National and Kapodistrian University of Athens
Computational methods for the characterization, modelling and reconstruction of complex heterogenous nanomaterials with emphasis on methods of dynamical systems and machine learning
Abstract
dc:descriptionThis thesis develops computational methods for two metrological problems in nanotechnology research concerning geometrical and chemical properties of nanostructured surfaces: a) reliable characterization of surface nanomorphologies with complex geometries and b) non-destructive thickness estimation of coatings on rough substrates used to change their chemistry. As characteristic dimensions shrink to tens of nanometers, interfacial phenomena dominate performance. However, standard roughness tools compress images to a few metrics based on amplitude distributions and correlations and can miss multiscale organization usually encountered in complex micro and nanostructured surfaces with high functionality. Also, traditional metrology methods of thin-film thickness assume lateral uniformity and near planarity of substrates; their accuracy drops at high roughness. The first contribution targeting the geometry problem is a pair of image-based complexity metrics grounded on entropy and chaos-based concepts and designed to complement the conventional approaches such as rms roughness, autocorrelation, height–height correlation, and power spectra. The multiscale Shannon-entropy metric quantifies distance from homogeneity by aggregating scale-wise departures from uniformity while the second proposal of the chaos-based complexity metric quantifies distance from noise-like randomness. The multiscale Shannon-entropy metric computes the entropy of local average-intensity fields over neighborhood sizes and aggregates across scale. To control amplitude sensitivity, three discretization schemes are analyzed: fixed-variable, which increases with amplitude range; variable-fixed, which largely removes amplitude effects and isolates spatial organization; and fixed-fixed, which mimics the projection of surface morphology to finite gray-level quantization and yields a characteristic non-monotonic trend due to clipping at high roughness. The second metric is chaos-inspired: iterations of a hyperbolic chaotic map (Arnold cat map) stretch and fold image structure, amplifying high frequencies and mixing surface pixels. The standard deviation of high-frequency spectral slopes across early iterates, acts as a scalar that is small for near-order and near-noise and maximal for mixed textures with enhanced complexity. On synthetic self-affine and process-driven surfaces, the entropy metric separates amplitude-driven from organization-driven behavior and maps the role of the maximum neighborhood scale. The chaos-based metric peaks at intermediate correlation length and increases with the roughness exponent at fixed correlation length. Applied to SEM images of chemically etched aluminum, the chaos-based complexity evolves non-monotonically with etch time and exhibits a maximum, consistent with the synthetic trends. Together, the metrics provide compact, interpretable summaries of multiscale heterogeneity. The second contribution is a non-destructive estimator of deposited thickness of coatings on rough substrates from top-down SEM pairs acquired before and after coating. For thickness small relative to the spacing of nearby features, geometric relationships give an estimate based on the increase in bright area normalized by the average of the pre- and post-coating perimeters. An overlap correction adjusts for feature coalescence after deposition that would otherwise inflate the estimate. Validation on synthetic shapes and binarized rough landscapes shows that the corrected form reduces bias and variance, with robustness improving as correlation length increases, while independent experimental SEM image pairs indicate agreement with external thickness indications. Two machine-learning modules complement the analytic route. A supervised regressor that ingests paired before–after images predicts an effective dilation calibrated to nanometers and reduces sensitivity to thresholding and overlap heuristics. A U-Net performs an inverse check by reconstructing a plausible pre-deposition image from a post-deposition image and a scalar thickness, providing internal consistency with the thickness-as-dilation assumption. Overall, this thesis combines information theory, dynamical systems, and geometrical concepts with targeted learning to convert SEM images into quantitative measures of nanosurface complexity and coating thickness on rough substrates. The entropy-based and chaos-based metrics add complexity-aware information that complements rms, correlation, and spectral descriptors. The thickness estimator provides non-destructive thickness values that agree with synthetic and experimental tests, while two ML modules improve robustness and enable an inverse consistency check. Limitations include spatially varying deposition while extensions include alternative chaotic maps and entropy families.
Degree
thesis:*- Grantor dc:publisher
- National and Kapodistrian University of Athens
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kondi, Aleksandros
- Κόντι, Αλέξανδρος
Subjects
dc:subject × 16Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/60646