Multitask learning with semiempirical orbital charges enables sample-efficient MLIPs
Researchers have developed a new method for training machine learning interatomic potentials (MLIPs) that significantly improves sample efficiency and accuracy. By incorporating semiempirical orbital charges, the model achieves a 46% reduction in energy mean absolute error and requires five times less data than traditional energy-only models. This approach, which uses computationally inexpensive orbital charges during training, allows for the development of accurate foundation models for complex chemical systems without sacrificing inference efficiency. AI
IMPACT Enhances accuracy and data efficiency for developing foundation models in chemistry and materials science.