We are applying for compute resources on Tetralith/Arrhenius to support the project "Learning Admissible Abstractions for Optimal Planning." The project investigates how machine learning can be used to automatically learn abstractions that are inherently admissible — that is, abstractions that provably preserve optimality guarantees in classical planning. Our approach combines machine learning and symbolic methods in a hybrid neural-symbolic framework. A key challenge in optimal planning is the construction of informative and admissible heuristics; hand-crafted abstractions such as pattern databases require substantial domain expertise and do not generalise across tasks. We aim to overcome this limitation by learning abstractions directly from problem instances using LLMs and later on other ML approaches, ensuring that the learned representations remain admissible by construction. The compute resources are needed to run planning experiments across standard benchmark domains, evaluate the performance learned abstractions, and compare against state-of-the-art planning heuristics. This work has the potential to significantly advance the scalability of optimal planners by replacing brittle hand-crafted heuristics with flexible, learned ones.