Revisiting the Form of Attention with Positional Encoding for Molecular Structures
ICML 2026 Workshop on AI for Physics, Seoul, South Korea
“It takes all kinds of people to make the world beautiful. You can do it too.”
— Sadhguru
Hi, I’m Klea Ziu, a final-year PhD student in machine learning at MBZUAI, advised by Martin Takac. I develop reliable and data-efficient ML methods for scientific discovery, with a focus on chemistry, catalysis, and sustainable technologies.
My work sits at the intersection of machine learning and the chemical sciences. I am especially interested in models that respect scientific structure and remain useful when experimental data are limited, noisy, or expensive. Recent directions include Neural ODEs for reaction kinetics, graph neural networks for catalyst design, and ML prediction of density of states (DOS).
Research Focus
Methods for Scientific Discovery
- Neural ODEs for reaction kinetics Learning dynamics that can model chemical reaction behavior with structure and interpretability.
- Graph neural networks for catalyst design Representing materials and molecular structure for more efficient catalyst discovery.
- ML prediction of density of states (DOS) Building models that connect electronic structure to useful downstream scientific decisions.
- Data-driven sustainable chemistry Using ML to support clean energy, low-emission industrial processes, and scalable catalysis.
Selected Publications
Topic:
MirrorCheck: Efficient adversarial defense for vision-language models
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026
Distinguished Paper Award, 6th AdvML Workshop, CVPR 2026, Denver, Colorado
Thinking like a CHEMIST: Combined Heterogeneous Embedding Model Integrating Structure and Tokens
IEEE Access, 2026
ψDAG: Projected Stochastic Approximation Iteration for Linear DAG Structure Learning
SPIGM Workshop, ICML 2026, Seoul, South Korea
Accelerated adaptive cubic regularized Quasi-Newton methods
Journal of Optimization Theory and Applications, 2025
Revisiting the Form of Attention with Positional Encoding for Molecular Structures
ICML 2026 Workshop on AI for Physics, Seoul, South Korea
Thinking like a CHEMIST: Combined Heterogeneous Embedding Model Integrating Structure and Tokens
IEEE Access, 2026
A deep neural network for oxidative coupling of methane trained on high-throughput experimental data
Journal of Physics: Energy, 2023
Talks & Workshops
Wellbeing Meets AI
Designed and delivered for the MBZUAI Fifth Anniversary Celebration
Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE
ψDAG: Projected Stochastic Approximation Iteration for Linear DAG Structure Learning
-
Okinawa Institute of Science and Technology (OIST), Okinawa, Japan
-
Young Researchers’ Association for Machine Learning (YAML) 2025, Heartpia Atami, Japan
Deep Neural Network Kinetic Model of Oxidative Coupling of Methane Using High Throughput Experimental Data
ACS Fall 2023: Harnessing the Power of Data
Data Science for Catalysis session · San Francisco, California, USA
Class of 2023 Valedictorian Address
MBZUAI Commencement 2023
Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE
Cubic Regularized Quasi-Newton Methods
Order up! The Benefits of Higher-Order Optimization in Machine Learning, NeurIPS 2022 Workshop
New Orleans, Louisiana, USA
Teaching
Introduction to Neuroscience, Psychology, Human Factors, and HCI Theory and Models
Mohamed bin Zayed University of Artificial Intelligence
Linear Algebra and Python
Academic tutoring in linear algebra and Python
Advanced Topics in Continuous Optimization
Project and assignment assessment, and examination organization
Mohamed bin Zayed University of Artificial Intelligence
Probability and Statistical Inference
Preparation of laboratory material, project and assignment assessment, and examination organization
Mohamed bin Zayed University of Artificial Intelligence
Blog
Notes on ML, Chemistry, and Research Life
October 2025
How to monitor, diagnose, and solve gradient issues in foundation models
A practical guide to vanishing gradients, exploding gradients, layer-wise gradient tracking, and stabilization techniques using PyTorch and Colab.
Read article →
September 2025
Synthetic data for LLM training
Where synthetic data is useful, where it is risky, and how GANs, diffusion models, LLMs, and meta-learning approaches generate training data.
Read article →
August 2025
Multimodal large language models
An overview of how multimodal LLMs combine text, images, audio, and other modalities to reason across richer inputs.
Read article →