A scientist walks into the lab with a good idea.
By lunchtime, the idea is wrong.
That may sound like failure. In science, it is often progress.
The experiment did not behave as expected. A measurement came back outside the predicted range. A material failed under conditions where the model said it should remain stable.
Now the scientist has something more valuable than confirmation:
a reason to think again.
This cycle—propose, test, learn, revise—is one of the deepest structures in scientific work. It is also the structure that AI for Science must eventually learn to inhabit.
Prefer to watch? Continue with the video below, or keep reading.
Science Is Not a Straight Line
Textbooks often make scientific discovery look orderly.
Ask a question.
Form a hypothesis.
Run an experiment.
Reach a conclusion.
Real research is rarely so tidy.
A more realistic picture is a loop:
\[ \text{question} \rightarrow \text{hypothesis} \rightarrow \text{prediction} \rightarrow \text{experiment} \rightarrow \text{data} \rightarrow \text{evaluation} \rightarrow \text{revised hypothesis}. \]
Then the process starts again.
Sometimes the loop takes minutes.
Sometimes it takes years.
Sometimes the result sends the project back several steps.
And sometimes the most interesting discovery is not the answer to the original question at all, but an unexpected observation that changes the question itself.
If AI is going to participate meaningfully in science, it must do more than provide good answers at isolated points.
It must learn to operate inside this loop.
Step 1: The Question
Every research project begins with some form of uncertainty.
Why does this material become superconducting?
Which molecule binds most strongly to this target?
What causes this cell to change state?
Which control pulse best prepares this quantum state?
These questions are not interchangeable.
A good scientific question must be important enough to matter, specific enough to investigate, and structured enough that evidence can change our belief about the answer.
This is already a difficult task for AI.
A model can generate hundreds of questions.
But which one is worth spending six months of laboratory time on?
Scientific judgment begins before the first calculation.
Step 2: The Hypothesis
A hypothesis is more than a guess.
It connects a possible explanation to consequences that could be observed.
Suppose a researcher believes that a particular atomic arrangement makes a catalyst more active.
That idea becomes scientifically useful only when it leads to predictions:
If the arrangement is present, reaction rate should increase.
If a particular site is blocked, activity should decrease.
If temperature changes, the effect should follow a certain pattern.
An AI system can help here by combining literature, previous experiments, mechanistic models, and statistical evidence.
Modern models are already capable of generating plausible hypotheses.
The harder question is whether they can rank them well—and whether they can recognize when none of their existing ideas is adequate.
Step 3: The Prediction
A hypothesis becomes testable through prediction.
This is where scientific models become especially important.
The model might be:
- an equation,
- a numerical simulation,
- a molecular model,
- a neural network,
- a probabilistic model,
- a quantum simulation.
The prediction creates a bridge between an abstract explanation and a future observation.
For AI4S, this step is often relatively comfortable.
AI systems are already strong at computation, approximation, and pattern prediction.
But a prediction by itself does not close the loop.
Nature still gets a vote.
Step 4: The Experiment
The experiment is where the model leaves its information world.
A proposed molecule must be synthesized.
A device must be fabricated.
A sample must be measured.
A physical system must be perturbed.
This is the point at which AI4S encounters one of its hardest limitations.
Digital actions are cheap.
Physical actions are not.
A program can be run a thousand times overnight.
A thousand wet-lab experiments may require robots, materials, preparation, calibration, safety procedures, and days or weeks of work.
For AI to move from scientific reasoning to scientific action, it needs an interface to the physical world.
That may be a robotic laboratory.
It may be an automated microscope.
It may be a telescope scheduler.
It may be a quantum-control system.
But without that connection, the loop remains incomplete.
Step 5: The Data
When the experiment finishes, the system receives evidence.
Often the evidence is messy.
Measurements have noise.
Sensors drift.
Samples vary.
Experiments fail.
Metadata are incomplete.
And sometimes the most important feature is not in the variable that the researcher originally planned to analyze.
For an AI system, the challenge is therefore not merely to “read the result.”
It must understand what the result means in context.
Was the experiment executed correctly?
Is this an outlier?
Does the measurement contradict the hypothesis, or did the instrument malfunction?
Should the result be trusted enough to change the next experiment?
The scientific loop depends on these judgments.
This is why modern science increasingly emphasizes the FAIR data principles: data and metadata should be Findable, Accessible, Interoperable, and Reusable. A closed-loop agent cannot learn effectively from historical experiments if the records are missing context, stored in incompatible formats, or inaccessible to machines. Data infrastructure is therefore part of the scientific loop, not an administrative afterthought.
Step 6: Evaluation
This is the moment where many AI systems become interesting.
Suppose the prediction was
\[ y_{\mathrm{pred}} = 0.82 \]
and the experiment returned
\[ y_{\mathrm{obs}} = 0.37. \]
A simple optimizer sees an error.
A scientist asks why.
Was the model wrong?
Was an assumption violated?
Is there an unmeasured variable?
Did a new mechanism appear?
Was the result produced by noise?
Evaluation is not merely assigning a score. It is interpreting disagreement.
That difference becomes crucial when AI moves from optimization toward discovery.
Step 7: What Should We Do Next?
Now the loop becomes active.
The system has to choose the next action.
Repeat the same experiment?
Change one variable?
Try the best remaining candidate?
Explore a surprising region?
Collect a different kind of measurement?
Abandon the hypothesis?
This is where scientific reasoning meets decision-making.
The next experiment should not simply be the one most likely to succeed.
Sometimes the most valuable experiment is the one most likely to teach us something.
That distinction will become the focus of a later blog on exploration versus exploitation.
In pharmaceutical and materials research, a closely related closed-loop workflow is often summarized as DMTA: Design, Make, Test, Analyze. AI is especially strong at Design and Analyze because those stages are information-rich and computational. Make and Test require contact with physical reality. That asymmetry is one reason closed-loop discovery is so powerful—and why the physical laboratory remains such an important bottleneck.
Closing the Loop Changes the Role of AI
Consider two AI systems.
System A receives a dataset from a scientist and predicts which material will perform best.
System B does something more.
It makes a prediction.
It chooses a material to synthesize.
A robot makes the sample.
An instrument measures it.
The AI compares the measurement with the prediction.
Then it chooses the next sample.
Both use machine learning.
But System B is participating in the research process in a fundamentally different way.
It is not only analyzing the past.
It is helping create the next piece of evidence.
That is the importance of the closed loop.
This is often described as closed-loop discovery: the output of one experiment becomes the input to the next decision, so the system updates not only a parameter setting but also its working model of the world. That is the scientific form of recursive improvement that matters most for AI4S.
The Loop Can Exist at Many Speeds
Not every scientific loop needs a robot arm.
In mathematics, the loop may occur entirely in software:
\[ \text{conjecture} \rightarrow \text{proof attempt} \rightarrow \text{formal check} \rightarrow \text{revision}. \]
In programming:
\[ \text{code} \rightarrow \text{test} \rightarrow \text{error} \rightarrow \text{revised code}. \]
In materials discovery:
\[ \text{candidate} \rightarrow \text{synthesis} \rightarrow \text{measurement} \rightarrow \text{new candidate}. \]
The structure is the same.
What changes is the cost and speed of feedback.
And that difference determines how quickly AI can improve.
The Fastest Loop Often Wins
Imagine two research systems with equal intelligence.
One receives reliable feedback every ten seconds.
The other receives feedback once a month.
After a year, they have lived through very different amounts of scientific experience.
The first has completed millions of cycles.
The second has completed twelve.
This simple observation helps explain why AI has advanced so quickly in domains such as code generation.
It also explains why automated experimentation matters so much for AI4S.
The goal is not merely to make AI smarter.
It is to increase the clock speed of the scientific loop.
But Not Every Loop Should Be Fully Autonomous
There is an important qualification.
A loop can optimize the wrong thing very efficiently.
If the system has a poorly defined objective, an unreliable measurement, or a hidden bias in its experimental process, closing the loop may accelerate error rather than discovery.
Human oversight matters especially when experiments are expensive, dangerous, ethically sensitive, or difficult to reverse.
The goal should not be autonomy for its own sake.
The goal is a scientific process in which the right parts can proceed faster, while uncertainty and risk remain visible.
Science as an Information Engine
Viewed this way, science is an engine that converts uncertainty into knowledge.
Each experiment is not merely an event.
It is a question asked of the world.
The answer changes what we ask next.
AI for Science becomes powerful when AI participates in that conversation:
\[ \text{model asks nature} \rightarrow \text{nature answers} \rightarrow \text{model changes}. \]
But there is one obvious problem.
Nature often answers slowly.
That may be the most important bottleneck in the entire AI4S story.
Next in the series: Why Verification Is the Bottleneck
0 comments