In various engineering fields it is required to design material layouts with improved performance or tai- lored properties. Hereby, the structural integrity should not be compromised, such that material failure is avoided. Topology optimization enables the systematic design of structures presenting enhanced performance with respect to criteria, such as mass, volume, stiffness, or eigenfrequencies. In the majority of research works in topology optimization, material failure is addressed by constraining or minimizing the stress state in simplified analyses, in which the material is modeled as isotropic linear elastic. However, brittle and quasi-brittle materials, such as concrete and rock, progressively degrade on a micro-scale, before material rupture occurs macroscopically. Using continuum damage mechan- ics, the state of material degradation may be modeled through one or multiple internal variables that are thermodynamically irreversible. However, modeling damage leads to a computationally intensive, non-linear and path-dependent optimization problem. The limited previous research suggests, though, that accounting for material degradation throughout the design phase can significantly improve the per- formance and reliability of the resulting designs. In this work, a level set-based topology optimization approach is proposed to generate steel- reinforced concrete layouts with enhanced damage resistance. Additionally, a thorough comparison of addressing material failure in a linear stress-based and a non-linear damage-based TO approach is conducted. The geometry of the designs is represented implicitly through a level set function. The level set and state fields are discretized using the extended finite element method, which alleviates the need to remesh in each optimization iteration. Assuming isotropic damage, strain softening degradation in the concrete material is modeled through a single scalar damage variable. The steel as well as the un- damaged concrete are modeled as isotropic linear elastic materials. To avoid instability and localization of deformation, non-locality is introduced through a gradient-enhanced damage model. Boundary and interface conditions are imposed weakly using Nitsche’s method. Numerical instabilities due to small material subdomains are mitigated through face-oriented ghost stabilization. The optimization prob- lems are solved through mathematical programming using the globally convergent method of moving asymptotes. Accounting for the irreversibility of damage, the required shape-sensitivities are evaluated semi-analytically through an adjoint approach. The proposed optimization approach is applied to two-dimensional numerical problems from the literature. The results show that resorting to a simplified linear analysis is insufficient to generate struc- turally reliable designs of brittle material. When modeling the non-linear damaged material behavior throughout the optimization process, though, resulting layouts show an enhanced load-bearing capacity and resistance to damage. It is further demonstrated that the maximum damage level in steel-reinforced concrete layouts can be controlled, through an optimized distribution of reinforcing steel material.

Neuromorphic computing, a novel computing configuration inspired by the brain, aims to perform calculations based on physical neurons and synapses, attracting significant attention in recent years. Resistive random access memory (RRAM) shows great potential in this field, demonstrating high operation speed, nanoscale scalability, long retention time, non-volatile performance, and a simple structure.

Despite the promising performance of RRAM, a high forming voltage potentially hinders the widespread application of the device. This thesis aims to diminish and eliminate the forming voltage. To achieve this, different metals were inserted between the insulator layer and the bottom electrode of the RRAM, serving as an interface metal layer. The interface metal was expected to introduce oxygen vacancies to the insulator, thereby decreasing the forming voltage. Advanced nanofabrication processes were employed in the cleanroom, and a related recipe was developed. The influence of layer thickness and device area was also studied to gain a comprehensive understanding. Among all the samples, Ru-based devices were observed to be forming-free.

Data analysis methods were applied to model the data, with the random forest method found to be the most suitable, achieving an accuracy of 82.4%. The model was verified by measurements of 10 nm Ru-based devices. Feature importance was then calculated to interpret the model. The four most important features determining the forming voltage are the thickness, standard electrode potential, area, and work function of the interface metal. This work adopts a new approach to eliminating the forming voltage, not only providing a forming-free device but also offering a guideline for future research on forming voltage.

Artificial intelligence has a strong need for faster and more energy-efficient solutions, especially for computation performed at the sensor edge. On-chip photonic neural networks (PNNs) offer a promising solution for high speeds and energy efficiency. A less explored side of PNNs is their application to time-series data, which is often the case for real-world sensor applications. While PNNs promise high speeds and energy-efficient solutions, no good use cases have been proposed. This report will first review the state of the art of PNNs. It will be seen that to solve time-series tasks, photonic continuous-time recurrent neural networks (CTRNNs) are required. The dynamics of CTRNNs are thoroughly explored to leverage obtained insights to recommend novel practical applications. This is done through simple examples, and applied to a real machine learning task. It was seen that on a real classification task the network learned two distinct fixed points corresponding to the classes. A link between the time constant of the continuous-time neurons and the temporal dynamics of the task is also found. Two general directions for novel applications are then proposed. Firstly, photonic PNNs can be slowed down to match the task. Opto-electronic PNNs allow for more control of the time constant, and on-chip photonic filter neurons are suggested. Secondly, the high speeds of photonic neural networks can also be directly leveraged. Extremely fast convergence of photonic CTRNNs can be utilized for Modern Hopfield networks, pathfinding algorithms, and time-dependent optimization problems such as for example Model Predictive Control.

In this study, a promising design for a compact, ultra-linear Compliant Torsion Reinforced Sarrus mechanism (CORS), capable of achieving ultra-linear motion is presented. The CORS prototype, made entirely of aluminum, is produced monolithically using electric discharge machining (EDM). The design incorporates four torsion-reinforced folded leaf springs, effectively reducing parasitic motion and enhancing support stiffness. To meet the specified requirements, a design optimization process is undertaken, carefully considering constraints to attain an optimal CORS configuration. Integration of the CORS with a voice coil actuator for driving force and confocal chromatic sensors for detecting parasitic motion is carried out. Experimental results demonstrate and validate the performance of the CORS.