This project focuses on instrument control for characterization of Quantum Cascade Lasers (QCLs) using MATLAB. QCLs are semiconductor lasers which emit in the Terahertz. These lasers can emit light between 1 and 6 THz, making up much of a region known as the “terahertz gap”. QCLs serve as critical components in applications such as astrophysics, astronomy, and biomedical imaging. In order to characterize a device, different instruments are used to analyze the performance of the laser. The characterization setup for this project employs a boxcar averager, pulse generator, lock-in amplifier and a pyroelectric detector to evaluate the lasers performance under both pulsed and continuous wave operation. A MATLAB script automates data acquisition by sweeping source voltage, reading the measured THz radiation from the device, collecting the voltage and current across the device and generating relevant figures in real time. The results provide a framework for understanding optimal laser performance. 

Optical imaging is limited by photon penetration depth in tissue, restricting imaging depth to ~1–2 mm. Ballistic photons retain structural information; successful imaging requires that ballistic photons dominate scattering. Scatter increases with penetration depth, reducing encoded spatial information. Fluorescence is a material property whereby excitation occurs at a specific wavelength and is absorbed, producing an excited energy state; emission of visible light returns the sample to ground state.  Fluorescent proteins are useful imaging tools due to their monochromatic emission and controllable localization, supported by established protein modeling and conjugation techniques. We propose a method for overcoming the ballistic limit to capture images in deep, scattering tissues. A time-variant magnetic field is applied to the sample during imaging with a standard wide-field epifluorescent microscope to encode spatial density of fluorescent proteins in a signal independent of scattering effects. This leverages the magneto-fluorescent response recently characterized in several common fluorescent proteins.  The magnetic field is generated with a custom control circuit using microcontroller-dictated pulse-width modulation and MOSFET power amplification, enabling independent control of multiple electromagnets. An RLC low-pass filter mitigates higher-order harmonics, producing a cleaner signal. Component selection accounted for 20 V DC supply and 1–2 A current, with aluminum heat sinks where needed. An Arduino IDE (C/C++) script allows user-specified waveform amplitude, DC offset, and frequency.  We observe a nonlinear decrease in fluorescence intensity proportional to the sinusoidal driving voltage of the applied magnetic field. This approach may enable the use of fluorescent proteins as contrast for deep-tissue functional imaging. 

Lithium-ion batteries are widely used in electric vehicles, smartphones, etc. However, most of the batteries in these products are at rest for a long period and are used for a small portion of their lifetimes. During the time they are not used, batteries will undergo degradation induced by calendar aging, the process of loss of capacity and power that occurs when the battery remains idle. As a crucial mechanism that drives the degradation of batteries, it is important to quantify the effect of calendar aging so we can optimize the battery management system. Battery degradation introduced by calendar aging is complex, and depends on several factors, such as state of charge (SOC) and storage temperature of batteries. As proved by previous studies, the conventional model that used to model and predict the calendar aging is not accurate for the long term. There are efforts to use machine learning to enhance the calendar aging prediction models such as semi-empirical models, but they do not have a good generalizability. In this study, we develop a data-driven recurrent neural network model for early prediction of calendar aging lifetime. The model is trained on a lithium-ion battery dataset with 142 distinct cell degradation trajectories, 3 different storage temperatures, and 12 different cell types. We investigated the interpolation and extrapolation ability of the model across different storage conditions, and found that it can accurately predict the battery degradation trajectory, but it still has limited ability to predict the end-of-life trend across different storage temperatures.

Soft structures provide a wealth of benefits over rigid structures for use in a variety of applications, especially soft robotics. However, in order to take advantage of these benefits for robotics or other applications, it is necessary to accurately control the position of soft structures as they move. Methods such as model predictive control and reinforcement learning, amongst others, have been explored for soft robotics control tasks, but can be complex to implement for highly nonlinear models or when training data is limited. To develop the control method shown here, we draw on a physical model and machine learning principles, specifically principles from the Discrete Elastic Rods (DER) model from Bergou et. al. and principles from the Physics Informed Neural Network (PINN) from Raissi et. al. These concepts are used to develop a method to manipulate a two-dimensional beam modeled as discrete nodes, using neural networks in which the DER model informs optimization. The final approach uses a separate neural network for each time step and assumes the absence of dynamics. This method is used to simulate the manipulation of an end of the beam, aiming to make the position of a node along it follow a target trajectory. The method qualitatively achieves this successfully for beams modeled as three nodes and five nodes for a specific movement shown here.