Research On Catalysts Screening Mechanism Based On Semantic Representation

Material science is the premise of modern industry.If we cannot make a breakthrough within the field of materials,the improvement in other innovations will be limited.As a significant area in material science,catalyst design has been attracted by many fields.However,the traditional catalyst screening strategies are time consuming and expensive.With the development of material science,numerous countries have started to effectively enhance,and always created intrigue science and innovation,and searched for modern strategies and thoughts to handle issues of catalyst screening.Recently,the data-driven methods associated with Artificial Intelligence(AI)in the various applications have received much more attention and make a huge progress.From related work,I found that semantic representation in NLP of AI and catalyst representation are interlinked.In semantic representation area,diversity,variability and ambiguity of language are the main issues which determine the difficulties of scene description;while in the representation of catalyst which also meets the following issues: complexity of catalyst structure,diversity of atomic properties and variability of reaction process.Natural language processing,as a significant field of AI,has make a great progress relying on the methodology.It is a challenge but also an opportunity for catalyst screening to share the wisdom of NLP.In this thesis,I mainly introduce several representation strategies from the NLP point of view and draw some conclusions.After that,I analyze the characteristics of 2D catalysts and 3D catalysts.To tackle the less labeled information of 3D catalysts,I propose a strategy to design 3D catalysts from the knowledge extracted from 2D catalysts.The achievements of this thesis are listed as follows:1.A Graph-Transformer based molecular representation model of materials is proposed,which can learn the nonlinear relationship between the structure(including geometric structure and atomic properties)and properties of catalysts and represent the implied information of catalytic materials.2.A Knowledge Distillation model is used to learn properties from 2D catalysts to tackle the problem of few-shot learning caused by more screening objects but less labeling information in 3D catalysts.3.In the practical application scenario,the combined model is used to predict the adsorption energy of three-dimensional catalyst.The experimental results show the effectiveness of the model.