Design And Synthesis Of Half-heusler Type Thermoelectric Materials Based On Machine Learning

The ever-increasing energy consumption and the consequent environmental pollution problems have attached widespread attention.Thermoelectric(TE)materials are qualified for thermal and electrical conversion and energy efficiency promotion.The half-Heusler(HH)type compound is a medium to high temperature TE material with both stability and high strength.And the TE device is expected to be employed in industrial waste heat collection.However,the enhancement in energy conversion efficiency is the bottleneck for TE materials,due to the cost of trial and error optimization in traditional experiments,as well as the computing power limitation in first principles calculations.In order to accelerate the research of HH TE materials,the existing literature and calculation data were studied by machine learning methods.The TE theory was understood in a mathematical perspect,and the TE performance of HH materials was optimized and explored at multiple scales.Based on the optimization strategy of isoelectronic alloying to suppress lattice thermal conductivity,machine learning models for predicting the TE transport parameters of n-type XNi Sn(X=Ti,Zr,Hf)compounds were designed.The composition range and parameter influence law for optimal TE performance were determined.The prior knowledge of TE was adopted for feature design.Active learning method elevated the prediction ability of the model through two iterations,where the prediction results and divergence of multiple algorithms were concerned for experiment.The regression scores(R2)of the test set and verification set is higher than 0.9.The optimal composition interval of TE figure of merit(ZT)>1.2 was determined and experimentally verified.The optimal sample in the iterative experiment obtained ZT of 1.3 at 873 K,and the influencing factors of transport parameters were summarized from the feature importance analysis.In order to determine the global optimal judgment conditions in the experiment optimization of HH materials,a ZT prediction model based on all HH data was established.The descriptor and range for obtaining the optimal ZT were searched and determined.Data segmentation enhances the understanding of tags in small data sets.The random features introduced by feature engineering unearthed potential factors affecting TE performance.The R2 score of the random forest algorithm in the test set was increased from 0.76 to 0.84.Based on the feature analysis,the tuning lower limits of Seebeck coefficient in n-type and p-type HH are defined as 190?V·K^-1 and 210?V·K^-1 respectively,and the effect of ground state volume and space group deviation range GSV1?8 and SGN1>30 for components regulation is discovered and confirmed.Discovering new types of TE materials is another research focus in exploring the promotion of TE performance.Combining the calculation results of the high-throughput computing database,multiple clustering models for classifying and screening potential TE materials have been established.TE-related compounds features and electronic structure features were designed.61 types of TE materials from 520 candidate samples calculated in the database were iteratively screened during 4 clustering.4 systems and a total of 20 potential TE materials were further determined.The energy and nature of the those samples in models'decision-making were surmmerized.Experiments have verified and confirmed the TE properties of Sc Ni Sb and Sc Co Te.The single-phase and intrinsic p-type Sc Ni Sb has a ZT of 0.47at 923 K after optimization.