Japanese Kyushu University researchers used deep learning to make a precise population forecast for all of Japan through 2100.
With half the country turning into empty land, South Korea’s fertility rate of 0.80 children per woman places it on an even steeper slope.
Japan’s Kyushu University researchers divided the entire country into more than 1.55 million grid cells, each 500 meters by 500 meters, and used artificial intelligence to predict population through 2100. The hit rate exceeded 99 percent.
The year 2100 is projected to bring a Japan in which more than half of those grid cells are no longer inhabited. Population growth will effectively be limited to Tokyo, while even other major cities will survive only in their city centers.
Average household size, which was 2.26 people in 2020, is forecast to fall to 1.39 by 2100, while the foreign population share rises above 6 percent. It is a shift toward a society of people living alone and a society shaped by immigration.
South Korea’s total fertility rate in 2025 stood at 0.80, far below Japan’s 1.15. If the same map were drawn for Korea, the clock for population decline would move even faster.
An artificial intelligence has drawn a population map 100 years into the future. It did so by breaking all of Japan into tiny 500-meter-by-500-meter squares, then calculating how many people would live in each square through 2100. That was the result released in March this year by a Kyushu University research team in a international academic journal.
The method may sound unfamiliar, but the principle is straightforward. The researchers combined “deep learning,” an AI technology, with grid-level population data that the Japanese government has accumulated every five years since 1995.
Deep learning is a method in which a machine learns vast amounts of data on its own and finds hidden patterns without a person entering rules one by one. The team also added nighttime satellite light data, using brightness as a clue that people and economic activity are concentrated there.
To test whether the model could really predict the future, the researchers checked it by having it reproduce the past. When they asked the model to identify older population data whose answers were already known, the accuracy exceeded 99 percent. The results also broadly matched the Japanese government’s official population projections. In other words, the AI-generated future map was not a fanciful imagination.
Until now, population forecasts have generally started with large figures for the country as a whole or for administrative regions, then broken them down into smaller parts. That often obscured local conditions.
This study took the opposite approach: it first calculated changes in each small cell, then combined them into a national picture. It is different in that it captures fine-grained differences at the neighborhood level.
◆ Japan in 2100, as drawn by AI
The future shown on the map is stark and harsh. Japan’s population is expected to fall below 100 million around 2055, decline into the 90 million range by 2070, and only near the end of the century begin to level off and then rebound slightly. Over 80 years, that amounts to the disappearance of an entire generation.
Even more shocking is where people remain. The model found that by 2100, more than half of Japan’s grid cells will have become completely uninhabited empty land.
Population growth will effectively be seen only in Tokyo. Even major cities such as Osaka, Nagoya, and Sapporo will see their outskirts empty first, with only city centers barely holding on. The pattern is one in which young people from the regions move to major cities in search of jobs, and then even people in those cities are eventually drawn toward Tokyo.
Viewed by age group, the shadow is even deeper. The number of children under 15 will quickly decline even in Tokyo. The working-age population will keep falling until 2070, only then stabilizing. The elderly population will rise for a while and then begin to decline from mid-century onward, but that is not something to celebrate, as it means older adults are also beginning to pass away.
Family structures will also change. The average number of people in a household will shrink from 2.26 in 2020 to 1.39 in 2100. The paradox is that while the population falls, the number of households will actually keep rising. That signals more people living alone and a trend toward not marrying becoming entrenched. The researchers pointed out that the risk of isolation among older adults would grow.
The empty spaces will be filled by foreigners. The model predicts that by 2100, foreigners will make up more than 6 percent of Japan’s total population, and more than 3,000 grid cells will become majority-foreign. The number of foreigners living in Japan is projected to exceed 5 million around 2075. The population crisis is expected to drive a transition into an immigration society.
◆ With a fertility rate of 0.80, Korea is already ahead of Japan
There is a reason this map does not feel like someone else’s problem. South Korea’s population clock is moving even faster than Japan’s.
Japan’s total fertility rate in 2024 was 1.15 children per woman, the expected number of children a woman will have over her lifetime. That same year, Japan’s number of births fell below 700,000 for the first time. But South Korea’s total fertility rate in 2025 was only 0.80. Although it has recovered to the 0.8 range for the first time in four years, it remains far below Japan’s and is the lowest level in the world. Korea is descending the same slope Japan is walking, but at a steeper angle.
The gap widens further when viewed by region. Last year, the highest fertility rate in South Korea was 1.10 in South Jeolla Province, while Seoul had the lowest at 0.63. The AI map’s pattern of concentration in the center and disappearance at the edges is already a reality in Korea.
The government’s long-term population projections point in the same direction. South Korea’s population, which stood at 51.67 million, will fall below 50 million in 2041 and decline to 36.22 million in 2072.
That is a scale equivalent to returning to 1977 levels. The share of people aged 65 and older will surge to 47.7 percent that year, making Korea a country where one in every two people is elderly. The median age will rise from 44.9 to 63.4, and the dependent population supported by every 100 working people will increase from 41 to 119, nearly tripling.
The provinces are already flashing warning signs. By the “extinction risk index,” which divides the number of women aged 20 to 39 by the number of people aged 65 and older, 130 municipalities and districts nationwide were classified as extinction-risk areas, more than half of the total. Of those, 57 were in the highest-risk category.
Even the metropolitan city of Busan has entered the risk stage for the first time. Since the Seoul metropolitan area overtook the non-capital region in population in 2019, the gap has continued to widen, and now 51 percent of the country’s population is concentrated on one side of the nation. Korea is following Japan’s “Tokyo-centric” pattern almost 그대로, but as a “capital-area-centric” country.
◆ The true value of a detailed map is preparation
So what does this forecast leave us with? The key point is not that AI will stop population decline, but that AI can help us prepare for it.
A 500-meter grid forecast turns vague anxiety into concrete administrative information. It can show in advance which neighborhood schools will close for lack of students, which bus routes will be cut, and which local hospitals may disappear because of fewer patients.
It becomes a precise compass for deciding where limited budgets should go first. In Japan, analyses have already emerged showing that medical service gaps will widen as population decline advances.
The conclusion reached by the Japanese researchers is along those same lines. Rather than trying to support every region equally, they argue for consolidating and strengthening places that can survive, orderly restructuring where decline is unavoidable, and tailored support for elderly populations and immigrant communities. In other words, they suggest rebuilding the social framework to match the trend rather than denying the collapse.
South Korea has not stood still. Over 18 years, it has poured nearly 380 trillion won into low-birthrate policies, and each year it has put about 1 trillion won from the local extinction response fund into population-declining regions to support jobs and the inflow of young people.
Even so, the fact that the fertility rate has barely risen exposes the limits of a broad-brush spending approach. Money spent without precisely knowing which areas are emptying makes it difficult to judge effectiveness.
The fact that the 2025 fertility rate rebounded for a second consecutive year is a small source of hope. It is understood as the result of the large cohort born in the early 1990s entering marriage age and of gradually changing attitudes toward marriage and childbirth. But the rebound is still too shallow to reverse such a steep decline.
The empty map drawn by AI for Japan is ultimately a mirror reflecting Korea. If Korea is already moving ahead down the same slope with an even lower fertility rate, the first step is to draw that map for ourselves in advance.
Knowing where the emptiness will spread means gaining time to decide how to fill it, or whether to let it empty in a managed way.