Blog · Research · Oct 21, 2026 · 5 min
Living Worlds predicted how real rooms sound in arrangements and positions nobody had measured, and how daylight falls on office desks on days nobody had read: met for a room never seen (20 of 20), four rooms never seen (4 of 4), sound beyond what was measured and daylight at the desk; not met in a second room.
How will a room sound in an arrangement nobody has measured? Put panels on the ceiling, open a wall, stand somewhere new, and the way sound dies away and how clearly speech carries both change. Measuring every arrangement and every seat is slow; a hall has thousands of seats and a variable room has more configurations than anyone will record.
Living Worlds asks whether it can predict those changes from a handful of measurements, in rooms and positions it has not heard. The same question applies to light: how much daylight reaches a desk, every ten minutes, on days nobody has read.
Each test was registered before the held-out measurements were opened, and scored once. In the room tests, "never seen" means no decay in that room had been computed, or no response at those positions opened, when the rule was written. In the daylight test the data was not read until the prediction was registered.
Acoustics has a natural trap here. A model that has heard nearby measurements can look excellent simply by remembering them. So most of these tests compare against memory itself: the nearest measured arrangement, or the nearest measured positions. A prediction counts only if it beats what was already known.
This matters more as the field leans on generated data. Large collections such as RIR-Mega, 50,000 simulated room impulse responses released in 2025, make it easy to train on rooms that were computed rather than measured. The tests below are all scored on measured rooms and measured light.
A room never seen, one configuration at a time used the public dEchorate room, whose walls, ceiling and floor can each be bare or covered in absorbing panels, measured with 30 microphones and 7 loudspeakers. Living Worlds predicted each new arrangement before it was measured, from those measured so far, across 2,100 measurements in 10 configurations.
It beat memory of the nearest measurements in all 20 seeded orders, where 18 were needed, and in 133 of the 160 configurations it predicted. Met (20 of 20). Errors sit near 10% of the decay time.
Before that result, A second room had tested a second measured room, and neither registered condition was met. Memory learned the room's own bias, and within a single room memory is as good as learned practice. Not met. The lesson set what came next: physics earns its place where measurements run out.
Sound beyond what was measured tested exactly that edge. In a real variable-acoustics room, fitted on the middle 70% of configurations, Living Worlds predicted the 15% at each end, arrangements outside the range it learned from. It beat memory alone by +0.138 (95% interval 0.134 to 0.142), winning 6,846 of 7,869 held-out measurements. Met. No run had read the ends before scoring.
Four rooms never seen, from twelve measured positions moved from arrangements to positions. Each of four public rooms, a classroom, a dampened room, a hallway and a complex room, was measured at 12 positions and predicted at every other one: how long sound takes to die away (T and EDT) and how clear it is (C50), in each octave band, scored against the smallest difference a listener can hear. 1,282 positions in all, none opened before registration.
The calibrated formula carried to new positions by a learned field scored 0.394, against 0.325 for the same formula corrected by its five nearest measurements, the method that had met twice before: 0.069 higher (95% interval 0.066 to 0.073), higher at 1,089 positions and lower at 193, and higher in all four rooms, where three were needed. Met (4 of 4 rooms). Image sources with the same memory scored 0.312.
The study reports one more number in full. The mean of the five nearest measured positions alone scored 0.417, above the field in every room: classroom 0.418 against 0.416, dampened 0.273 against 0.271, hallway 0.467 against 0.405, complex 0.452 against 0.444.
Light falls through a room as sound does, shaped by what it meets. Daylight at the desk, on days nobody had read predicted the daylight on six desks of a real office at Berkeley Lab every ten minutes, for eleven days it had never seen. It reached a root-mean-square error of 58.1 lux, ahead of calibrated physics (115.3) and a learned model (64.3), and roughly halved the physics error at the desks nearest the window, where the light changes most. Met.
Together these results describe where prediction pays. Inside a single well-measured room, memory is hard to beat, and the second room showed it. Where measurements run out, in arrangements beyond the measured range, in new configurations and in rooms never seen, the predictions beat what was already known, and the daylight test showed the same at the desks where light changes fastest. A little real measurement goes further when the world carries it to the places nobody has stood.
In the four-rooms test, the registered comparison was the learned field against the formula with its memory. It does not show the field ahead of memory alone: the mean of the five nearest measured positions scored higher in every room, by a hair in three and clearly in the hallway. The single-room result rests on one public room with panels on or off, and its errors sit near 10% of the decay time. The second room was not met. The daylight result covers six desks in one office over eleven days.
Living Worlds simulates the places people live in as well as the people: how a room sounds and how light falls among them. Its site lists a venue digital twin, where a production tests a show in the hall before load-in, and states what is shown so far: how a room sounds in arrangements beyond those measured. Each result is checked against real measurements before it is relied on, and used within the range where it held.
Blog
A feeder model that learned from Living Worlds' checked simulation first, then from a handful of real measurements, predicted 30 unmeasured low-voltage feeders better than the same model on real measurements alone; power through 40 unmetered feeders beat physics and a learned model; thermal comfort in fifteen countries came close but was not met.
Spatial audio got a full international standard in 2025, wave-based simulation reached open datasets, and learned acoustic fields now predict how a room sounds from a handful of measurements. The open question is how far those predictions hold beyond what was measured, and how to show it.
Europe's grid operators, its cities and its climate programme all moved digital twins from plans to infrastructure in 2025-26. The architecture is now well specified; the step that decides whether a twin can be relied on is a prediction scored against what it never measured.