By: George E. Froudakis - University of Crete, Greece

When: November, 6 - 10 AM

Where: Salón de Actos

Abstract: 

In the era of big data, machine learning (ML), a subfield of artificial intelligence, offers powerful tools widely used in science and industry due to their significantly lower computational cost compared to conventional methods. However, the accuracy of ML predictions relies heavily on identifying the right parameters (features) that enable effective learning from past data. Additionally, the quality and size of the dataset used to train the algorithm are critical for reliable predictions.

In traditional ML, feature extraction involves manually identifying and engineering key characteristics from raw data to enhance model performance. In contrast, deep learning (DL) automates this process through layered neural networks that learn relevant features directly from data, making it especially effective for complex tasks like image and speech recognition. This end-to-end approach allows DL models to process high-dimensional data with minimal preprocessing.

Here, we propose a deep learning framework for predicting gas adsorption properties of materials, using the Potential Energy Surface (PES) as a descriptor. This descriptor is universally applicable to various materials and properties, as PES uniquely combines the material s classical structural characteristics with its quantum electronic structure, encapsulating the material s information in both classical and quantum worlds - a fusion of chemistry and physics.

Furthermore, we developed a generalized deep learning framework that predicts gas adsorption properties from a point cloud representation of the material’s structure, eliminating the need for hand-crafted features. This approach requires only the unit cell structure, offering a complete yet simplified material representation.

Our algorithms were tested on CO2 uptake predictions in MOFs, where our deep learning models significantly outperformed the best ML-Random Forest model in the literature, which relied on geometric descriptors. This methodology was further validated across different materials (e.g., COFs), gases (e.g., CH4), and datasets, demonstrating broad applicability for diverse materials and properties.