- Title
- Heterogeneous catalyst discovery for renewable energy applications using machine learning
- Creator
- Li, Xinyu
- Relation
- University of Newcastle Research Higher Degree Thesis
- Resource Type
- thesis
- Date
- 2022
- Description
- Research Doctorate - Doctor of Philosophy (PhD)
- Description
- The world desperately needs more clean and sustainable energy technologies to cope with energy shortages and man-made climate change. Currently, certain renewable energy technologies, such as fuel cells, biomass reforming and CO2 reduction, are inhibited by a lack of efficiency and safe and abundant catalyst materials. Novel material discovery is an expensive, trivial and time-consuming process, compared to the development of commercially viable materials, which generally requires around 10 years from the time of initial discovery. This thesis aims to use machine learning (ML) techniques to accelerate the discovery of new heterogeneous catalyst materials. It focuses on heterogeneous catalyst materials that accelerate chemical reactions, where adsorption energy is an important descriptor of catalysis activity and selectivity. Density functional theory (DFT) is used as the primary method to calculate this value, but its high computational cost limits its application for screening a large amount of potential candidates. ML techniques model the relationship between complex inputs and targets and represent one promising alternative method for catalyst design and discovery as they offer a significant reduction in computational cost compared to DFT. However, high-accuracy and robust ML methods, including both representations and algorithms, are relatively unexplored. This thesis presents four studies investigating high-accuracy ML models and their application in catalyst material discovery, to address this knowledge gap. The first study proposes three new representations of adsorbate-catalyst systems by combining the distinct representations of the adsorbates and the catalysts. These representations can be used to predict the adsorption energies of intermediates in a complex reaction network to a higher degree of accuracy than the conventional linear scaling relations model. Notably, these representations overcome the problem of previous models, which can only be applied to one material or one adsorbate, thereby enabling 'cross-surface' and 'cross-adsorbate' prediction. Among the three representations investigated, elemental properties and spectral of London and Axilrod-Teller-Muto potential achieves a mean absolute error (MAE) of 0.19 eV for reactions related to C2 reforming. It is worth noting that two of the combined representations achieve their results using the empirical adsorbate structure only, yielding a significant reduction in computational cost, compared to full DFT approaches. The second study proposes group and period-based representations. The first study uses elemental properties to describe the catalysts, which are only applicable to pure metals. To predict adsorption on more complex materials such as alloys, representations such as coordination atom fingerprints should be used. However, the atom identifier (atomic number) used in these representations could not reserve their group and period information that relates to their catalytic activity. The proposed group and period-based representations, which replace atomic numbers with group and period numbers, achieve improved accuracy relative to the original atomic number–based representations. Notably, the group and period-based labelled site crystal graph achieves an MAE value of 0.05 eV for weak-binding adsorbates and 0.10 eV for strong-binding adsorbates - the lowest values achieved on bimetallic alloy surfaces to date. These group and period-based representations also prove robust in out-of-distribution prediction (predict adsorption energies on surface facets, elements, and alloys not included in the initial training set). While studies one and two relied on feature engineering to design the representations for the ML models, the third study takes a different approach. After an investigation of current ML algorithms, the third study develops an algorithm - graph neural network (GNN) with the local environment pooling (LEPool) method - to predict adsorption energies. Previously, crystal graphs and GNNs have been used to find novel catalysts by predicting adsorption energies on different catalysts. These GNNs use global pooling such as mean and sum pooling to obtain a prediction value. However, adsorption energy is a local property determined by the slab atoms near the adsorbate. The LEPool gives node importance scores (based on distance to the adsorbate) and readout the prediction value only from those important nodes. A Top-K pooling is also designed, and this allows the GNN to learn the importance score by itself. The LEPool obtains MAE values of 0.07 eV and 0.10 eV, for H and CO adsorption energies, respectively. These values are lower than those obtained by global pooling and Top-K pooling. The first three studies investigated ML models by comparing their accuracy using different data sets. The fourth study investigates the application of ML models by predicting low-cost platinum alloys. Platinum is a high-performing catalyst for biomass conversion and water splitting, but as a noble metal, it has limited reserves and a high price. This study uses the labelled site crystal graph and GNN to find alternative catalyst materials to Pt(111). Active learning, whereby an ML model improves itself iteratively by predicting and querying new reference data points, is utilised to overcome the difficulty of relatively limited training samples. This study identifies 12 binary and ternary catalysts that have similar catalytic performance (via catalytic descriptors - adsorption energies of H, CO and O) with Pt(111).
- Subject
- machine learning; catalysis; materials discovery; computational chemistry; thesis by publication
- Identifier
- http://hdl.handle.net/1959.13/1509125
- Identifier
- uon:56211
- Rights
- Copyright 2022 Xinyu Li
- Language
- eng
- Full Text
- Hits: 81
- Visitors: 85
- Downloads: 7
| Thumbnail | File | Description | Size | Format | |||
|---|---|---|---|---|---|---|---|
| View Details Download | ATTACHMENT01 | Thesis | 8 MB | Adobe Acrobat PDF | View Details Download | ||
| View Details Download | ATTACHMENT02 | Abstract | 374 KB | Adobe Acrobat PDF | View Details Download |