MAS ETH DFAB - Thesis

Thesis Supervisors: Dr. Lloret-Fritschi Ena, Szabo Anna

Introduction - Objective:

During Digital Casting, concrete is deposited into a form-work in a spatially controlled way with a robotic arm. By the digital control of the fluidity and hardening of concrete, the material sets on demand. As the concrete hardens fast during fabrication, it exerts only minimal pressure on the form-work walls allowing for weakly supported form-work constructions from materials with low bending stiffness such as foils, sheets or textiles. These light form-works could provide a low-cost, reusable alternative to traditional form-works.

However this process takes time, from mix preparation to form-work construction and experiment to clean up. Furthermore due to the casting parameters of Digital Concrete such as fluidity and curing speed the final results of cast elements can look drastically different (fig 1.1, 1.2). How to configure or adapt the cast parameters for the different form-works is somehow a process, intuitively understood over the course of many experiments. This serves a wonderful premise and framework for my Thesis objective. If adequate behaviour is conceptually matched in the simulation of complex material (such as accelerated concrete) and reality of the mix, then a powerful tool can be envisioned to simulate and predict cast concrete members.

The thesis objective is as follows:

- Conceptually simulate the behaviour of Digitally Cast concrete within the framework of Anna Szabo’s PhD investigating Thin folded concrete members through the process of Digital Casting.

- Extrapolate Design potentials from the Digital Casting process.

State of the Art:

Many interfaces on fresh concrete flow simulation have been studied in the last 50 years. Most of the work carried out relates to correlation between mechanical properties and mix design of concrete, or in the field of structural engineering in order to associate the needed properties of the concrete to be cast with the structure to be built. However, it is only recently, that researchers from across the globe have started to work on casting prediction tools (Roussel and Gram, Simulation of Fresh Concrete Flow 2014).

Among the researched topics within simulation of concrete, methodologies have been developed from Computational Fluid Dynamics (CFD), Discrete Element Method (DEM), Finite Element Model (FEMLIP – Hybrid FEM) and Two-Phase Model to simulate the behaviour of material flow. DEM and CFD  methods use either the Lagrangian or the Eulerian description for motion (fig2.1). However given the time constraints of this thesis, building new computational methodologies to use in the simulation of concrete would be time consuming and inefficient for the final outcome of a design tool. Hence for speed purposes the simulation method of Flip Solver in Houdini will be used with a particle system. This methodology/ simulation solver utilizes Navier–Stokes equations for fluid motion rather than Lagrangian or Eulerian. These equations come from applying Isaac Newton’s second law of fluid motion (Hosch n.d.).

Despite this being less formalized than other methodologies in simulating the behaviour of poured concrete, the outcome is on understanding how to simulate the aesthetic properties of concrete flow in a form-work and to validate the results in a physical prototype. Therefore, a conceptual approach was taken combining the DEM and CFD. Its primarily point based, however when calculating pressure, the particle velocities are transferred to a grid and a volume is used to perform the fluid projection. This approach is known as a FLIP or Fluid Implicit Particle Approach. This approach is often seen in the Animation industry due to its fast calculations and much quicker simulation times, and due to it being used for film and movies, fluid aesthetics and fluid behaviours are often very art directed giving tremendous control (fig 2.2).

‍

Problem Statement:

Looking at it from a material science point of view, there are many problems with this methodology. In simulation of viscous material and of course the accuracy of such simulations they are often questioned and answered with digital slump tests. However, the biggest issues in my research methodology was correlating digital tests with physical prototypes using both casting parameters and simulation parameters. This was followed by the computational power needed to simulate a complete Digital Casting PhD prototype and the integration of a design tool which can simulate and create predictable results rapidly.

Correlating digital and physical tests may be complicated due to meshing of the particle data at the end of the simulation. A separate study for this might be needed. Multiple iterations were simulated to slowly narrow down the aesthetics of the simulated concrete and the physical prototypes. Computational power can be solved by understanding the simulation solver and fine tuning it to get the results we need that can be simulated rapidly. Another possible solution will be a faster timestamps for simulation and then matching meshing aesthetics to reality in post-simulation treatment of the simulated particles. Although this thesis covers quite a large scope, due to time constraints, exploration of the simulation tool and furthermore design potentials were completed for speed rather than accuracy.

The final problem was the implementation of a weakly supported form-work material resulting in not just a fluid simulation but a hybrid of 2 simulation solvers, cloth and fluid, as particles interact with the cloth geometry and vice versa.

‍

‍Research Methodology:

Using the existing physical prototypes and the noted casting parameters, a theoretical digital model was designed as an initial step, incorporating simple list of attributes that can change fluid behaviour. This model was not to scale or to speed, but rather the goal with the primitive digital model is to understand the behaviours that can be exploited using the simulation parameters, such as Viscosity, particle separation, particle radius, density, curing age and more. These were then used to explore design potentials for the fabrication process of Digital Casting.  To begin creating coherence between the physical and digital, three calibration tests are proposed: Flow-rate, Slump test and Spread test.
‍

Flowrate: With the knowledge of Anna Szabo’s existing cast prototypes a constant flowrate of 1.85L/min is used. Therefore a minute was simulated and a volume of 1.85L was filled. The most important parameters to get the flowrate accuracy were size of emission object and particle separation in millimetres.

‍

Slump test: To make the behaviour match further we ran the second calibration test often used to determine the consistency of concrete as-well as its yield stress. To do the slump test, we use the Hagerman’s cone and we deposit accelerated concrete in the cone. After a measured time period (in our case 5 seconds, 30 seconds, 60 seconds, 120 seconds) we lift the cone and measure the slump spread of the concrete. With the standard accelerator amount concrete hardens and keeps its cone shape perfectly after 2 mins. In order to match the physical reality my simulated fluid particles needed to slow down and eventually stop moving all together. This can be simulated with time dependant viscosity.

A simple mathematical formula was developed to create this Vc (viscosity change). Created parameter space include, Viscosity Initial (Vi), Viscosity Change (Vc), Density (D), Age (A), Curing Age Multiplier (Cam), Particle Separation (Ps), Particle Scale (Pss), and Frame counter (F). With this formula I exponentially increase the curing of simulated particles, matching the behaviour of accelerated concrete.

‍

‍

Spread/Flow test: The final calibration tests done to narrow down particle/fluid behaviour was Spread test. In Spread test we deposit concrete from a height of 1m and experimented with different robot speeds looping back and forth along a straight path. The slower the robot speed, the more material is deposited, causing larger material spreads. We tested first physically, measured a width range of 8mm to 25mm respectively from fastest robot speed of 200mm/s to 20mm/s. Initial digital test conveyed matching results other than one fact. Due to the particle density being too great the fluid simulation appeared to go in on itself. Differing to reality where Digital Concrete mostly flows on top. Therefore after a small study in density we had very accurate material behaviour (fig 3.3). Though this required re-running of the calibration tests.

Further investigation methodology: Once an informal coherence is developed among both the physical and the digital, we move to part two of the research which is correlating the casting parameters and the simulation parameters and creating a neutral behaviour. This was the first step to analyse the previous studies and create a baseline. This baseline was a highly fluid, form filling concrete without any errors or artefacts. Step two, a new setup with larger simulation model (regarding the size of the foam-work and the number of particles simulated for accuracy) and further defined list of simulation parameters was developed to further correlate the parameters on the two separate interfaces. Here, a small selection of physical experiments was done to better understand the behaviour of the concrete flow by varying two independent parameters that are needed to describe the rheological behaviour of fresh cementitious material: curing speed (accelerator amount) and the plastic viscosity (N. Roussel, Three-dimensional numerical simulations of slump tests 2004).  Understanding these casting parameters and how they can be manifested digitally in the simulation was the main point of this thesis. Upon the existing experiments done the two most coherent digital parameters matching the fluidity and accelerator amount are viscosity (change of viscosity over time as calibrated in the slump tests) and curing age multiplier (exponential speed of hardening). Though paramount, a physical to digital coherence needs more time to be developed completely and as such for thesis a more complete simulation package was prioritised than accuracy between interactions between different elements.

‍

Case Study / Results:

In order to evaluate the usefulness of the simulation tool, three case studies were developed with significant jump in scope between the last two. Despite a crude simplification on what can ultimately be achieved through a simulation tool such as this, these simple case studies provide enough insight in the Digital Casting Process. The first benefit of the simulation tool is the functionality aspect. Initially the goal was to test out something relatively simple – to simulate and perfectly fill a non-standard form-work. I knew from Anna Szabo’s existing prototypes that when she changes the amount of accelerator (the speed of hardening) she gets drastically different looks. Often times, she sees defects such as layering, improper form-work filling.

Case Study 01:

With a small study on curing speed a quick link was made between the accelerator amount and the curing speed in simulation (fig 4.2). With these studies a simple form-work was simulated with slightly wider outer edges. This means the standard mix that may fill a smaller form-work will not completely fill the larger counterpart due to reduced flow.

As expected, the pre-existing simulation parameters fail to fill the form-work perfectly. This causes’ defects such as layering and unwanted cavities. The simulation tool aims here to provide insight (though a little intuitively) on what aspects of the casting parameters need to change in order to fill a given form-work effectively. After many careful tests exploring controlled range for simulation viscosity and curing speed multiplier we were finally able to fill the form-work perfectly... Almost perfectly.

‍

Case Study 002:

Pushing the existing simulation but with more art directed control, I created a grid and coloured areas where I need an or some form of change (fig 4.4). I remapped this colour value to create data per frame that can be used to tweak any parameter. In this particular case I chose the curing speed of the mix. This resulted in the remapped values causing particles to cure much faster in areas of intervention

This resulted in approximately 10-15 simulations that showed no signs of implemented design constraints of the grid from 20 simulation studies. However, once again due to the power of simulation, many iterations were done where physically it would take the same amount of time to do 2-3. Similar to the previous process a careful range was discovered which exposited the design potential of this novel system (fig 4.6). This serves as a starting point for a huge range of explorations possible by evaluating individually the casting/simulation parameters.

‍

Case Study 003:

Instead of continuing along the path of the previously documented case study 2, the simulation tool was explored further and prioritised to encapsulate all aspects of the process of Digital Casting. As Digital casting celebrates the use of weakly supported form-work such as foil, textile or cloth, it only made sense to try incorporate the interaction of particle and deformable collision geometry with cloth/soft-body dynamics. This created a particular challenge as the collision geometry had to involve its own cloth solver in addition to the existing flip/fluid solver. Nonetheless, a cloth solver and its calibration, is a topic worth its own study. A conceptual approach was sought for the completion of the entire casting process.

Once the cloth solver was implemented it was used to bind the particles within the form-work and to deform by the particles themselves. This was done by giving each particle a small vector force from particle projection step which then affected the cloth geometry, in-turn affecting the particles. In respect to continuing a methodological approach once again a simple form-work was filled with the exception that one side is treated as deformable cloth. I then set constraints (seen in red in fig 4.7) to the cloth geometry restricting and casing the material to bulge under particle pressure. Studies in fig 4.7 investigate what happens when robot speed is slower towards the edges compared to slower in the middle.

Research Outlook:

Finally, a larger prototype was simulated with the idea of creating controlled deformations through Digital Casting parameters (fig 4.9). The research in simulation of material behaviour is at its birth. Based on the amount of exploration that is still left I believe with conceptual strategies such as these we can start understanding new and interesting fabrication processes. For instance, the cloth component, though rather important in the Digital Casting needs tremendous amount of calibration just on its own. While the two systems used in this research thesis, work together well enough conceptually, there are still many areas where I believe we can push for precision and accuracy and get very predictable and realistic results.

‍

Simulation Tool - Side FX Houdini

The simulation setup requires many iterations with varying aspects. In order to incorporate and explore the various options a completely procedural setup was designed using a procedural animation software called Houdini. Due to its encapsulated nature Houdini works with pure code under its hood, allowing for maximum control over all aspects of the setup. In Houdini node-based network graph is often used within different contexts or network types. In order to simulate fluid behaviour, you need: 1) Geometry network containing collision object and particle source objects as closed meshes. 2)  Dynamic network containing fluid object and voxel/point data which can be linked to the geometry network. Geometry such as particle can be cached with all necessary per-point attributes and can be visualised with various aesthetics for age, curing speed, and turbulent concrete (fig 4.1).

The simulation setup requires many iterations with varying aspects. In order to incorporate and explore the various options a completely procedural setup was designed using a procedural animation software called Houdini. Due to its encapsulated nature Houdini works with pure code under its hood, allowing for maximum control over all aspects of the setup. In Houdini node-based network graph is often used within different contexts or network types. In order to simulate fluid behaviour, you need: 1) Geometry network containing collision object and particle source objects as closed meshes. 2)  Dynamic network containing fluid object and voxel/point data which can be linked to the geometry network. Geometry such as particle can be cached with all necessary per-point attributes and can be visualised with various aesthetics for age, curing speed, and turbulent concrete (fig 4.1).

‍

Bibliography:
‍

Bhooshan, S., Ladinig, J., Van Male, T, and Block, P. 2018. “Function representation for robotic 3D printed concrete.” In ROBARCH 2018 - Robotic Fabrication in Architecture, Art and Design 2018, 98-109. Zurich: Springer.

Flatt, Robert. 2016. “Evolution of strength and failure of SCC during early hydration.” Cement and Concrete Research 288-296.

Hosch, William L. n.d. Navier-Stokes equation. Accessed July 3, 2019. https://www.britannica.com/science/Navier-Stokes-equation.

Huynh, H T, N Roussel, and P Coussot. 2005. “Aging and free surface flow of a thixotropic fluid.” Physics of Fluids 1-9.

Mechtcherine, V, A Gram, K Krenzer, J, -H Schwabe, and N Roussel. 2014. “Simulation of fresh concrete flow using Discrete Element Method (DEM): theory and applications.” Materials and Structures 615-630.

n.d. NSDesign. Accessed 08 20, 2019. http://nicksullivandesign.blogspot.com/2015/06/dop-multiple-cloth-objects-using-copy.html.

Remond, Sebastien, and Patrick Pizette. 2014. “A DEM hard-core soft-shell model for simulation of concrete flow.” Cement and Concrete Research 169-178.

Roussel, N, and P Coussot. 2005. “’Fifty-cent rheometer’ for yeild stress measurements: From slump to spreading flow.” Journal of Rheology 705-718.

Roussel, Nicolas. 2004. “Three-dimensional numerical simulations of slump tests.” Annual Transaction of the Nordic Rheology Society. Paris. 55-62.

Roussel, Nicolas, and Annika Gram. 2014. Simulation of Fresh Concrete Flow. State-of-the-Art, Paris: Springer.

Roussel, Nicolas, Mette R Geiker, Frederic Dufour, Lars N Thrane, and Peter Szabo. 2007. “Computational modeling of concrete flow: General overview.” Cement and Concrete Research 1298-1307.

Side FX. 2019. “FLIP Solver dynamics node.” Side FX. Accessed July 2, 2019. https://www.sidefx.com/docs/houdini/nodes/dop/flipsolver.html.