Ahmad Arif

Energy Transition Specialization · Cohort 2026

The Role of Geometric Representations in Coarse-Grained Graph Neural Networks for Metal–Organic Frameworks

This project investigates whether and how geometric information can be reintroduced into coarse-grained graph neural networks (GNNs) for metal-organic frameworks (MOFs). Starting from a CG2NN baseline that encodes only building-block identity and connectivity, the study systematically ablates three levels of geometric detail: (1) RBF-encoded pairwise metal–linker distances, (2) line-graph bond angles, and (3) global persistent-homology descriptors of the pore space. All variants are evaluated on two structurally distinct datasets — QMOF (dense, building-block-dominated) and hMOF GA-51k (porous, geometrically diverse) — across nine property targets, and benchmarked against atomistic GNNs (CGCNN, MEGNet, ALIGNN) and a classical random forest. The central finding is a **decoupling principle**: geometric features help exactly when continuum geometry varies independently of building-block identity (hMOF), and are neutral or harmful when chemistry fixes geometry (QMOF).

Read the final report (PDF)

Artificial Intelligence Specialization · Cohort 2025

Core-EMS: A Reinforcement Learning Approach for Intelligent Building Energy Management

Commercial and residential buildings consume a significant share of global electricity. Traditional rule-based HVAC schedules are static and seldom optimise for varyingoccupancy or weather. We aim to learn an adaptive control policy with Deep Reinforcement Learning (DRL) that minimises energy use while keeping indoor comfortwithin bounds. Our system should quickly adapt to unseen buildings or climates by transferring experiencelearned elsewhere.

Mentor: Elias Al Ghazal

Read the final report (PDF)

See it in the program report (PDF)