The presented research aims to design and optimize a timber observation tower, with a primary focus being the influence of topological and curvature parameters on its stability and lateral stiffness in resistance to non-uniform wind load profiles.

Having recognised the environmental benefits and urgency to find other alternatives, there's a clear necessity to incorporate wood as the main construction material into the infrastructure projects, like observational towers. By conducting a study on hyperboloid towers implemented in the last 100 years, this project questions the necessity of in plane stiff platforms, flexurally stiff rings and continuous vertical members as being pivotal to the stability and lateral stiffness of the global structure when subjected to non-uniform wind loads. In addition, the study intends to investigate how the narrowness of the hyperboloid might affect the lateral stiffness of the structure. Thus, by utilizing parametric tools (Grasshopper, Karamba3D and Beaver plug-in), the study aims to design a timber tower structure comprised of fully segmented members as well as incorporation of circular (flexurally stiff) rings, aiming to address the questions raised. In terms of structural member arrangement, the study will investigate two topologies: one featuring a diagrid pattern that emulates a geometrical shape of an antiprism, and a custom pattern inspired by the post and beam approach, which resembles a regular prism shape. The study emphasizes the tower's multi-functionality and adaptability throughout its lifespan. Tower's main structural framework is a pivotal element in providing required stiffness and strength by excluding the need for in-plane reinforcement provided by arbitrarily placed platforms.

The key realisation of this research was the kinematic behaviour exhibited by the segmented, triangular tower, adversely impacting its stability characteristics. The study used a combination of analytical and graphic kinematics techniques, along with a physical mock-up model, to confirm the kinematic behavior of the tower given that even sided polygons for rings are incorporated. This revelation would have an impact on how a structure like that would perform as well the way it would be built. Further research unveils the strong influence of ring type on structural stiffness showing that segmented rings render tower structures less stiff than ones employing curved ring members, regardless of the pattern. In terms of the direct comparison between tower patterns, custom one demonstrated a more consistent and stable performance in general, particularly achieving higher stiffness levels when employing segmented rings. As regards the triangular pattern, the stiffest response against wind loads has been exhibited through the use of curved rings. In addition, the study validates that adopting a hyperboloid shape along with smaller shape factors for the global tower geometry yields more favorable lateral stiffness characteristics. Finally, the study navigates through the exploration of the most optimal connection design, illustrating how considerations related to detailing have necessitated a re-evaluation of the most optimal tower configuration, which initially was chosen to be a triangular one equipped with curved rings. A qualitative assessment of a potential joint within this specific tower variant has confirmed that designing such a connection is significantly more complex. This is due to the necessity of ensuring the continuous flow of the curved ring, emergence of a kink within the insertion of the plate and how that is needed to be addressed.

Alternatively, these considerations have motivated the design process to converge into a new hybrid design, integrating segmented rings with a curved top ring defined by the custom pattern. This choice has been made by conducting a separate parametric study of the new tower design and ensuring that the new connection design fulfills elastic slipping modulus and ultimate strength requirements. The final topology that showed higher stiffness metrics was the custom one. It's also highlighted that maintaining the top of the tower constrained leads to favorable effects on stability and stiffness, irrespective of the chosen topology. The resulting structure is optimized for mass by strategically reducing member cross-sections in accordance with connection scheming while adhering to both SLS and ULS criteria.

This research addresses the implementation of learning algorithms and generative design in string-based topology exploration methods. It aims to generate diverse structural patterns for shells and surface structures that align architectural, engineering, and construction objectives. By integrating reinforcement learning (RL) and quad-mesh grammars, surface topology is explored through quantitative metrics, demonstrating the strength and generality of this approach. The research ultimately promotes creative exploration during the conceptual stages of structural design, emphasizing collaboration between form-designers and form-analyzers to harness emerging computational techniques.

The quad-mesh grammar was first formulated within a Markovian decision framework to integrate with open-source RL Python packages. States, actions, and rewards were defined with sufficient generality to avoid over fitting while evaluating the RL agent's ability to navigate between two specified string-action sequences and their associated mesh layouts. Initially, simple tasks involving four design steps were tested, followed by more generalized target terminal states with longer design sequences. The impact of different reward structures and varied model parameter setups on convergence and accumulated rewards was also analyzed.

The findings indicate that reward functions based solely on topological and grammatical characteristics did not fully guide the agent from an initial coarse mesh to a target state. However, extended design episodes demonstrated potential for improved RL outcomes. The DQN struggled with non-optimal policies due to negative rewards and sparse positive reinforcement, suggesting that customized model architectures or alternative RL algorithms could enhance performance. The exploration phases yielded suboptimal but diverse mesh configurations, highlighting the need for additional structural and geometric parameters, as well as more complex grammar operations to improve diversity while mitigating computational challenges. These insights underscore the importance of balancing feasibility, exploration, and optimization in computational design workflows.

In many structures, space frames are used as the main load-bearing system. Especially for structures that are designed to have large spans or free-form in design. The reason is that space frames have a very high structural performance. Besides, they are also material and cost-efficient. Challenges for space frames are that they often require complex designs and the elements used are unique. Joint design is also difficult for these irregular constructions. This thesis explores the possibility of structurally optimising space frame design by using topologically reconfigurable modules, taking into account circularity. The focus lies on planar, square on square double-layered grids. The research question to be answered is the following:

”What kind of topologically reconfigurable modular system enables the generation of efficient space frames that are suitable for circular construction?”

The first design step is the initial topological design of steel cubic modules. This forms the basis of the catalog. With Grasshopper and Karamba3D (parametric FEM software) single-span trusses are then designed to determine the required cross-sections of the elements. The beams and columns are given SHS cross-sections for different spans and the diagonals are given various CHS cross-sections. With the software IDEA StatiCa various intra- and intermodular joints are designed and the rotational and translational stifffnesses analysed for
the different modules.

The former Grasshopper model is extended with a GA called Galapagos which is a plugin just like Karamba3D. This is used for topological optimisation of a given space frame structure. A GA is an evolutionary algorithm that is inspired by the natural selection process that Darwin described. The fittest genes in every population of solutions are used in further iterations. This process continues until the total number of iterations is reached or a threshold is met [1]. The algorithm reconfigures the different modules and minimises the weight of the structure
while staying within the constraints of maximum deflection, maximum material utilisation, and avoiding buckling. Logarithmic barrier functions are used for these constraints and together with the minimisation of the weight a fitness function is made. The lower the value of the fitness function, the better the solution is. The joint stiffness for each of the joint types for different |M|/N ratios is known. In the model, an initial stiffness is assigned to the joints. Then, depending on the moments and forces found in the joints, a new stiffness is assigned corresponding with the relations found in the stiffness analysis. The resulting stiffness loop describes non-linear behavior.

A verification is performed for the joints in the model. It was observed that the model was sensitive to small differences in stiffness within the joints of a standard truss. There were torsional moments observed that also resulted in an uneven distribution of the forces in the truss. To overcome this problem the stiffness of all joint groups is changed at once instead of individually. This resulted in more logical results where the support reactions were symmetric again. The model is then verified with a simple structure to see if the performance of the algorithm matches the expectations. A simple 2 x 2 grid is constructed and the model is tested. The simulation could not be performed until 50 generations because the stiffness loop led to a large accumulation of memory on the computer. It was concluded that the loop was not needed and therefore omitted because mostly only 1 iteration was needed. The model then was tested again on the small-scale model and compared to a couple of intuitively good-performing structures. This resulted in the conclusion that the model performed well and converged towards a final solution. However, since there were 106 different configurations possible the best solution was not reached in the end. It is recommended to always perform a couple of simulations to get a good solution.

After verification, the model is validated with a different FEM software called RFEM 5. Single grasshopper modules are structurally analysed in Karamba3D and imported in RFEM as well. The results show similarity in the order of magnitude of the stresses and the way the stresses are distributed. Furthermore, the reaction forces are also similar. The percentual differences between both models for the minimum and maximum stress are all below 5% except for module type 4. The absolute differences are in the order of 0.01 kN/cm2. Besides, a small frame is analysed in Karamba3D and RFEM as well to check if the deflection is in the right order of magnitude which is also the case. The percentual differences for the maximum and minimum stress are 6.40% and 7.85%, and for the deflection 12.15% and thus slightly larger than for the modules. However, for the deflection, the difference is just 1.5 mm for a structure with a span of 16 m which is a small difference. The Karamba3D model is valid and also conservative because the observed stresses are larger than in the RFEM model for the same
load condition.

A literature case study is performed on a space frame located in India. It is compared to the topologically reconfigurable model to check if the order of magnitude of the deflections and stresses is normal for this type of truss. First, a standard Pratt truss is modelled in Grasshopper to see if it is possible to design a feasible structure. This structure stayed within the limits of the constraints for the applied load which meant it worked. The case study truss is of a different size than the small structure used in the verification. With the catalog of 6 different modules and 66 possible locations for the modules as many as 10101 configurations would be possible. For this reason, the simulations are performed with 1 module type first (module type 6) and then with 2 (module types 3 and 4), greatly reducing the amount of possible
configurations. These are respectively 10^71 and 10^19. There are more solutions for the simulation with 1 module type because this module has a larger number of orientations than the other two modules combined. The outcome of the case study was that the algorithm did
not find better or equal solutions than the truss structure. This was because the convergence of the model was very slow, even for a simulation of more than 200 generations (10050 solutions) the best solution had a fitness value much larger than the truss structures. The issue was mainly that the extreme utilisation of some of the members was too large which made them fail. The BLF and displacement were within the boundaries. Comparing it with the case study the maximum stresses were indeed larger but the deflection was in the same
order of magnitude. The Pratt truss with reconfigurable members which resembled a regular truss was in the right order of magnitude for deflection as well. This truss does not have extra parallel beams since not entire cubes are joined together but single beams and columns.

An application is analysed to see how the model acts when the supports are irregular. From this setup, it was unclear what would be an optimal solution. The selected modules for this problem did not give the option to create a regular truss-like structure. The model found a slightly better solution than the one initially found using intuition and expert judgment. The value of the fitness function decreased from 0.84 to 0.80, which is a 5.5% decrease. The conclusion from this case study is that when the optimal solution is unclear and the design
space is large the model can find a relatively good solution. Combining this expert judgment and computational design can improve a model even more.

Overall the model can be used to design planar steel space frames with varying support conditions. These space frames are circular in design because their parts can easily be deconstructed and reused in similar structures. This is a huge advantage because it can save a lot of material and construction time as well. The modules themselves have the extra advantage that they can also easily be changed to a different type by varying the diagonals. This makes the system flexible in design. However, the optimisation efficiency is not very good.
The algorithm does not find optimal solutions that can be used in practice for a large design space. The type of structure is also much heavier in general compared to conventional space frames. This is because joining entire modular cubes together requires more steel. There are numerous parallel members that do not appear in regular space frames where beams, columns, and diagonals are individually placed in the structure. This makes stress on the members relatively high and therefore unfeasible. To use the system is therefore a consideration for the designer. When focussing on modular and circular construction this system is very useful but it is less structurally efficient. The model with ”modular elements” could potentially be used for a more structurally efficient system if the convergence towards a solution can be improved (with a different algorithm) and the speed at which solutions are calculated as well.