What Is Science as a Service?
Bringing Scientific Expertise to Industry and Society

What Science as a Service means here

In this article, Science as a Service (ScaaS) means a proposed approach that combines researchers' expertise, measurement, data processing, physics-based models, and AI to support work from defining a problem through to testing possible solutions.

The starting point does not have to be “We want to introduce AI.” It can be a specific problem: “We want to understand when defects occur,” “We want to reduce the work involved in measurement,” or “We want to detect changes in equipment sooner.” Deciding what to measure and what to compare to make a useful judgment is itself a way to put scientific expertise to work.

Connecting observation to improvement

The basic process is a loop: observation → hypothesis → model → testing → improvement.

First, observe what is happening in practice. Look at the conditions under which readings were taken, including the time, the equipment's state, and the surrounding temperature, as well as the readings themselves. If units or recording intervals differ, they need to be aligned before making comparisons.

Next, form a hypothesis, such as “Could temperature changes be related to variation in dimensions?” A model expresses that hypothesis through equations, calculations, or a system that learns from data.

Once a model has been built, test it using data that was not used to build it, or additional measurements. Examine the conditions in which its predictions were wrong, then revise the measurement method or the hypothesis. Repeating this process in light of the results brings the work closer to evidence that can inform practical decisions.

Combining AI and physics

Putting AI × Physics into practice is important here. A physics-based model represents known mechanisms behind a phenomenon, such as heat transfer or the balance of forces. AI can help identify relationships among many variables, or patterns in data that are difficult to describe with simple equations.

For example, one possible approach is to use a physics-based model to calculate the approximate course of temperature changes, then use AI to correct the difference between the calculation and the measurements. AI predictions can also be checked for consistency with basic physical conditions, such as conservation laws.

The right combination depends on the problem. Sometimes a simple comparison on a graph reveals possible causes. Whether complex AI is needed is also something to test.

Physics Simulation as a Service

One concrete way to put ScaaS into practice is Physics Simulation as a Service. It uses physical models and numerical simulations to compare designs before experiments or prototypes, evaluate how instruments and sensors respond, and explore conditions that could test a hypothesis.

For example, a model can compare how a sensor's response might change with its shape or operating conditions. Such predictions help decide which conditions to try in an experiment. They still need to be checked against measurements and the assumptions of the model.

• Inputs: The problem to solve; the equipment's geometry, materials, and operating conditions; and any available measurements to compare with the simulation.
• Process: Agree on the model's assumptions, the scope of the calculation, and evaluation criteria. Compare candidate conditions, check the numerical calculations, and compare predictions with existing measurements or reference experiments to examine where the model is valid.
• Deliverables: Reproducible code, models, and calculation settings; figures comparing the conditions; and a report describing predictions, uncertainty, conditions of use, and limitations that remain untested.

Scientific Data Analysis as a Service

Scientific Data Analysis as a Service is another concrete form of ScaaS. It starts with measurement data to clarify a practical problem: organizing records, checking noise and unusual readings, applying statistical analysis and physics-based estimation, visualizing results, and assessing uncertainty.

Rather than treating an unusual value as a cause in itself, the analysis asks whether it reflects a change in the phenomenon, a measurement error, or a difference in recording conditions. Finding an association can suggest a hypothesis; it does not by itself prove a cause.

• Inputs: The question to answer, equipment and measurement conditions, and the available data, including units, timestamps, and relevant operating records.
• Process: Agree on what counts as improvement, the reference for comparison, and how measurement errors will be considered. Check missing values and units, document any cleaning, analyze the data, and test the findings against separate data or new measurements.
• Deliverables: Reproducible analysis code and settings, documented data preparation, figures, and a report explaining the findings, their uncertainty and limitations, and what should be measured next.

The phenomena to be studied, the available data, and the extent of validation are agreed for each project. Phenomena that cannot be measured and models that have not been validated limit what either service can establish.

The two services complement each other: simulate a prediction, compare it with real measurements, and use the difference to refine the model or decide what to measure next. This loop connects AI × Physics with decisions in manufacturing and society. AI can assist calculation and analysis, while people remain responsible for judging the assumptions, validation, and use of the results.

What kinds of problems could it help with?

The following three examples are illustrative scenarios, not accounts of completed projects.

Variation in manufacturing

Compare records of materials, processing conditions, temperature, and quality to narrow down the conditions associated with variation. Change the candidate conditions in small trials to test whether they actually lead to improvement.

Unusual sensor readings

Compare readings from several sensors with equipment operating records to detect changes that depart from the usual state. Inspections and reference measurements are also needed to distinguish a change in the equipment from drift in the sensor itself.

Environmental measurement

Organize measurements such as temperature and humidity by location and time, then examine how they vary. Check the effects of differences in measurement positions or instruments before using the results to decide on additional measurements or changes in placement.

What could the deliverables be?

A deliverable need not be limited to a report summarizing the analysis. It could also be a measurement plan, organized data, an analysis procedure that can be run again, a predictive model, or an interface for viewing the results.

What matters is that the recipient can decide what to do next. Alongside “What did we learn?”, document “Under what conditions can this be used?”, “What remains untested?”, and “What should we measure next?” Agree in advance on evaluation criteria to check in practice, including false detections, time spent on the task, and the effort required to operate the system, as well as accuracy.

Sharing limitations and data-handling rules

Finding a relationship in data does not, by itself, establish a cause. Bias in the records or measurement error can also affect the results. Predictions may fail under conditions the model has not encountered, and the benefits of a proposed improvement remain unknown until tested. Start with small trials and establish the conditions in which the approach is useful.

Manufacturing conditions and equipment records may also contain confidential information. Before sharing data, check what is needed, where it will be stored, who can access it, how long it will be kept, and how it will be deleted. Sending it to an external AI service or reusing it for another purpose should also be based on prior agreement.

Starting with the problem you want to solve

When starting a discussion, the following information helps define the problem:

• What is causing difficulty, and what would count as an improvement
• What data is available, and when and how it was measured
• What has already been tried, and the results
• Deadlines, budget, and what can be changed in practice
• Information that cannot be shared, and conditions for handling the data

You do not need to have everything ready. If there is no data yet, the discussion can begin with what to measure. The aim is to connect scientific methods with everyday decisions and improvements. Science as a Service, as defined here, is a proposal for creating that connection.