A Study On The Thermal Transport Properties Of SnTe Using Deep Neural Networks Learning

In recent years,due to the full use of energy and the improvement of environmental awareness,thermoelectric materials have been paid more and more attention by the government and research institutions.Pb Te as a kind of classic thermoelectric materials for its excellent performance has been widely used,but the Pb Te itself contains toxic lead elements will bring pollution to the environment,so people began to look for alternative to lead telluride,Sn Te and Pb Te have the same chemical structure at room temperature,thus gradually entered the view of the researchers,but the intrinsic Sn Te containing Sn vacancies(10²⁰～10²¹cm^-3)led to the thermoelectric performance is poorer,this makes us to study in the micro-nano scales.The micro-nano-scale research methods include first-principles and molecular dynamics simulation.First principles calculation has high precision and slow speed;The molecular dynamics simulation method has low accuracy and high speed.And we don’t have an empirical field for tin telluride,which bogged us down.We think in machine learning method is introduced into our research,through the data from first principles,through the machine learning method to get the telluride tin machine learning potential function,finally to use machine learning potential function of molecular dynamics simulation process,not only can make up the vacancy of telluride tin field experience,and can make the calculation results are accurate.For this reason,we first applied this idea to the calculation of silicon.As a classical semiconductor,we first obtained the machine learning potential function of silicon and verified the accuracy of the machine learning potential function,so as to verify the feasibility of this method.We collect the primary principle of silicon in different temperature molecular dynamics data,data sets collected a total of 45089 frames,finally got its potential function of machine learning,we verify the energy and force does the predicted results,the system crashes,phonon dispersion curve fitting and thermal conductivity fitting,after the results of machine learning potential function can reproduce the first principles calculation results,machine learning potential function of thermal conductivity is 140.3 W/m K,than primary principle results compared to 145.3 W/m K error less than 5%.Finally,we used the same idea to get the machine learning potential function of Sn Te,and collected data set of 7922 frames.The energy and force prediction results,system collapse,phonon dispersion curve fitting and thermal conductivity fitting were also verified.The thermal conductivity of the machine learning potential function was 4.825W/m K,while the first-principles result was 5.151W/m K.