Edit | Radish peel
Despite recent advances in data acquisition and algorithm development, the application of machine learning (ML) in real-world catalyst designs faces significant challenges, largely due to its limited ubiquity and poor interpretability.
Researchers at virginia tech have developed a fusion-theory neural network (TinNet) method that combines deep learning algorithms with well-established d-band chemosorption theory for reactive prediction of transition metal surfaces.
Using simple adsorbents from active site sets (e.g., OH, O, and *N) as representative descriptor species, the team demonstrated that TinNet is comparable to purely data-driven ML methods in predicting performance while being inherently interpretable.
Incorporating scientific knowledge of physical interactions into learning from data further elucidates the properties of chemical bonds and opens up new avenues for ML to discover new motifs with the desired catalytic properties.
The study, titled "Infusing theory into deep learning for interpretable reactivity prediction," was published in Nature Communications on September 6, 2021.
The adsorption energy of simple molecules or their fragments on solid surfaces is often used as a reactivity descriptor in heterogeneous catalysis. Because of the computational cost of accurately solving the multi-electron Schrödinger equation, it is attractive to quickly discover structural motifs with kinetic favorable descriptor values (e.g., using quantum chemistry calculations) but still a daunting task.
In this regard, the theory of D-band chemical adsorption pioneered by Hammer and N rskov has been widely used to understand the reaction trends of D-zone metals and their compounds. However, due to the perturbative nature and large variations of the theoretical framework, its quantitative prediction accuracy using individual d-band features such as the number of d-band electrons, the center of the d-band, and the edge of the d-band is limited. Site characteristics in high-throughput catalyst screening.
In recent years, machine learning (ML) has become an alternative method to predict the chemical reactivity of catalytic sites with hand-crafted or algorithm-derived features. By learning the associated interactions of atoms, ions, or molecules with substrates from a sufficient amount of ab initio data, it is possible to calculate adsorption characteristic orders of magnitude faster than traditional practice and narrow down the range of candidate materials before experimental testing.
A major limitation of black-box ML models, especially for re-emerging deep learning algorithms, is that it's easy to learn some correlations that look good on both training and test samples, but don't generalize well outside of labeled data. To alleviate this problem, active learning workflows, led by key performance indicators or model uncertainty, have been used to accelerate the exploration of the enormous, essentially infinite size of the accessible design space.
However, the large data samples required for model development and the difficulty of interpreting model predictions posed a huge challenge to its automated search for high-performance catalytic materials.
Illustration: Schematic of a neural network (TinNet) injected into the theory. (Source: Thesis)
Here, the researchers propose a fusion-theory neural network (TinNet) method to predict the chemical reactivity of transition metal surfaces; more importantly, extract physical insights into the properties of chemical bonds that can be translated into catalyst design strategies. Incorporating scientific knowledge of physical interactions into data-driven machine learning methods is an emerging area of research in catalytic science.
Currently, no such alternative model of chemical adsorption mixing has been developed in a fully integrated ML framework that is fairly accurate (~0.1-0.2 eV error) and can be transferred between samples. By using deep learning algorithms such as convolutional neural networks to learn adsorption properties from scratch, while respecting the accepted d-band theory of chemical adsorption in structural design, TinNet can be applied to a wide range of physical aspects of the interaction of d-block metal loci and naturally coded bonds.
Illustration: Model development. (Source: Thesis)
The researchers demonstrated methods for using adsorbed hydroxyl (*OH) as a representative descriptor species on end-terminated intermetallic compounds and near-surface alloys, such as looking for an effective electrocatalyst for metal catalytic O2 reduction, CO2 reduction, and H2 oxidation in alkaline electrolytes.
Illustration: Out-of-sample validation of a TinNet model. (Source: Thesis)
The frame can be applied directly to other adsorbents (e.g., *O) or to the collection of active sites of multiple bonded atoms, as shown in *N adsorption on the surface of the end metal. TinNet not only achieves predictive performance comparable to pure regression-based ML methods, especially for out-of-sample systems with unseen structures and electronic features, but also enables physical interpretation, paving the way for ML discovery of new motifs with the desired catalytic properties.
Illustration: Physical insights into chemical bonding. (Source: Thesis)
"Most machine learning models developed for material property prediction or classification are often considered 'black boxes' that provide only limited physical insights." Researcher Hemanth Pillai said.
"The TinNet method expands its predictive and interpretive capabilities, both of which are critical in catalyst design." Siwen Wang said he was also the leader of the study.
Illustration: TinNet model of other adsorbs/surfaces. (Source: Thesis)
As a hybrid approach, TinNet combines advanced catalysis theory with artificial intelligence to help researchers gain insight into this "black box" of materials design to understand what is happening and why, and can help researchers break new ground in many fields.
"Hopefully we can make this approach universally available to the community, where others can use the technology and really further develop renewable energy and decarbonization technologies that are vital to society." Xin, the team's head, said, "I think this is really a key technology that can make some breakthroughs."
"I really enjoy seeing different aspects of chemical engineering outside of the classroom." Researcher Athawale said, "It has a lot of apps, you know, it could be really revolutionary. So it's great to be a part of that."
Related: https://techxplore.com/news/2021-11-artificial-intelligence-advance-energy-technologies.html
Artificial Intelligence × [ Biological Neuroscience Mathematics Physics Materials ]
"ScienceAI" focuses on the intersection and integration of artificial intelligence with other cutting-edge technologies and basic sciences.
Welcome to follow the stars and click Likes and Likes and Are Watching in the bottom right corner.