超宽禁带半导体β-氧化镓（β-Ga2O3）因具有优异的电学性质已成为最具应用前景的下一代新型半导体材料。但同时，由于β-Ga2O3的本征热导率较低，当器件功率密度激增时，β-Ga2O3基器件将产生严峻的散热问题，严重制约其进一步发展。本文以β-Ga2O3基器件的散热问题为背景，通过发展具有量子力学精度的机器学习势函数，系统地研究了β-Ga2O3、无定形氧化镓（a-Ga2O3）以及氧化镓基界面的热输运性质和导热机理，并探讨了调控异质结界面热阻的方法。针对氧化镓体系开发了具有高精度、高迁移性的高斯近似势（GAP）。该势函数可广泛用于β-Ga2O3单晶、缺陷、无定形、晶界以及β-Ga2O3/碳化硅（4H-SiC）异质结等复杂体系的分子动力学（MD）模拟研究。通过与第一性原理计算结果对比证明了本文所发展的GAP势达到了量子力学的计算精度，且计算速度比第一性原理快4个量级以上。通过对β-Ga2O3单晶声子热输运性质的研究进一步证明了GAP势可以准确描述声子间的相互作用。基于GAP-MD模拟、非晶导热理论和实验揭示了a-Ga2O3的原子结构特征、导热性质及机理。通过三电极3ω-2ω法测得a-Ga2O3室温下的热导率约为1.25 W/mK。模拟结果与实验值吻合较好，证明了GAP势处理复杂体系的能力。研究表明振动模态间的耦合作用是影响a-Ga2O3导热性质的主要机制，其对热导率的贡献超过95%。基于对结构特征的分析，提出了具有可解释性的无定形材料的结构描述器，并建立了a-Ga2O3结构与热导率之间的映射关系。基于GAP-MD模拟和实验研究了β-Ga2O3晶界、β-Ga2O3/a-Ga2O3界面及β-Ga2O3/4H-SiC异质结的原子结构特征及热输运性质，并提出了MolNet-3D和GANN两种深度学习模型用于加速界面性质的预测。在晶界研究中，利用高通量计算发现了30余种稳定的特殊大角度晶界。在相界的模拟研究中发现β-Ga2O3/a-Ga2O3界面处形成了化学结构平滑过渡区，而β-Ga2O3/4H-SiC异质界面附近则形成了丰富的Si-O共价键，结果表明以上两种界面结构特征均有利于界面热量的传输。通过实验测得β-Ga2O3/4H-SiC异质结的界面热阻约为12.2 m2K/GW，与其他实验结果对比发现，引入合适的过渡层材料或减小无定形层厚度均可有效降低异质结的界面热阻。最后在a-Ga2O3/a-SiC异质结势能面的预测任务中，证实了MolNet-3D和GANN两个深度学习模型的可靠性和高效性。

β-phase gallium oxide （β-Ga2O3） is an emerging ultrawide bandgap semiconductor that is promising for applications of next-generation power electronics. High heat generation in β-Ga2O3-based devices can cause excessive temperature, resulting from the relatively low thermal conductivity of β-Ga2O3. Overheating has been identified as a major bottleneck to the development of β-Ga2O3-based devices. This work systematically studies the thermal transport properties of β-Ga2O3, amorphous gallium oxide （a-Ga2O3）, and gallium oxide-related interfaces. Moreover, this work explores the method to reduce the thermal boundary resistance of β-Ga2O3-based heterostructures.A high-accuracy and high-transferability Gaussian approximation potential （GAP） is developed for gallium oxide-related systems, including crystal, liquid, and amorphous bulk phases, as well as defects, grain boundaries, and β-Ga2O3/4H-SiC heterostructures. The GAP model exhibits remarkable accuracy in reproducing the ab initio potential energy surface. Molecular dynamics （MD） simulations with the GAP model, however, are approximately four orders of magnitude faster than ab initio calculations. The GAP model is then employed to predict the phonon properties of β-Ga2O3. It is demonstrated that the GAP model can well describe the lattice dynamics of β-Ga2O3.The atomic structure, thermal conductivity, and underlying mechanism of amorphous gallium oxide （a-Ga2O3） are revealed by GAP-MD simulations and experiments. The thermal conductivity of a-Ga2O3 at room temperature is measured to be 1.25 W/mK by the three-sensor 3ω-2ω method. The simulation results agree well with the experimental values, which shows the capability of the GAP model to tackle complex systems. It is elucidated that the coherence between vibrational modes dominates the thermal conductivity of a-Ga2O3, and its contribution to thermal conductivity is more than 95%. Based on the analysis regarding atomic structures, an interpretable descriptor for amorphous materials is proposed. Then the quantitative relationship between structures and thermal conductivities of a-Ga2O3 is established. The atomic structures and thermal transport properties of β-Ga2O3 grain boundaries, β-Ga2O3/a-Ga2O3 interfaces, and β-Ga2O3/4H-SiC heterostructures are investigated by GAP-MD simulations and experiments. Moreover, two deep learning models, i.e., MolNet-3D and GANN, are proposed for accelerated prediction of interface properties. In the study of β-Ga2O3 grain boundaries, more than 30 stable and ordered high-angle grain boundaries are discovered by high-throughput calculations. In the study of phase boundaries, it is found that smooth chemical structure transition zones are formed at the β-Ga2O3/a-Ga2O3 interface, while abundant Si-O covalent bonds are formed near the β-Ga2O3/4H-SiC interface. The results show that the above two structural characteristics are conducive to interfacial heat transport. The thermal boundary resistance of the β-Ga2O3/4H-SiC heterostructure is measured to be 12.2 m2K/GW by the three-sensor 3ω-2ω method. In comparison with other experimental results, it is found that introducing suitable interfacial transition materials or decreasing the thickness of the interfacial amorphous layer can effectively reduce the thermal boundary resistance of the β-Ga2O3/4H-SiC heterostructure. Finally, the reliability and efficiency of the two deep learning models, i.e., MolNet-3D and GANN, are verified in the prediction task of the potential energy surface of a-Ga2O3/a-SiC heterostructures.