A little real data goes a long way

Real measurement is the most valuable thing a model can learn from, and the scarcest. The useful question is how much each real measurement can be made to carry. A model that learns from a checked simulation first, and from real data second, needs far fewer real measurements to reach the same accuracy. One registered study puts numbers on how far that goes, where it stops, and what a wrong simulation costs.

The question

Low-voltage feeders carry power the last stretch to homes and businesses. Most of them have no meter. A grid operator who wants to know the power through an unmetered feeder has two options: install a meter and wait a year, or predict it.

A learned model can predict it, if it has seen enough metered feeders. The difficulty is that "enough" is often more than exists. So the study asked a precise question: if a model learns from Living Worlds' simulation of feeders first, and then from a handful of real measured feeders, does it predict feeders nobody has measured better than the same model trained on those real measurements alone?

What was tested and what was held out

The test set was 30 real feeders that no model had measured, over a full year that no model had seen. The real data came from the FeederBW dataset of Netze BW and the Karlsruhe Institute of Technology, published by Treutlein and colleagues in Scientific Data in 2026 under CC BY 4.0.

Two versions of the same model were compared at each budget of real data. One learned from the simulation first, then from the real feeders. The other learned from the real feeders alone. A third arm mattered as much as the first two: the same model taught by a deliberately wrong simulation.

The study was registered before the test year was read, and scored once.

The result

The verdict was Met.

With 5 measured feeders, the model that learned from the simulation first made a normalised error of 0.712, against 0.880 for the model trained on real measurements alone. That is about 19% less error.

With 10 measured feeders, 0.605 against 0.695.

With 80 measured feeders, the two were level.

The model taught by a deliberately wrong simulation did worse than either.

What it means

Three things follow, and each is as important as the headline.

First, the advantage is largest exactly where real data is scarcest. A handful of measurements is the usual starting point for anything new. That is where the simulation earns its keep.

Second, the advantage shrinks as real data grows, and at eighty feeders it is gone. That is what a sound simulation should do. Once there is enough of the world on hand, the world speaks for itself. A simulation that still claimed a large advantage with plenty of real data would deserve suspicion.

Third, the wrong simulation made things worse. That arm is the control that gives the result its meaning. If any prior helped, the gain would say nothing about the simulation. Because a wrong one hurt, the gain from the checked one is evidence that it carried real information about how feeders behave. Simulation helps only when it has been held to measurement.

What it does not show

The study covers one region and one operator in Baden-Wuerttemberg, 30 test feeders, scored on hourly means. It does not show the same gain in another grid, at fifteen-minute resolution, or in a different domain. It does not show that simulation can stand in for measurement: at every budget, the model still learned from real feeders, and with enough of them the simulation added nothing.

It is one result, registered in advance and reported as it came out. Its strength is that it says exactly how far the simulation took the model, and no further.

The wider pattern

The study sits inside a broader shift in how models are built. Learn broad structure from a large body of data first, then adapt to the specific task with a little of the right data.

Microsoft's Aurora, published in Nature in 2025, was trained on more than a million hours of diverse geophysical data and then fine-tuned, at modest computational cost, to forecast air quality, ocean waves, tropical cyclone tracks and high-resolution weather, outperforming operational forecasts on those tasks. The large first stage carries the general physics of the atmosphere; the small second stage carries the particulars of the task.

Work on generated data points the same way. Shumailov and colleagues showed in Nature in 2024 that models trained recursively on their own output lose the tails of the real distribution. Kazdan and colleagues, in work revised in March 2025, found that models stay stable when real data is kept and accumulated alongside what is generated. Real data stays in the loop. Generated data, if it is good, makes it go further.

The feeder study adds the part those results leave open: a simulation that has itself been checked against the world, a test on units and a year never seen, and a deliberately wrong simulation to show that the help comes from being right.

A second result on the same feeders

An earlier registered study took on the same kind of problem directly. Living Worlds predicted the power flowing every fifteen minutes through 40 real unmetered feeders, from the metered feeders around them, for a year it had never seen. It came out ahead of calibrated physics on 38 of 40 feeders and of a learned model on 30 of 40, with 7% lower error overall than the learned model and 29% lower than physics, and ahead of both in every season. Verdict: Met.

Taken together, the two results describe a method that is useful on its own and that makes a little real data go much further when combined with it.

Where Livgens fits

Livgens represents the thing exactly, in sound, light, matter and motion, and Living Worlds runs it forward through its lifecycle: a feeder, a building or a grid lives for years, month by month, ageing and wearing. What comes out is data a model can learn from and be tested against, in places where none was ever collected. Every thing carries its measurements and the range they hold for, so a simulation stays inside what is known. And each result is registered before its data is read and scored once on measurements no model has seen, with the misses published too. Simulation makes a little measurement go much further, and the results say exactly how far.