Super-Resolution Processing of FDTR Measurement Data for Defect Detection

The semiconductor industry is vital for technological advancement and economic development. The emergence of heterogeneous integrated electronics packaging is revolutionizing the field by enabling the seamless integration of diverse components, enhancing performance, and driving the next generation of smart technologies. Recent rapid advancements in semiconductor manufacturing technologies have necessitated the development of new nondestructive characterization techniques for quality assurance testing of heterogeneously integrated microelectronics packages. These have included scanning acoustic microscopy (SAM), scanning electron microscopy (SEM), high-resolution X-ray imaging, Terahetz tomography, time domain thermoreflectance (TDTR), and frequency domain thermoreflectance (FDTR). In this work, we focus on the postprocessing of FDTR measurement data for the purposes of defect detection. FDTR is a non-contact microscopy technique which has been used to measure bulk thermophysical properties. FDTR measurements involve locally heating a sample with a periodically modulated laser while a second probe beam monitors the surface reflectivity, which is correlated with the surface temperature. By varying the modulation frequency of the laser, measurements may be made at different depths into the sample. By rastering the pump and probe lasers across the sample surface, 3D maps of thermophysical properties at resolutions of 1μm may be generated. These maps are useful for identifying defects in heterogeneous integrated electronics. FDTR is capable of sub-micron resolution but high-resolution scans are slow, requiring a large number of independent measurements performed in sequence. Traditional postproccessing techniques for FDTR data examine each measurement independently. This limits the output resolution to the lateral resolution of the measurement equipment. For detecting small defects, O(1μm), many measurements are required and scan times are long. The scan time is inversely proportional to the square of the measurement resolution. Long scan times limit the applicability of the technique by restricting the measurement to either small specimens or lower lateral resolution. In this work, we present a super-resolution postprocessing technique which identifies defects with 8x the measurement resolution, decreasing scan times by 64x without sacrificing accuracy. To achieve this, we take advantage of the sensitivity of the FDTR measurement to laterally misaligned defects due to the pump laser spot size and lateral heat spreading. Leveraging convolutional neural networks and a novel synthetic data generation pipeline, we produce models capable of accurately identifying defects ranging from 1μm to 3μm with measurement resolutions as low as 8μm. The models are shown to be robust to measurement noise and reliable across the range of defect sizes considered.