
However, the article does not provide evidence that agents or language models have resolved these constraints. Rather, it frames them as open design questions: How is a successful outcome defined? How is the result connected to the subsequent behavior? And when is human intervention necessary? Therefore, the fundamental value of the proposed shift does not lie in fully automating research, but in redesigning the research cycle so that the use of artificial intelligence can be examined and corrected.