High-resolution images often contain large amounts of data, which limits their ability to fit into the context windows of Large Language Models (LLMs) and Visual Question Answering (VQA) systems. To address this limitation, image compression is used to reduce file size and while this may reduce quality, the full-resolution image is often unnecessary. This is because models can process heavily compressed images with lower cost and maintained or improved performance. However, current methods are based on pixel-level redundancy, generic saliency or general semantic interests rather than what is useful for a given task. We investigate task-driven compression guided by personalized human semantic needs in a user study (n = 5). Participants marked semantically important regions for different images under the following conditions: without any task and with explicit two task prompts. Participants then evaluated how well visual saliency and attention maps generated by three different models (TranSalNet, DINO v1, and CLIP GradCAM) aligned with their selected regions. During our results we find that users all value different parts of images differently based on personal experiences and the given tasks. We identified the following key design principles for future image compression: usefulness is based on a user’s familiarity with the environment and personal experiences, usefulness is dependent on the desired task with the image, and usefulness is not always correlated with visual saliency. Our results also indicate that the CLIP GradCAM model is perceived as better and more representative.

Passive radiative cooling enables energy-free heat dissipation by emitting thermal radiation to cold outer space, but its static nature limits performance in environments with varying thermal demands. Active radiative cooling overcomes this constraint by dynamically tuning thermal emissivity. However, current approaches based on paints, polymer films, and photonic structures have limited tunability and are generally opaque in the visible spectrum. We present a micro-electro-mechanical systems (MEMS) micro-mirror array for active radiative cooling with real-time emissivity control. The device employs electrostatic actuation to switch between an open state, exposing a high-emissivity surface for cooling, and a closed state, where a reflective film suppresses thermal emission. A custom outdoor experimental platform was developed to evaluate thermal performance under diverse daylight conditions. Continuous temperature monitoring compared the actuated (closed, low-emissivity) and unactuated (open, high-emissivity) states, using ambient temperature as a baseline. Outdoor testing confirmed effective thermal modulation. In full and partial sunlight, the open state achieved up to a 5 °C reduction compared with the closed state. These results verify that MEMS-based emissivity switching can deliver significant, on-demand radiative cooling without additional energy input. This proof-of-concept demonstrates the potential for scalable, adaptive thermal control systems. Future applications include smart building facades, responsive textiles, advanced electronics cooling, and aerospace thermal management. The integration of MEMS-engineered emissivity control could enable a new generation of intelligent, energy-efficient climate regulation technologies.

Performing time-resolved spectroscopy at terahertz (THz) frequencies is desirable to be able to detect hazardous situations such as gas leaks from lithium ion battery fires. This remains challenging due to a lack of sophisticated THz sources (which often generate low output power) and detectors (which are often slow and unable to detect low power).    In self-mixing interferometry, the light emitted from a laser reflects off an external target and is reinjected back into the laser cavity, a process also known as optical feedback. Properties of the reinjected light, such as the level of attenuation, can be measured via the terminal voltage with a fast response time. Thus, self-mixing spectroscopy avoids the need for slow external detectors and operates independent of laser output power. We use a quantum cascade vertical-external-cavity surface-emitting laser (QC-VECSEL) as our source since it generates THz radiation, and the near gaussian beam pattern and large radiating aperture allow us to achieve a robust self-mixing signal.    In our work, we build a self-mixing setup and place a silicon etalon in the optical path. We are able to confirm the presence of optical feedback in our system and detect the presence of the etalon by observing changes in the laser voltage. Furthermore, we show that increasing the distance of the external mirror from the laser likely improves the spectral resolution of our setup. These results demonstrate the feasibility of using self-mixing in QC-VECSELs to perform time-resolved spectroscopy.

Viscoelastic materials have characteristics of both solid and fluid-like deformation, which means its stress strain relations are time dependent. They tend to be stiff under fast loading and very soft under slow loading. Under typical linear elastic fracture mechanics (LEFM), the relation between the critical stress, strain, and energy release rate at fracture is clear; however, due to the time dependency of viscoelastic materials and extra bulk dissipation, this is not as clear and cannot be solved analytically. Through this research we aim to see how bulk dissipation affects the fracture energy through measuring both post cut and pre cut crack propagation velocities – the former being with little bulk dissipation effects and the latter being more extreme. Through this, we find bulk dissipation to nonlinearly increase the toughness of the VHB, yet it takes away from the fracture energy. Further research is needed in different, more fluid-like materials, as well as FEM modelling to capture a clear image of the deformation around the crack tip.