Researchers at the National University of Singapore have developed a novel artificial intelligence framework capable of extracting macroscopic behavioral laws from microscopic data, addressing a longstanding computational bottleneck in materials science. Led by Associate Professor Qianxiao Li from the Department of Mathematics, the team published their findings in Physical Review Materials and presented the work at ICML 2026, alongside an arXiv preprint. Traditional material simulation requires tracking trillions of atomic interactions over extended periods, a process that remains computationally prohibitive even on advanced supercomputers. The NUS methodology circumvents this limitation by training AI models on small-scale systems and identifying a compact set of hidden variables that capture the collective dynamics of the entire assembly. Once these latent rules are learned, the models accurately extrapolate to predict the evolution of systems orders of magnitude larger than those used during training. The research introduces two complementary breakthroughs. The first enables accurate forecasting of large stochastic systems using simulations restricted to a fraction of the constituent particles. The team provided mathematical proofs validating this approximation and demonstrated its efficacy across biological population models and simulated alloys containing over 500,000 atoms. The second advancement equips the AI to process unordered particle assemblies, such as fluids and polymers. Rather than tracking individual constituents, the algorithm learns distribution-level patterns, ensuring permutation invariance and consistent outputs regardless of particle labeling. These capabilities deliver substantial improvements in both scalability and computational efficiency. By inferring large-scale behavior from limited microscopic data, the framework preserves essential physical properties while eliminating the need for exhaustive particle-by-particle calculations. Researchers can now model realistic materials at previously inaccessible scales without sacrificing accuracy. Looking forward, the NUS team aims to integrate these AI methodologies with experimental datasets and mesoscopic simulations. The long-term objective is to deploy automated material discovery platforms capable of rapidly forecasting system behavior, thereby accelerating innovation in energy storage, next-generation electronics, and advanced manufacturing.