The field of atomistic modeling and materials discovery is moving towards the development of more efficient and accurate models that can handle large and diverse datasets. Researchers are exploring new approaches to pre-training graph foundation models, including multi-task parallelism and novel architectural designs, to improve their transferability and scalability. Additionally, there is a growing interest in developing universal models that can quickly and accurately compute properties from atomic simulations, with applications in chemistry, materials science, and beyond. Reinforcement learning frameworks are also being proposed to address the challenges of inverse materials design under data scarcity, leveraging expert knowledge and automated refinement strategies to improve model reliability. Furthermore, multimodal learning approaches are being developed to learn from elemental composition and X-ray diffraction data, without requiring crystal structure input. Noteworthy papers in this area include:

• UMA: A Family of Universal Models for Atoms, which presents a family of universal models for atoms that can perform similarly or better than specialized models without fine-tuning.
• XxaCT-NN: Structure Agnostic Multimodal Learning for Materials Science, which proposes a scalable multimodal framework that learns directly from elemental composition and X-ray diffraction data.
• AIMatDesign: Knowledge-Augmented Reinforcement Learning for Inverse Materials Design under Data Scarcity, which introduces a reinforcement learning framework that addresses the limitations of machine learning models in high-dimensional spaces and effectively incorporates domain expert knowledge.