AI learns hidden rules from microscopic data for large-scale material behavior
Researchers have developed AI methods that can predict the macroscopic behavior of complex materials by learning from microscopic data, significantly reducing the need for extensive simulations.
Scientists at the National University of Singapore (NUS) have devised novel artificial intelligence (AI) techniques capable of deducing the large-scale behavior of intricate materials directly from microscopic observations. These methods automatically identify a limited set of underlying variables that encapsulate the collective dynamics of a system, enabling predictions of material evolution over time with reduced reliance on computationally expensive simulations.
Understanding material properties at a macroscopic level is crucial for developing advanced technologies, but the sheer number of atomic interactions makes simulating every atom over extended periods often infeasible. A significant hurdle in materials science has been linking microscopic processes, such as atomic movement, to observable macroscopic properties, with existing approaches frequently demanding prohibitively costly large-scale simulations.
The NUS team, led by Associate Professor Qianxiao Li, has introduced AI methods that learn system dynamics from microscopic data without tracking each individual particle. Instead, the AI identifies a small number of hidden variables that represent the system's collective behavior and forecasts their temporal evolution. This approach has been validated on systems ranging from biological models to simulated alloys comprising over 500,000 atoms, proving that it can accurately predict the behavior of very large stochastic systems using simulations performed on smaller counterparts.
Furthermore, a second breakthrough allows these AI models to learn from systems lacking inherent ordering among their microscopic components, such as fluids or particle assemblies. The AI achieves this by learning the overall distribution patterns of particles, rather than individual tracking, ensuring consistent results regardless of particle labeling order. These combined advancements enable the creation of efficient macroscopic models from microscopic data while preserving essential physical accuracy, offering scalability to study previously inaccessible material scales and applicability to a wide array of scientific challenges.
This development presents a significant leap in computational materials science by enabling AI to infer effective macroscopic laws from microscopic data. By reducing the computational burden of simulations, these methods could accelerate the design and discovery of new materials for applications in energy, electronics, and manufacturing, potentially impacting areas requiring in-situ material characterization and production.
Edited by the news editor with AI from the original report — please refer to the original source.