High-resolution XRF scanning of soft lake cores can overinterpret minerogenic noise without biological context. We align elemental profiles with quantitative diatom assemblage zones in Andean alpine cores to separate catchment-driven metal pulses from productivity-driven organic dilution. The coupled interpretation revises inferred twentieth-century lead peaks relative to element-only models.

Clients at the network edge differ in compute capability and remaining battery, which skews model updates toward well-provisioned devices. We formulate participation as a constrained optimization problem with per-round energy budgets and a variance penalty on client contributions. Simulations on non-IID vision partitions show improved worst-class accuracy relative to FedAvg without large increases in wall-clock training time.

Temperature-sensitive alleles are useful genetic tools yet their structural basis is often inferred indirectly. We combine microsecond-scale molecular dynamics with hydrogen-exchange proxies for a set of GroEL surface substitutions that shift the permissive temperature range. Electrostatic and loop-flexibility metrics correlate with experimental growth phenotypes better than static contact counts alone.

Clinical-grade caps constrain field studies of cocktail-party listening. We evaluate a compact transformer encoder on 16-channel dry-electrode recordings from a two-speaker selective attention task. With subject-specific fine-tuning, decoding accuracy approaches that reported for wet systems on matched trials, suggesting feasibility for portable neurofeedback prototypes.

Full-network Kalman updates become expensive as smart meter density grows. We approximate the admittance Laplacian with a spectrally similar sparse graph using effective-resistance sampling, then run a reduced-order estimator on the sparsified model. Case studies on medium-voltage feeders show bounded error in voltage estimates while cutting average update time substantially compared to the dense formulation.

Soft modular end-effectors tolerate uncertainty but are difficult to tune across payloads. We combine a high-level task planner with a low-level policy trained from simulated contact statistics gathered over randomized object sets. Hardware trials on irregular produce and glassware show fewer slip events and shorter settling times than a flat policy trained on the same data budget.

Annotating every frame in long microscopy runs is costly, so many labs rely on sparse phase labels. We train a contrastive encoder on unlabeled trajectories and align its embeddings to a small set of weak phase tags using a temperature-scaled loss. On three public datasets, semi-supervised accuracy improves over a purely supervised baseline while reducing labeled frame requirements by roughly half.