UC Irvine

Electron microscopy provides direct access to nanoscale structure, but quantitative interpretation is difficult when specimens are dynamic, weakly scattering, surrounded by liquid or vitreous ice, or sampled incompletely. This dissertation develops a traceable framework that combines liquid-phase transmission electron microscopy (LPTEM), cryogenic electron microscopy (cryo-EM), molecular dynamics, multislice image simulation, and deep learning. Across four studies, each conclusion is matched to directly measured image observables and to the reference information actually available.First, electrochemical LPTEM was used to examine FFssFF-derived peptide coacervates through one three-cycle cyclic-voltammetry (CV) recording before electrochemical impedance spectroscopy (EIS) and two three-cycle recordings after EIS in one liquid-cell field. Initial domains were approximately 100–500 nm in projected diameter. Recording A showed no prominent resolved CV peak and minimal morphological change under the combined electrochemical and imaging conditions. The largest qualitative transition occurred across EIS: coacervate boundaries appeared sharper, while numerous smaller, lower-contrast aggregates appeared in the surrounding field. The impedance response had a predominantly capacitive appearance and lacked a resolved charge-transfer semicircle over the measured frequency range. During the subsequent CV recordings, small aggregates progressively diminished and the same four tracked domains lost projected area. Median endpoint area changes were +0.98%, −7.41%, and −20.67% in Recordings A, B, and C, respectively. These observations establish a time-associated sequence of projected morphological change, but they do not isolate bias from electron exposure, identify a unique chemical mechanism, or equate projected contrast with mass or composition. Second, cryo-EM measurements were used to constrain atomistic CSH–CSSC fiber models, which were converted into multislice TEM images for comparison with experiment in a common image domain. The experimentally measured individual-fiber diameter was 5.42 ± 1.55 nm, whereas the defocus-weighted simulated diameter was 4.78 ± 0.66 nm. Systematic simulations showed that defocus and water thickness altered apparent fiber width and contrast. Because the experimental diameter also informed the radial restraint, this agreement is interpreted as an image-domain consistency test rather than independent structural validation. The forward-modeling strategy was then extended to generate 500 physics-informed synthetic TEM image–reference pairs for low-contrast nanofiber segmentation. A U-Net trained exclusively on these data achieved Dice 0.9547, intersection over union 0.9134, and centerline Dice 0.9969 on the 78-image synthetic test set. In unlabeled experimental TEM images, its predictions followed many visually recognizable fiber-like trajectories, providing qualitative correspondence but not a quantitative estimate of experimental accuracy. Finally, TSGNet was evaluated for generating intermediate projections in sparse simulated scanning transmission electron microscopy tilt series. Relative to linear interpolation, TSGNet improved the reported image-level mean-squared error, peak signal-to-noise ratio, and structural similarity. Insertion of the generated projections reduced both surface Chamfer distance and volumetric root-mean-square error relative to matched sparse baselines at every reported nominal tilt increment from 2° to 20° under both simulated configurations. These comparisons were made against a dense-series reconstruction reference and do not establish experimental dose reduction or acquisition-time savings. Collectively, the studies show how synchronized in situ imaging, physically grounded forward simulation, and learning-based inference can extend electron microscopy toward traceable quantitative analysis while preserving the distinction between direct observations, model-based references, synthetic-domain evaluation, and experimental validation.