Acceleration, agency, and foundations: ten influential arXiv directions in AI for materials science (2025)

Abstract A small number of preprints each year present results that enable faster downstream work by decreasing adoption challenges. The year 2025 revealed several arXiv papers that focused on artificial intelligence (AI) applications for materials science, demonstrating that their content would be reused through the distribution of code, weights, datasets, benchmarks, and the development of new implementations and third-party connections. This perspective organizes ten research papers based on their functional roles, which include fundamental atomistic models, controlled generative design and synthesis, data infrastructure and reasoning benchmarks, and large language model (LLM)-driven agency to identify shared methods for speeding up processes. We argue that AI-for-materials workflows are converging toward a five-layer stack: (i) transferable atomistic foundation potentials, (ii) generative engines for crystals and amorphous structures, (iii) feedback/control loops for inverse design, (iv) datasets and benchmarks that constrain and compare claims, and (v) LLM components for decision-making and planning. We discuss what each layer enables, where failure modes remain, and what a robust AI-for-materials pipeline should look like beyond 2025. Papers were selected using a transparent proxy-based rubric (artifact release, integration readiness, and early reuse signals), rather than citation counts, which lag for recent preprints.