← 返回论坛
科学探讨

When Neural Networks Tackle the Three-Body Problem: 100 Million Times Faster

三

三体粉丝

2026年7月11日 17:48

When Neural Networks Tackle the Three-Body Problem: 100 Million Times Faster


The three-body problem — predicting the motion of three celestial bodies under mutual gravity — has haunted physicists for over 300 years. Newton himself recognized its unsolvability in closed form. Poincaré later showed it was chaotic: tiny differences in initial conditions explode exponentially over time.


But here's what's fascinating: neural networks don't care about closed-form solutions.


In 2019, researchers from the University of Edinburgh and the University of Cambridge trained a neural network on 9,900 simulated three-body trajectories. The result? Their network predicted future positions 100 million times faster than the best numerical integrators, with only a tiny accuracy trade-off (about 1 part in 10⁵ for short-term predictions).


Why This Matters for Three-Body Fans


1. The Trisolaran's computational advantage makes sense now.

In Liu Cixin's world, the Trisolarans' sophons could monitor Earth in real-time by exploiting high-speed computation. The idea that brute-force machine learning can "solve" chaotic systems is no longer science fiction — it's 2019 science.


2. Chaos isn't the end of prediction — it's the beginning of a different kind of prediction.

Traditional physics says: give me the equations, I'll give you the answer. Machine learning says: show me enough examples, I'll give you a good guess. The three-body problem becomes less "unsolvable" and more "computationally tractable with enough training data."


3. This is exactly the kind of science that makes the Three-Body universe feel real.

The best sci-fi takes real scientific breakthroughs and asks "what if this went further?" A neural net that solves the three-body problem 100M times faster is a small step today. Extrapolate that a century — and you're in Trisolaran territory.


*Originally reported by MIT Technology Review (2019), republished on HN (2026).*

游客可读 · 发言需登录 [登录]

分享到:𝕏Reddit

登录后回复

登录