Accurate forecasting of separation performance in seawater reverse osmosis (SWRO) systems is critical for minimizing operating costs and enforcing water quality standards. While most research focuses on optimizing industrial systems, this project targets portable SWRO units vital to long-term military operations and disaster relief. We propose a physics-informed long short-term memory (LSTM-PINN) framework to model membrane degradation as a function of salt rejection decay. Three architectures were evaluated: a data-driven LSTM, a soft-constraint physics-informed loss function, and a hard-constraint differentiable physics layer. Our operational data includes water quality parameters such as pH, electrical conductivity, oxidation-reduction potential, flow, and temperature. Embedding the solution-diffusion transport equations allows the model to capture additional physical parameters, including water permeability, salt permeability, and flux. Our findings suggest that purely data-driven and soft-constraint models over-extrapolate the initial drop in salt rejection per cycle during the transient conditioning phase, triggering premature estimates of membrane failure. In contrast, the hard-constraint physics model offers superior prognostic performance by identifying outliers relative to the theoretical rejection trend and effectively capturing the gradual decay in degradation. The integration of this model into a portable SWRO unit enables real-time estimates of the membrane’s remaining useful life and water quality monitoring.

The foundations of quantum physics is one of the most enigmatic puzzles, upon which depends our philosophical understanding of physical reality. The paper presents the mathematical and conceptual fundamentals regarding the measurement problem. The plethora of different interpretations of the formalism is summarized. In particular, the distinction between 𝞧-ontic and 𝞧-epistemic theories are presented. The former sees the quantum state as referring to physical entities; the latter sees the quantum state as referring to information of the physical entity doing the measuring. The paper aims to elucidate the theoretical adequacy of the QBist (quantum bayesian) view: quantum states are expressions of degrees of personal belief and the Born rule is a normative consistency condition on those beliefs. To learn the material, I found many lectures online from leading researchers all around the world; furthermore, at UCLA, I took multiple physics classes and finishing up a math minor. Additionally, a special acknowledgement is due to the many papers and books that contributed to my thinking.

California has mandated the transition from compressed natural gas (CNG) to battery electric buses (BEBs) to eliminate transit greenhouse gas emissions. While BEBs offer significant environmental benefits, operational efficiency varies based on route and terrain. We evaluated the performance of the Santa Monica Big Blue Bus (BBB) fleet, comparing new generation BEBs against the previous BEB fleet. We consolidated two datasets: Swiftly, which contains real-time route information, and Chargepoint, which contains energy consumption metrics. Using Python-based linear models, we analyzed select performance metrics. The models quantify the relationships between the distance driven, change in state of charge, and energy used. This study will identify efficiency differences across 17 routes while controlling for confounding variables. Findings will optimize fleet distribution, ensuring that the most efficient buses are assigned to the most energy-demanding time blocks. This work will help prepare Santa Monica BBB for high traffic events, including the 2028 Los Angeles Olympics.

Longitudinal count data are often irregular, with missing observations, gaps, and varying assessment frequencies, complicating modeling under standard assumptions and motivating more flexible approaches. This project developed a reproducible preprocessing framework that transforms raw, irregular time series of activity count data into structured sequences suitable for advanced modeling, using professional basketball as an example. The framework includes: (1) structuring longitudinal activity counts into fixed-length sequences based on domain-specific meaning, (2) normalizing within individuals to account for baseline differences, (3) engineering temporal features including exposure and normalized measures, and (4) applying bootstrap-inspired sequence augmentation to increase effective sample size. An LSTM model with padded sequences was trained to capture temporal dependencies and serve as a vehicle for evaluating the framework. The dataset included 85 players, with an average of 19 complete and 21 incomplete games per player and substantial variability in data availability across individuals. Player-level longitudinal points per game served as the prediction target. The proposed framework was applied, and LSTM performance was evaluated accordingly. This work highlights data structuring as a critical and often overlooked step, and demonstrates that data augmentation can bridge raw, irregular high-frequency data and downstream modeling.

Covariance estimation is a fundamental component to many psychological models, including regression, factor analysis, and structural equation modeling more broadly. However, the estimation of covariances is often complicated by the presence of missing data, which are always prevalent with behavioral measures. Missing data can typically be handled by through auxiliary models, which either weight with a response (or missingness) probability model or by imputing data with a prediction model. Both these approaches assume that each of their respective models are correctly specified. We investigate an augmented inverse probability weighted estimator (AIPW) which combines these two methods. AIPW only requires at least one among the response model and prediction model to be correctly specified to yield unbiased estimates in theory, achieving a so called doubly-robust property. We investigate this property and compare AIPW to other commonly used methods of handling missing data in covariance structure models. Our results indicate that AIPW outperforms the current practice when the prediction model is misspecified, while the current standard methods show significant bias. This demonstrates an alternative modeling approach for applied settings where auxiliary models cannot be specified with certainty.

The Roman Space Telescope Galactic Bulge Time Domain Survey will monitor approximately 100 million stars toward the Galactic center, producing on the order of 20,000 variable candidates per day. Identifying planetary microlensing signals at this rate requires classification and parameter inference faster than traditional forward-modeling permits. Existing training sets often simplify the detector physics of Roman’s H4RG-10 sensors and the noise realism needed for transfer to flight data. We present SMIG, a synthetic microlensing image generator that couples a GPU-accelerated finite-source engine (microlux) to a physical model of Roman’s H4RG-10 detector. The detector model includes MultiAccum reads, saturation, 1/f noise, and persistence. The pilot run targets magnification agreement with VBMicrolensing benchmarks at fractional error below 10^-4, and evaluates synthetic detector outputs against H4RG-10 calibration data for persistence decay and saturation behavior. The pipeline generates a pilot catalog of approximately 100,000 events spanning single-lens, binary-lens, binary-source, and contaminant classes, with a path toward scaling to a two-million-event training catalog. Downstream ML stages for domain adaptation and real-time inference are specified and will be trained once the catalog and post-commissioning Roman data are available, enabling autonomous exoplanet discovery for Roman’s first observing season.

Shear Alfvén waves play a central role in energy transport in magnetized plasmas, but resolving their full structure experimentally is challenging due to limited measurements.  Physics-informed neural networks (PINNs) can help address these limitations by providing robust reconstruction of the plasma state with only sparse measurements, but the training process can be computationally expensive for large 3D datasets.  This work develops a quasi-3D axisymmetric reduced model of the magnetic field structure from data on the Large Plasma Device to enable direct comparison with particle-in-cell simulations used to train PINNs. A processing pipeline performs noise characterization, unit normalization, interpolation to cylindrical coordinates, and azimuthal Fourier decomposition, yielding a compact and physically interpretable representation of the field. Results show that low-order modes dominate, with the m=1 mode consistent with the expected dipole structure and five modes capturing over 95% of the total wave energy. This approach enables quantitative comparison between experiment and simulation and provides a basis for reconstructing unmeasured fields from sparse data.

Understanding and quantifying the carbon footprint of healthcare systems is essential for driving sustainable policies and decarbonization strategies. This study estimates total CO₂ emissions from procurement activities at UCSF Health using a combination of Life Cycle Assessment (LCA) data and statistical inference. Scope 3 emissions—arising from the manufacturing, use, and disposal of goods—are substantial but difficult to measure at scale. UCSF Health procures approximately 25,000 distinct items, and while metadata such as item name, manufacturer, price, and quantity are available, emissions data require labor-intensive, item-by-item LCA, making full coverage infeasible. Existing approaches that rely on industry-average emissions per unit cost often produce inaccurate product-level estimates and do not quantify uncertainty, leading to potential misallocation of decarbonization efforts. To address these limitations, we propose a hybrid framework that integrates LCA data with data-driven techniques, including TF-IDF–based grouping to consolidate similar items and propensity score reweighting to approximate a probability proportional to size (PPS) sampling design. Using this approach, we produce improved estimates of total procurement-related emissions along with 95% confidence intervals, offering a more robust and scalable alternative to traditional methods. This framework reduces estimation bias and provides a generalizable solution for large-scale procurement systems.

Mesh simplification is a fundamental operation in computer graphics, yet classical greedy algorithms like Quadric Error Metric (QEM)-based decimation follow fixed heuristics that cannot adapt to task-specific qual- ity objectives. In this work, we formulate 3D mesh simplification as a reinforcement learning problem, where a policy network learns to make per-face keep-or-remove decisions that jointly optimize geometric fidelity and reduction ratio. We model the mesh as a face adjacency graph and employ a Graph Attention Network (GAT) backbone within a Proximal Policy Optimization (PPO) framework. The agent observes per-face ge- ometric features—including area, aspect ratio, normals, and QEM er- ror—and outputs binary actions for each face, triggering edge collapse operations when removal is selected. A composite reward signal balances progress toward a target face count against the QEM error incurred per collapse, encouraging the agent to preserve geometric quality while meet- ing simplification targets. We describe the environment design, action space, reward shaping, and training procedure, and discuss the challenges of applying RL to combinatorial mesh operations.