Harvard researchers created a new machine learning tool that predicts how materials react to electric fields with extreme accuracy, matching quantum-level precision. This tool can handle systems as large as a million atoms, far surpassing traditional methods that are limited to just a few hundred atoms. Machine learning is a type of computer program that learns patterns from data to make predictions.
This breakthrough could help scientists design advanced materials for things like energy storage or new technologies by running highly accurate, large-scale simulations. The research is published in Nature Communications.
For over 30 years, scientists have used density functional theory to study how atoms and molecules behave. Density functional theory is a physics-based method that solves quantum mechanical equations to predict material properties with high accuracy. However, this method is slow and can only model small systems because it requires a lot of computing power.
Machine learning has recently helped study larger systems while keeping the same accuracy, but it struggled to predict how materials respond to external forces, like electric fields. Earlier machine learning models often ignored important physical rules, leading to less accurate predictions.
Advancing material simulations
To fix these problems, the researchers developed a new machine learning method called Allegro-pol. This method combines different quantum behaviors, like energy and polarization, into a single model. The model uses data from density functional theory to train and check its predictions, ensuring it follows the correct physical laws. Allegro-pol builds on an earlier tool called Allegro, which was good at simulating energy and forces in atoms. The new version also captures how atoms react to external changes, like electric fields, which is key for designing materials like ferroelectrics or dielectrics.
The researchers tested their method on silicon dioxide and barium titanate, successfully predicting their electrical properties and behaviors. This shows the tool can handle real-world materials used in technologies like capacitors or non-volatile memory. Research co-leader Stefano Falletta noted that traditional quantum methods are limited to small systems, but this machine learning approach can scale up dramatically, making it possible to study much larger systems. Now at Radical AI, Falletta believes this work could push materials science forward, combining better theories, faster computers, and advanced models to speed up discoveries in exciting ways.