What Is AI for Science?
How AI Is Changing Scientific Research

AI for Science uses AI to advance scientific research

AI for Science is an umbrella term for efforts to use artificial intelligence in scientific research. To understand how nature works and explore new materials or technologies, researchers work with large amounts of data and complex calculations. AI can help them find patterns in that data and narrow down what to investigate next.

Generative AI that produces conversations or text can contribute to this work. It can help organize research papers, write programs for calculations, or explain ideas. However, plausible writing is not necessarily scientifically correct. Science deals with more than words: it also involves measurements, images, molecular shapes, and mathematical equations. The key is to combine AI suited to the task with experimental equipment and appropriate computational methods.

Connecting observation, prediction, and experiment

Consider a question such as, "Why does a particular material conduct heat well?" Researchers first take measurements and record the temperature and the material's composition. When analyzing these observations, AI can identify common trends and highlight differences worth investigating. It can also learn quirks of the measuring equipment or random variation as if they were meaningful features.

The next step is to build a predictive model from the data. A model describes the relationship between conditions and outcomes. It can estimate how well a material with a particular composition will conduct heat, helping researchers compare candidates for experiments. Researchers also use AI to approximate parts of a simulation—a computer calculation that represents a natural phenomenon—so they can investigate many conditions. Even when calculations become faster, they still need to check the approximation's error and the conditions under which it can be used.

When planning experiments, it matters to choose both promising candidates and conditions that are not yet well understood. Researchers compare the resulting measurements with predictions and revise the model or hypothesis if they find differences. A hypothesis is a proposed explanation of a phenomenon that has yet to be tested.

Ask a question → Observe → Analyze and predict → Test through experiments → Update the hypothesis.
This repeated cycle is central to research. AI can support the connections between its stages.

Results that do not match a prediction are valuable too. They offer clues about whether the material's conditions differed, whether there was a measurement problem, or whether the explanation itself was incomplete. Collecting only successful results can distort our view of a phenomenon. Keeping unexpected results and using them to inform the next question is essential.

AlphaFold and AI that helps choose the next experiment

One prominent example is AlphaFold, which predicts the three-dimensional structures of proteins. Proteins are molecules that perform many roles inside cells, and their shapes offer clues to how they work. The 2021 paper on AlphaFold 2 reported highly accurate structure predictions for many targets, using information including the sequence of building blocks called amino acids[1].

Structure predictions can provide clues for understanding function and developing drugs. They do not establish how molecules behave inside living organisms, or whether a drug will be effective or safe. The later AlphaFold 3 expanded its scope to complexes involving proteins and small molecules, among other components. It still has limitations, including that it does not directly reproduce the dynamic behavior of molecules[2]. Researchers need to check the confidence of predictions and compare them with experimental results as appropriate for their purpose.

In another example, researchers combined a search algorithm with a mobile robot to find materials that help a reaction produce hydrogen. The system used experimental results to choose the next conditions to test[3]. This was an example of automating a repeated process within the limits of a specified setup and task. It does not mean that any kind of research can proceed without people.

Respecting physical laws and asking how certain a result is

Natural phenomena follow physical laws, such as the conservation of energy. A prediction that merely fits the data can still contradict those laws. This is why researchers are developing methods that incorporate physical laws into learning and prediction, an approach known as physics-informed learning[4]. Even when those laws are incorporated, computational errors and incomplete assumptions do not disappear.

It is important to describe uncertainty: how much a result could vary and what remains unknown. Measurement variation, a shortage of training data, and application to unfamiliar conditions create different kinds of uncertainty. Researchers need to test predictions with data that was not used for training or with independent measurements, and investigate the range of conditions in which the predictions hold. Polished numbers and figures are not evidence by themselves.

Reproducibility is essential too. Researchers record where the data came from, measurement conditions, preprocessing, programs, model versions, and computational settings so that others can check the procedure. The FAIR principles provide a foundation for making scientific data easier to find, use, and reuse under appropriate conditions[5]. When some information cannot be made public, the permitted uses and methods of verification still need to be clear.

People remain responsible for questions and verification

As the use of AI grows, the role of researchers becomes more important. People decide what they want to know, which experiments can test it, and what the results allow them to conclude. They consider safety and ethics, distinguish predictions from experimentally established facts, and take responsibility for their conclusions. Accepting an AI suggestion—or rejecting it—is part of research.

The research introduced on this site also involves analyzing radiation measurement data, checking consistency with physics, and developing systems that support scientific computing. These are research directions; they do not imply a completed capability that works under every condition.

One connection to society is the idea of Science as a Service (ScaaS): an effort to apply scientific knowledge, measurement, and data analysis to challenges in manufacturing and society. For example, researchers might measure variation on site, consider possible causes, and test whether a proposed improvement works. This, too, rests on the cycle between questions, data, and verification.

The potential of AI for Science lies in helping people explore more possibilities and deepen their understanding through sound evidence, rather than revealing the truth in one step. When AI meets science, it changes both the answers and the way we search for them while continually checking what we find.