Accelerating and scaling MLIP inference | LeMaterial Reading Group скачать в хорошем качестве

Accelerating and scaling MLIP inference | LeMaterial Reading Group

Paper Link: https://www.arxiv.org/abs/2601.21147 MLIPs face two primary bottlenecks preventing them from reaching realistic simulation scale: inference time and memory consumption. In this talk, I will discuss two recent works focused on addressing both problems: 1) DistMLIP (ICLR 2026), a distributed inference platform for MLIPs predicated upon zero-redundancy 1-hop graph partitioning and 2) smooth dynamic cutoffs, a novel method to effectively prune edges of the underlying atomic graph while maintaining accuracy and accelerating inference. Kevin is a second year PhD student at Carnegie Mellon University researching machine learning methods in scaling molecular simulation as well as LLM agentic post-training for scientific discovery applications. On molecular simulation, he is particularly interested in developing machine learning tools to simulate atomic systems of real-world sizes and ultra-long timescales. LeMaterial Reading Group is a recurring gathering where we discuss recent papers at the intersection of AI, chemistry and materials science