Skip to content

Multiscale investigation of thermal transport in β-Ga2O3-based heterointerfaces enabled by machine learning potential - cross-scale parameter

FieldValue
Journalnpj Computational Materials
Published2026-02-18
AuthorsZhanpeng Sun, Zijun Qi, Yunfei Song, Lijie Li, Sheng Liu
DOI10.1038/s41524-026-02007-y
Citations2

A β-Ga2O3(010)/diamond heterostructure reaches a simulated maximum power density of 26 W/mm under the model’s 500 K damage limit. The authors report that this is approximately 226% of the capability of the β-Ga2O3/Si device. Diamond gave the lowest peak temperature and thermal stress for a fixed β-Ga2O3 orientation. Its interface nevertheless had the highest thermal boundary resistance and became the dominant bottleneck as substrate conductivity and operating temperature increased.

The finite element model set 500 K as the device damage limit and calculated the maximum allowable power density. The resulting allowable power densities are listed below.

Heat-spreading substrateMaximum allowable power density at 500 K
Siapproximately 10 W/mm
SiC16 to 20 W/mm
Diamond18 to 26 W/mm

The ranges span the four β-Ga2O3 orientations. At the 500 K limit, the diamond orientation difference reached 8 W/mm, corresponding to about a 44% increase. The (010) case gave the highest value. At 300 K, the model assigned single-crystal diamond 1185 W·m⁻¹·K⁻¹ in-plane and 1480 W·m⁻¹·K⁻¹ through-plane. The model used 150 for Si and 420 for 4H-SiC at 300 K. This conductivity ranking explains why diamond produced the lowest peak temperature at equal power and orientation. The need for a heat-spreading layer also follows from β-Ga2O3’s room-temperature average conductivity of about 15 W·m⁻¹·K⁻¹.

The framework transfers DFT-derived atomic quantities into mesoscopic molecular dynamics, then into finite elements. The training set covered bulk β-Ga2O3, Si, 4H-SiC, and diamond plus their interfaces. VASP AIMD used projector augmented waves, the PBE functional, and a 700 eV cutoff. The authors sampled 2400 training structures and 1200 test structures over 10 K to 1000 K. Atomic displacements and strains broadened the sampled environments, while interface energy tests set the starting separations. After 1,000,000 iterations, the neuroevolution potential supplied energies and forces for NEMD. Each NEMD model exceeded 5 nm by 5 nm by 30 nm and directed heat from β-Ga2O3 to the substrate. Thermostats maintained a 50 K source-to-sink difference. Thermostat pairs were 175 K and 225 K, 275 K and 325 K, 375 K and 425 K, and 475 K and 525 K. The 300 K analysis used 275 K and 325 K. Each run used 0.5 ns to establish the gradient and 2 ns for production, sampled every 50 ps. The protocol was repeated 80 times, and 50 consecutive samples were averaged. Fitted TBR over 200 K to 500 K and orientation-dependent β-Ga2O3 conductivity then entered the finite element model.

The NEP reproduced DFT phonon dispersions and radial distribution functions for the four materials. Its 300 K thermal conductivities also agreed with published values for diamond, Si, 4H-SiC, and β-Ga2O3. The table lists root mean square errors.

Predicted quantityTraining RMSETest RMSE
Energy4.9 meV/atom4.8 meV/atom
Forces229.0 meV/Å212.2 meV/Å
Virial stress37.2 meV/atom34.4 meV/atom

The test error was equal to or lower than the training error for every output. The authors judged the force error acceptable for the subsequent transport calculations. The potential was then used for bulk, interface, and temperature-dependent molecular dynamics calculations.

At a fixed β-Ga2O3 orientation, TBR was lowest with Si, intermediate with SiC, and highest with diamond. The 300 K comparison used 275 K and 325 K thermostats. For a fixed substrate, (010) and (001) interfaces had lower TBR than (100) and (201). The authors associate this pattern with anisotropic conductivity in β-Ga2O3. At 300 K, the fitted conductivity values were 20.07 and 13.18 W·m⁻¹·K⁻¹ for (010) and (001). They were 10.85 and 10.13 W·m⁻¹·K⁻¹ for (100) and (201). TBR decreased from 200 K to 500 K for every interface structure. This differs from prior studies that treated TBR as constant. The finite element model therefore used temperature-dependent TBR fits instead of fixed interface resistance.

The vibrational density of states, or VDOS, overlap ratio was 61.65% for β-Ga2O3/Si. It was 50.03% for β-Ga2O3/SiC and 21.58% for β-Ga2O3/diamond. The authors propose that greater spectral overlap supports elastic phonon transmission and partly explains the resistance ranking. Modes from 0 to 10 THz dominated conductance for β-Ga2O3/Si. For SiC, the main peak shifted upward, and 10 to 20 THz modes contributed appreciably. For diamond, 25 to 40 THz modes also contributed, beyond β-Ga2O3’s intrinsic phonon range. Cumulative conductance reached its maximum before 22.5 THz for Si and SiC, but only above 25 THz for diamond. The authors suggest that the added high-frequency modes may come from diamond or inelastic scattering. As temperature rose, VDOS overlap increased for all three material pairings. Phonon participation ratio also increased at 10 to 20 THz for Si and SiC, and 10 to 30 THz for diamond. These observations offer a possible explanation for declining TBR, rather than direct experimental proof.

Finite element simulations used about 1.3 million hexahedral and prismatic elements. The heat sink bottom was fixed at 300 K, while the remaining boundaries were adiabatic. A 200 nm minimum mesh resolved the steep gradients near each gate finger. The modeled device was 670 μm by 530 μm, with a 240 μm long and 0.5 μm wide heat source. For a fixed β-Ga2O3 orientation, Si produced the highest von Mises stress at 2 W/mm, followed by SiC and diamond. The less conductive substrate retained more heat and developed higher thermal stress. Under a fixed substrate, β-Ga2O3(010) had the lowest surface temperature but the highest von Mises stress. Its high conductivity and relatively low TBR aided heat dissipation. The authors associate the high stress with lattice mismatch and coefficient of thermal expansion mismatch at that interface. A region-by-region decomposition showed substrate resistance falling as substrate conductivity rose. For β-Ga2O3(100), the β-Ga2O3 contribution increased, and the interface contribution rose moderately. For diamond devices, interface resistance changed strongly with power, making interface optimization the main remaining thermal constraint. High substrate conductivity and rising temperature therefore move the bottleneck from the substrate toward the interface. These results are computational predictions; the paper reports no fabricated-device power or stress measurements. The 500 K damage threshold is a model criterion rather than a measured failure point. The analysis does not test surface roughness or interface fabrication variability. The authors call for further optimization of interfacial transport and thermal stress management. Device selection should therefore consider substrate conductivity, temperature-dependent TBR, crystal orientation, and stress limits together.

The modeled β-Ga2O3 high-power heterointerface needs a diamond heat-spreading substrate, with bulk transport and interface resistance assessed separately. The 6CCVD diamond materials catalog lists single-crystal and polycrystalline MPCVD diamond plates and wafers up to 125 mm, with thickness from 0.1 to 500 micrometers. The engineering support link provides custom fabrication support for the selected substrate geometry.

View Original Abstract

The rising power density of advanced electronics demands improved thermal management, while traditional single-scale methods are unable to fully reveal the complex heat transfer mechanisms in heterostructures. This work establishes a multiscale simulation framework by constructing a machine learning potential, enabling accurate cross-scale parameter transfer from atomic to mesoscopic and then to macroscopic levels. Results show that the thermal boundary resistance (TBR) at the β-Ga 2 O 3 /diamond interface is higher than that at the β-Ga 2 O 3 /Si and β-Ga 2 O 3 /SiC interfaces, and that the TBR decreases with increasing temperature, which contradicts conventional understanding. Vibrational density of states and interface conductance modal analysis elucidate the underlying mechanisms. These mesoscale insights are incorporated into macroscopic simulations, showing the β-Ga 2 O 3 /diamond heterostructure’s peak power capability reaches 226% of that of β-Ga 2 O 3 /Si. Further analysis reveals that although the thermal conductivity of the heat-spreading substrate remains the dominant factor in overall thermal performance, the thermal bottleneck gradually shifts toward the interface as both substrate conductivity and operating temperature rise. Moreover, crystal orientation significantly influences thermal performance and thermal stress distribution, necessitating careful trade-offs. This study not only provides effective strategies for optimizing β-Ga 2 O 3 -based devices but also establishes a generalizable paradigm for cross-scale thermal management research in heterogeneous material systems.