Optimising CIM Processes: Theoretical Physics vs. Real Data and Artificial Intelligence

Model-based calculations have been around since Newton formulated the falling of apples in a mathematical- physical equation with his Newton’s Law approach. Today, when we talk about model-based virtual optimization of Ceramic Injection Moulding (CIM), we mean the numerical solution of complex three-dimensional differential equations in DOE systems. The term Artificial Intelligence (AI) was first mentioned in 1956 at a workshop on an AI research project at Dartmouth College in New Hampshire/US. Data-based and AI-assisted optimization in industrial production is founded in an initiative of the German Federal Ministry of Education and Research in 2011, on the threshold of the so-called 4th Industrial Revolution. Driven by the internet, real and virtual world should grow together to an Internet of Things. Model-based optimization has become an established technology, also in ceramic injection moulding, on which reliable decisions in the planning of manufacturing processes can be based. In contrast, Big Databased approaches to process optimization (AI), especially for the injection moulding of polymers, metal or ceramic powders, are only a few years old. They are often developed and marketed by startups, with many questions about data structures and storage, standards of process and quality data still open, especially in production environments. In addition, there is the central issue of transforming data into valid information with which processes can be controlled and optimized. This paper will address how extremely complementary the two technologies are and what synergies can be gained when using both technologies: model-based virtual optimization (classical process simulation) offline during production planning, and data-based process optimization online during production. All this is documented along the CIM process chain from injection moulding to sintering.

CIM – process models based on 1st principles physics and thermodynamics
In principle, all phenomena occurring in CIM processes, which can be formulated in mathematical-physical equations and solved with the help of numerical methods on three-dimensional meshes, can be represented virtually on the computer, tested in detail and finally optimised. For more than 20 years, there have been many scientific works from academic institutes and software companies [1–3]. In the meantime, software applications that are used directly in industrial production planning and production itself are state-of-the-art. As an example of this, the process and mould design for a CIM watch case carried out with the aid of a model-based process assessment is shown below. The typical questions that arise at this stage, i.e. before the detailed design of the injection mould and before the exact manufacturing parameters are defined, include: 

• How can all cavities be filled simultaneously in a multiple mould? 
• How can the inhomogeneous and transient heat load be managed in terms of uniform temperatures during mould filling and solidification? 
• How can shrinkage errors, surface defects and dimensional inaccuracies be minimised or avoided altogether? 
• How can density differences in the green parts be avoided? 

And finally: 

• Which process parameters are relevant for achieving the desired part quality and which data must be recorded accordingly in the mould or on the injection moulding machine? 

This last question is elementary in the context of process data collection, process documentation and ultimately for databased online process optimization with AI approaches. These issues can be addressed with a single, well-prepared and thoughtout virtual DOE. In a methodical, lean approach, the first step after designing the watch case is to determine the gating point. On this basis, a simple geometric approach can be used to check where the locations of the final filling and possible weld lines might be based on this choice with a time expenditure of just a few minutes (Fig. 1). Already at this stage, it becomes apparent where the positions for pressure sensors should be located, with which the balance of the flow front can be controlled during mould filling in the later series production process. Then the dimensions of the injection mould and the shape of the runner and gate are determined. Watch cases have been produced in MIM or CIM for many years and in high numbers, so a lot of experience can be drawn from there. This means, that a probably good initial design of the injection mould – perhaps not the optimum one – can be made very quickly (Fig. 2). The pressure and temperature sensors in Fig. 2 are used to evaluate and compare the calculated (simulated) filling of the cavities: 

• The pressure sensor near the injection point is used to control the maximum injection pressure. 
• The two pressure sensors opposite the gating point are used to detect unbalanced filling, avoiding flash in order to ultimately assess and, if possible, minimise the weld lines. 
• The temperature sensor in the mould is used to control the mould temperature and thus to evaluate and ultimately optimise the measures for controlling mould temperatures. 

Today, model-based simulation of ceramic injection moulding processes is carried out by “virtual DOEs”, in which almost any number of reasonable variants are automatically calculated and automatically evaluated according to basic statistical rules. The different variants result from variable process parameters and their different combinations among each other. In the case of the CIM process for the watch case shown here as an example (Fig. 2), the virtual DOE was formed with five characteristic process parameters, each with three different nominal values (Tab. 1), resulting in a total of 35 = 243 variants. The calculation of the DOE with the 243 process variants and a semi-automatic statistical evaluation took 21 h on a normal engineering-workstation. In addition to the well-known results of a state-of-the-art process simulation, which will not be discussed in detail here, the main results directly available from DOE’s calculation are: 

1. Ranking of the calculated variants with regard to various criteria. These criteria can be product-related (e.g. air inclusions after mould filling, green part density or sink marks, Fig. 3) or business-related (e.g. cycle time, necessary clamping force of the injection moulding machine or necessary energy per part). Various criteria can also be evaluated as a combination with specific weighting. 
2. Diagrams of the main effects of process parameters (Tab. 1) on a wide variety of characteristics of both CIM processes and CIM parts that can be detected and documented by the simulations. 

From model-based to data-based process optimization
As mentioned above, model-based virtual optimization (classical process simulation) is performed offline as part of production planning. The virtual DOE’s as described above lead to a clear understanding of how and especially why different quality characteristics of the CIM parts depend on process parameters within a broad virtual test field. The parameters are clearly defined, 1st principles physics and numerics are validated, and the CIM processes are covered very comprehensively. Thus, one could say that process simulation can be seen as something like a “clean room production”. This means that the models used today lead to absolutely unambiguous correlations between the selected production parameters, the resulting conditions in the CIM process and quality characteristics of the CIM parts. A very important framework condition of the process is the thermal balance in the course of an injection moulding process in the thermally stable operating state. Each individual injection moulding process is highly transient, and the temperatures of feedstock and mould are inhomogeneous (Fig. 4). When viewing a film with the view selected here on the moving mould half, the areas of the mould where the mould temperature behaves highly dynamically or, on the contrary, with small fluctuations, become immediately apparent. This information is of central importance for the positioning of thermocouples to be used later for process control. This is exactly the case when it comes to clarifying in detail how pressure sensors should be positioned for process control. Starting from a certain pressure under which the feedstock is injected, there are naturally different effective pressures in the feedstock during the mould filling and cooling phases (Fig. 5). In addition to the methodical positioning of temperature and pressure sensors, the relevance of set process parameters on sensor data and quality characteristics of the CIM parts plays a major role with regard to databased process optimization. For the interaction of the model-based process optimization, which happens before production and off-line, and the data-based process optimization, which runs on-line with production, the already mentioned diagrams of the main effects represent the most important tool. First for the selection of sensors, which must be sensitive, characteristic and unambiguous in their response to changes in process parameters. Such sensors must be installed, other sensors can be installed. The Main Effect Diagram in Fig. 6 shows very clearly which sensors react sensitively to changes in process parameters. Blue and red represent a sensitive, relevant sensor. Gray means that changes in process parameters are not detected by the virtual sensor. Secondly, the Main Effect Diagrams have further significance for the assignment of quality characteristics of the CIM parts to process parameters. Here, too, there are strong couplings on the one hand, i.e. process parameters which must absolutely be kept in the specified range in the sense of product quality. On the other hand, there are weak couplings, i.e. the quality characteristics are robust against fluctuations of the compared process parameters (Fig. 7). 

Model-based virtual sinter process assessment and optimization
Up to this point, the first part of the production process chain for CIM parts, the feedstock injection, has been discussed. In the following, the end of this process chain, the sintering process, will be discussed. In all sintering processes, whether in a continuous furnace, batch furnace, under protective gas, reaction gas or in a vacuum, the temperatures in the individual sintered parts should rise as controlled as possible during heating, then be kept isothermal and fall again during cooling or quenching. To minimise stresses and the resulting cracks in the sintered parts, the temperature gradients within the sintered parts should also be kept as low as possible. If only the temperatures in the furnace and in the individual sintered parts are considered, an ideally slow heating and cooling in an ideally insulated furnace chamber would be the goal. This is, of course, impossible for metallurgical and economic reasons. In real sintering processes, sometimes significant temperature gradients prevail in furnaces and also within the sintered parts. It has been clearly demonstrated, both practically and virtually, that depending on the position in the sintering furnace and also depending on the heating and cooling rates, individual sintering curves apply to each individual sintered part [5–7]. Of course, this fact leads to differences in the quality characteristics of the sintered parts. Against this background, it must be examined whether this can lead to rejects, or if not, how high the potentials in time and energy savings can be if the sintering conditions in the furnace become more homogeneous and controllable within narrower limits. The watch case produced in CIM also serves as an example here. The sintering process takes place in a vacuum furnace in which a total of 1350 watch cases can be sintered in one batch in three racks on a total of 30 plates. The furnace chamber has a length of 1,5 m and a diameter of 1,1 m. There is a distance of 0,90 m between the two lateral heating surfaces, via which the heat radiation is transferred to the racks and the sintered parts (Fig. 8). In this example, too, two questions are at the forefront of the calculations. On the one hand, the homogeneity of the temperatures in the furnace chamber is to be assessed. The aim here is typically to determine the different temperature gradients resulting from different heating rates in order to find, among other things, the slowest heating temperature necessary to minimise stresses – and at the same time the fastest permitted heating rate with regard to the energy consumption of the furnace for the particular charge. On the other hand, the aim is to find out where thermocouples can be positioned for process monitoring and control, so that the characteristics of the sintering process can be recorded with as few as possible. As a representative of all possible sintering curves, the following one provides the parameter basis for the temperatures shown in Figs. 9 a–d: 

• Heating from 900 °C to 1450 °C, 2 different heating rates (5 and 15 K/min, respectively); 
• 8 h hold at 1450 °C; 
• Continue heating to 1580 °C, again different heating rates (5 and 15 K/min, respectively), then hold for 60 min. 

However, this certainly not optimal furnace and process lay-out clearly shows how sensitively the results of the model-based simulation react to different process parameters. Therefore, the tools and methods shown here can and should be used to evaluate different furnace and process layouts. With the help of the calculation of further variants, supported by the method of virtual DOE already shown above, furnaces and processes can of course be optimised even before the first “real” sintering process. 

From model-based to data-based process optimization with AI
Model-based virtual testing and optimization can therefore be used to develop extensive knowledge about the clear relationships between cause (process parameters) and effect (quality characteristics) in CIM processes. It is absolutely realistic to cover a virtual test field with 50–200 process variants within the scope of production planning, i.e. in the run-up to production. Today, a high level of reliability and informative value has been achieved in the forecasts of the CIM manufacturing process. However, there are cases where the forecasts do not exactly coincide with later results in production. This is mainly due to influences on production that are not represented in the simulation models – either because they occur erratically and unsystematically, or are simply not foremost in the minds of the simulation and manufacturing experts. In principle, such effects can be captured with the help of certain AI-based approaches. In contrast to the model-based virtual optimization described above, this databased process analysis and optimization takes place online during production. Certain prerequisites must be created for this: First of all, it is a question of positioning sensors in the injection mould or in the sintering furnace at the points relevant for the formation of the important quality characteristics of the CIM parts. For this purpose, model-based simulation as described above provides the best conceivable basis. Then the required sintered part specifications must be recorded in a measurable and digitally documentable manner. In particular, such data must be stored for each individual part produced and, moreover, must be assignable to each individual set of production parameters. Finally, fast algorithms must detect the patterns that may exist in the interaction between process parameters and quality characteristics. Once a sufficient amount of linked data from individual manufacturing processes is available, such patterns can be detected, discretized and documented [8]. The “sufficient amount” of data depends strongly on the complexity of the processes to be optimized, and this in turn on the amount of known and unknown qualityrelevant parameters. Usually, one speaks of several hundred thousand individual manufacturing processes. A significant reduction of the “sufficient quantity” can be achieved if the AI model is trained with the results from a virtual DOE [9]. 

Conclusions: about synergies
The described model- and data-based optimization tools and methods are used to obtain information on how to make the production start of CIM processes as smooth as possible on the one hand and how to make the running production as stable as possible within narrow limits on the other hand. It has been known for a long time that the availability and management of information is absolutely critical for technical and economic perspectives. It is also known that the consolidation of information can multiply the value of the individual pieces of information. This is exactly how it is with the technologies and methods described here: Both sides gain from the synergies between the proven model-based virtual optimization and the fairly new discipline of data-based process optimization. Specifically: • Using model-based virtual DOEs to generate incorruptible information about the cause-and-effect relationships between process parameters and quality characteristics of CIM products, which can be used to teach AI systems. • Using information from the simulation, where exactly which sensors need to be positioned and dimensioned to collect data for process control and the AI. • By discretizing process parameters that were unobserved yet have a quantifiable impact on part quality out of AI and in favour of advancing model-based optimization methods. 

References

[2] SIGMA Engineering: 3D moulding simulation – the “whole process” approach. Metal Powder Report 68 2013 

[3] 225–240 [3] Vietri, U.; et al.: Improving the predictions of injection molding simulation software. Polymer Engin. & Sci. 51 2011 [12] 2542–2551 

[4] Islam, A.; et al.: Experimental investigation of comparative process capabilities of metal and ceramic injection moulding for precision applications. J. of Micro Nano-Manufacturing 4 (2016) [3] 

[5] Hartmann, G.; et. al.: Fast and efficient optimization of the load temperature distribution in high temperature sinter furnaces. Euro PM2017, Conf. Proc., Milano, 2017 

[6] Hubmann, R.; Kukla, C.: Final report of the project IndAlMIM funded by the Austrian Funding Agency FFG under the number 8342. Private communication 

[7] Hartmann, G.; Thorborg, J.: Closing the virtual MIM process chain from design assessment to debinding and sintering. Euro PM2019, Conf. Proc., Maastricht, 2019 

[8] Salentin, F.G.: Data-driven system identification and fault-diagnosis in manufacturing processes. Master Thesis, University Duisburg- Essen, 2019 

[9] Salentin, F.G.: Entwicklung einer modularen datengetriebenen Optimierungskette für Fertigungsprozesse am Beispiel der Gießereiindustrie. PhD Thesis, Hochschule Kempten, 2023