Development Of Material Esign Software Based On Machine Learning Algorithm And Application In The Perovskite Material

The data-driven research paradigm has been widely applied in the field of material,which is benefited from the rapid development of artificial intelligence algorithms and the increasing abundance of data by theoretical computational and experimental characterization methods for material.The new research paradigm has promoted material research toward material informatics.Traditionally,the design and characterization of new material is usually performed by a combination of theoretical simulation and experiment,but it often needs a long experimental development cycle and a limited search of the material space.Material informatics methods can shorten the material development cycle and expand the searchable material space by constructing potential mapping relationship between material structure and properties through machine learning(ML)algorithm.Data mining and statistical learning methods have also become particularly important in analyzing material datasets and building machine learning models.Therefore,in this paper,we develop a ML data-driven program for material systems to analyze and mine the relationship between material structure and properties.Meanwhile,a new material ML framework is proposed to improve the ML fitting performance based on the problem of quantitative structural property relationship(QSPR)data set limitation.The following research advances were achieved:1.Developing ML data-driven material design method and program.The program is designed based on the supervised learning framework in the current version,covering material data pre-processing,material feature engineering,ML model construction and evaluation,and material data visualization function.The program was written by Python language and integrates and standardizes the use of various ML algorithms in Scikit-learn,XGBoost,and Light GBM software.The feature derivation algorithm was developed to allow the users to design material descriptors based on the physical/chemical intuition and dimensional constraint.Currently,this program was integrated into our group’s self-developed artificial-intelligenceaided material design software JAMIP(Jilin Artificial-intelligence aided Materials-design Integrated Package).To verify the effectiveness of our program,we applied it to construct hybrid organic-inorganic perovskite bandgap model and explore the factor affecting the bandgap.And differential assessment of the effectiveness of multiple ML models in the study of halide perovskite materials.2.Proposing a material ML framework that combined material "data increment" by pressurization and feature engineering oriented by feature volatility score.The QSPR research often restricts the prediction effect of ML model due to the amount of material data of specific system.We expand material dataset by pressurization,while the feature engineering method with feature volatility score as an indicator mines descriptors suitable for describing structural systems under different pressures simultaneously.We use mixed halogen perovskite as an example for validating the effectiveness of our framework,formation energy and bandgap were selected as the target properties.In the end,we not only improved the ML model prediction accuracy but also reduced the overfitting of the model.This work provides an effective method to solve the restriction of QSPR research by limited material dataset.