A Theoretical Study Of Nanostructured Phonon Thermal Transport Based On Machine Learning And Molecular Simulation

With the development of nanotechnology and decreasing feature size of electronic components,a serious challenge was posed to the thermal management of nanodevices.On the other hand,owing to microstructure and properties,nanomaterials have unique advantages in improving energy conversion efficiency and heat transfer performance.To dealing these challenges,it is hoped to develop accurate predictive models for the properties of these nanomaterials,especially thermal conductivity and thermal interfacial resistance.At present,these properties of materials can be measured by experimental methods on the one hand,but this has high requirements for experimental instruments and the like.On the other hand,many simulation methods were developed to predict various properties of materials,such as first-principles calculations,molecular dynamics simulations,and Monte Carlo simulations,but these methods still suffer from large amounts of computation.Recently,machine learning technology has been widely used in various fields,and has been favored by researchers because of its ability to fit high-dimensional and nonlinear functions.Combining machine learning methods with traditional simulation methods points out a feasible method for improving computational efficiency.In this paper,the interfacial thermal resistance of superlattice materials and low-dimensional materials,and the physicochemical properties of superlattice materials and low-dimensional materials are studied by combining machine learning and simulation methods.The research achieved the following results:（1）Since the thermal interfacial conductivity has become an important factor that cannot be ignored in the thermal management of electronic components,simulation and experiment have gradually become the main methods to study the interfacial thermal conductivity.And all interfacial thermal conductivities,at the microscopic level,have atomic-level contacts.Therefore,it is of great significance to study the interfacial thermal conductivity of point and surface contact.In this paper,the non-equilibrium molecular dynamics method is used to calculate the interfacial thermal conductivity of the point contact model of silicon block-silicon sphere-silicon block fixed at both ends.By changing the size of the silicon sphere in the model and the pressure exerted on the outside of the silicon block,The effects of the two on the thermal conductivity of the interface are simulated.The simulation results show that both are linear to the change of interfacial thermal conductance.（2）Superlattice materials have great potential in thermoelectric materials due to their excellent properties.However,it is still a challenge to design superlattice materials optimizing their thermal conductivity.In this paper,by simplifying the superlattice material into a one-dimensional atomic chain model,and using the Landau formula to calculate the thermal conductivity corresponding to different one-dimensional atomic chain structures,the dataset required for machine learning is obtained.The dataset was used to train the BP neural network.After that,the research on the optimization of the one-dimensional atomic chain structure of a specific length was completed.The optimization of the one-dimensional atomic chain structure helps to obtain superlattice materials with better properties.（3）Two-dimensional materials often have unique properties,but due to the high requirements of experiments on equipment and environment,simulation has gradually become a commonly used method.First-principles calculations in simulations have the highest reliability,but are computationally expensive and time-consuming.Machine learning can improve computational efficiency by orders of magnitude while ensuring accuracy close to first-principles calculations.In this study,the first-principles software VASP was used to calculate the interatomic interactions of molybdenum disulfide,and extract information such as interatomic forces as a data set for machine learning.A single-layer Mo S₂ machine learning potential function is trained using the GAP model.Finally,through molecular dynamics calculations,the reliability of the potential function in simulating low-dimensional materials is verified.