MODELLING OF POROSITY DEFECTS IN HIGH PRESSURE DIE CASTING WITH A NEURAL NETWORK

M. Imad Khan1, Yakov Frayman2 and Saeid Nahavandi 3 CRC for Cast Metals Manufacturing (CAST), Deakin University, Victoria 3217, Australia School of Engineering and Technology, Deakin University, Victoria 3217, Australia E-mail: mik1 , yfraym2, [email protected] Abstract: High Pressure Die Casting (HPDC) is a complex process that results in casting defects if configured improperly. However, finding out the optimal configuration is a non-trivial task as eliminating one of the casting defects (for example, porosity) can result in occurrence of other casting defects. The industry generally tries to eliminate the defects by trial and error which is an expensive and error -prone process. This paper aims to improve current modelling and understanding of defects formation in HPDC machines. We have conducted conventional die casting tests with a neural network model of HPDC machine and compared the obtained results with the current understanding of formation of porosity. While most of our findings correspond well to established knowledge in the field, some of our findings are in conflict with the previous studies of die casting. Keywords: Artificial Neural Network, High Pressure Die Casting, Porosity.

1. Introduction
High Pressure Die Casting Machine (HPDC) is a complex industrial system.

a change over position and the velocity is decreases at the end when nearly all the liquid metal is injected into the die. The metal is then injected through gate and runner system at a high velocity and pressure. The die is then opened and the solidified part is extracted out by a robotic arm. The die is lubricated with typically a liquid lubricant although there are some powder lubricants available as well. The role of a lubricant is to ease the extraction of casting and to avoid soldering of the metal with the die surface. The extracted casting is then cooled down with water and is placed on a conveyer belt for a further treatment, or otherwise stored on a rack for later quality control tests. The die casting process is a complex process consisting of over one hundred and fifty process parameters. The complexity of the process results in many problems like blistering, and porosity. However, the porosity is by far the most highly occurring defect in automobile engine castings [1].

2. Porosity
Porosity is the formation of voids inside the castings either through the entrapment of gas or improper pressure configuration in HPDC machines. Porosity is one of the most difficult defects to eliminate in die casting. The industry sometimes has to settle to move porosity to a different location in a casting rather than to remove it completely. It is in the best interest of industry and the consumer of die casting (for example, car manufacturers) that porosity is eliminated completely from the castings, but this is not always possible to do with the current level of process understanding. In addition, attempts to eliminate

Figure 1: A diagram of a die casting machine. In a typical die casting machine (Fig. 1) the molten metal is poured in the shot sleeve through a ladle. After the die is closed, a movement of a plunger (piston) forces the metal through the die resulting in that the moveable part coincides with the fixed part. Some die casting machines allow for this plunger movement to be completed in four stages [9], however, typically it is done in two stages only. The plunger starts initially with a low velocity, then the velocity increases during the pistons motion at

porosity defects can affect other process settings and result in other casting defects. The porosity can be dived int o three major types, which are: gas porosity; shrinkage porosity, and; flow porosity.

In HPDC, the first two types of porosity are mostly encountered. The gas porosity (Fig. 2) is the porosity in casting due to the presence of gas. This type can arise from: gas produced during process; entrapped air, and; melt composition. Figure 3: Interconnected shrinkage porosity. Formation of porosity in die castings is a combination of process parameters, melt composition, and solidification properties under high pressure. We will discuss process related porosity formation mechanisms in detail which covers solidification and gas related (entrapped) formation. Melt related porosity formation is not discussed in detail primarily because hydrogen entrapment in HPDC is not a big problem. Hydrogen can be considered seriously if the scrap is remelted which we assume is not the case. The specific reason for porosity formation is undesirable states of shot sleeve, cavity, vent and gates, runners, solidification pressure, lubricant quantity and steam formation from water during the process. Shot sleeve related parameters are perhaps the most sensitive one when it comes to entrapped air porosity. The parameters like acceleration, stage velocities, diameter, or even deceleration are all shot related parameters determining the formation of metal wave patterns which can be a crucial factor in deciding whether air becomes entrapped. As soon as the metal is ladled, the goal is to begin injection as soon as possible but still at the right time in the case of a cold chamber die casting machine. It should begin soon because the metal starts to solidify in the shot sleeve and if metal with solid particles is injected into the die, the high velocities can cause die wear and may contribute to die erosion and t o a deteriorated quality of the castings. It is not recommended to inject metal immediately because it can destroy the wave pattern and can entrap air in different forms. Hence 2

The shrinkage porosity (Fig. 3) is due to shrinking of metal, so that the metal loses volume and hence more metal is required to fill the gaps (voids) produced. In high pressure die casting, it is hoped that this problem can be minimized with the application of pressure to fill the voids when metal is in the solidification state.

Figure 2: Gas porosity with medium ( top figure ) and large sized pores (bottom figure).

shot command delay is the first process parameter to be selected carefully. Another process parameter to be optimized is the first stage velocity. If it is too low or too high, it can contribute to wrong wave formation. It is further explained with the help of a Fig. 4. The wave forms if slow shot velocity (1st stage velocity) is too slow in the Fig. 4. The wave gets on top of the air and the air is injected into the cavity.

If the biscuit is too small it can solidify first or even metal in shot sleeve can solidify which can take the pressure off from cavity. If the biscuit size is wrong then the pressure applied at the tip of plunger cannot reach the desired intensification pressure meant to remove porosity. The process is needed to be modelled and be understood well to reduce defect appearances in the castings. Thus the aim of this work is to enhance our understanding of por osity formation in die castings.

3. Methodology
Computational intelligence techniques that include artificial neural networks, genetic algorithms, simulated annealing and fuzzy logic have shown a promise in many areas including industrial engineering where the use of neural networks [6, 7, 13], g enetic algorithms [7] and fuzzy l ogic [3] is quite prom inent. The capability of Artificial Neural Networks (ANNs) to learn complex relationships well has made them a popular methodology for understanding the behaviour of complex systems like robot guidance [10], job shop scheduling [7] and die casting [6, 8, 13]. Computationally, ANNs in their most common form of a multilayer perceptron (MLP) are distributed parallel processing systems capable of a fault tolerance and efficient learning and are robust to noise and disturbances. They are connectionist structures composed of nodes called neurons and arcs connecting the neurons with weights associated with the arcs. Weights are adaptable and they are the main learning parameters of the network. The network learns typically by using a backpropagation learning algorithm [11] which updates the weights. The network has generally three types of layers called input, output and hidden layers. The information is presented in a pre-processed or raw format into the input layer of the network and the predictions are obtained at the output layer of the network. Mathematically, a MLP can be seen as either:

Figure 4: Entrapped air below the metal wave. The air in cavity can be entrapped due to the problems in runners or ventilation. The vents should be big enough to let the air escape and be located near the last spot to solidify. The runner should not have sharp corners in general. If the vents are working properly the air entrapped can escape to a sufficient extent. One of the main purposes pressure in die casting is to HPDC, no extra met al is reduce shrinkage porosity volumetric contraction. of application of high minimize shrinkage. In generally provided to which is a result of

Many die casters still find shrinkage related porosity despite of applying enough pressure. Research in the area suggested that the pressure applied can be different than the actual pressure developed inside cavity. It happens because of insufficient biscuit size or too big a size and unexpected solidification. Our neural network model has learned that the pressure which develops in the cavity is a decisive factor as shown in our results section later. The extreme temperature and presence of water at die surface can produce porosity due to steam. Water can accumulate on die from sprayer and leaking water cooling lines.

y j = ( xi wij ) ,
i =1 j =1

i =n j =m

where,

is a sigmoid function like 1 f ( x) = , 1 + exp ( x)

y j is the jth neuron in the curren t layer,

xi

is the feed-in from the previous layer, and

wij are the weights connecting the two layers,

y1 y 2 . = . . yn

w11 w12 . . . . . . . . wm1 .

. . . w1m . . . w2 m . . . . . . . . . . . . . . . wmn

x1 x 2 . . . xn

1st stage velocity 2nd stage velocity Changeover position Intensity of tip pressure Cavity pressure Squeeze tip pressure Squeeze cavity pressure Biscuit

The MLP adjust the weight matrix during the learning and is able to generalize for the data that it hasnt seen during the learning. Please note that sigmoid function is applied to each individual element of the matrix after multiplication. An MLP were selected for this work as the aim of the current research is an understanding and modelling of the casting defects in terms of machine parameters.

Figure 5: The neural network model used in this work to predict porosity at different physical locations in a 95mm AY2 Die Casting.

A C E F

5. Results
After modelling of die casting process with an MLP to a sufficient degree of accuracy, we conducted conventional die casting tests by varying one of the process parameters and keeping the others constant . This was done with a simulated process rather than on actual die casting machine as the experimentation on a die casting machine could result a considerable waste of resources in terms of metal, manpower and energy and incurs a significant cost. Fig. 6 shows a relationship between the quantity measures of porosity and the logarithm of slow stage velocity. The slow velocityporosity curve show that the obtained results are in agreement to what has been obtained by [4], modeled mathematically in [5] and as suggested by [12]. The neural network model of porosity show that the porosity decreases sharply with increas e in 1st stage velocity (slow speed velocity) and then the curves sharpness decreases as the velocity approaches Garbers critical velocity. It could be noticed that the network is generalizing well here and is able to model the correct porosity behavior. It is noticeable for all the results that the level of porosity is constant on a minimum level for location C and F at the casting. The data was collected from a machine which was producing castings with the requirement to move porosity away from locations A, C, E and F. It seems that the operation was successful to keep porosity away from locations C and F.

4. Experimental setup
We have used MLP to model location and quantity of porosity in a high pressure die casting. The data used to train the network consisted of process parameters related to porosity and location and quantity measures of porosity in the castings. The process parameters included 1st stage velocity, 2nd stage velocity, change-over position, intensification of tip pressure, cavity pressure, squeeze tip pressure, squeeze cavity pressure and biscuit thickness. The quality measures were X-Ray quality grades at four different locations used to represent porosity defects. The network was train ed using back-propagation algorithm [11]. The output of the network was the quality measures at four different locations named A, C, E and F. These quality measures are X-Ray grades ranging from one to four with one representing minimum level of porosity at the designated location and four as the worse. An occurrence of porosity level of four on any of the four locations of die casting results in a product being rejected. For accuracy and knowledge extraction reasons we have treated the porosity as a function approximation problem. The MLP that we have used to model die casting machine behaviour and to predict porosity defects is shown in Fig. 5.

Network Output(s) for Varied Input Average slow speed vel.

Network Output(s) for Varied Input Maximum Intensification of Tip pressure

Output(s)

Output(s)

A C E F

1.5 1 0.5 0 -0.638 -0.635 -0.632 -0.630 -0.627 -0.624 -0.622 -0.619 -0.616 -0.614 -0.611 -0.608

1.5 1 0.5 0 1.646 1.662 1.678 1.694 1.709 1.725 1.741 1.757 1.772 1.788 1.804 1.820

Varied Input Average slow speed vel.

Varied Input Maximum Intensification of Tip pressure

Figure 6: Relationship between the level of porosity and the logarithm of slow stage velocity (also known as 1st stage velocity).
Network Output(s) for Varied Input Average high speed vel.
4 3.5 3 A C E F

Figure 8: Relationship between the level of porosity and the logarithm of intensification of tip pressure.
Network Output(s) for Varied Input Maximum cav A Intensification pressure
3.5 3 2.5 A C E F

Output(s)

Output(s)

2.5 2

1.5 1 0.5 0 0.181 0.198 0.216 0.234 0.252 0.269 0.287 0.305 0.322 0.340 0.358 0.375

1.5 1 0.5 0 1.563 1.576 1.589 1.603 1.616 1.629 1.643 1.656 1.669 1.682 1.696 1.709

Varied Input Average high speed vel.

Varied Input Maximum cav A Intensification pressure

Figure 7: Relationship between the level of porosity and the logarithm of high stage velocity (also known as 2nd stage velocity). The average high speed velocity, also known as second stage velocity (Fig. 7) has a purpose to meet the resistance offered by gate and runner system and to inject the metal immediately as it reaches the end of shot sleeve. The critical high speed velocity is found to be with in the range of sensitivity analysis and it can be calculated as follows from the graph. Critical 2nd stage shot velocity is ALog(0.375)= 2.371 m/s, where ALog (.) is the Antilog to the base 10. This obse rvation is interesting in the sense that it shows clearly the ability of the neural network to model die casting process well along with the next observation .

Figure 9: Relationship between the level of porosity and the logarithm of maximum cavity A pressure. Note that the machine used to collect data was a multi cavity machine. Similar results were obtained for the other die cavity. The Fig. 8 should be seen in tandem with the Fig. 9 - the maximum cavity. A pressure in Fig. 9 is low in magnitude comparing to the intensification of tip pressure in Fig. 8. In fact, part of the tip pressure is transferred to the cavity. Please note the die under consideration is a multi cavity die, hence we are referring it as Cavity A. The neural network is able to learn that it is this factor which has the real effect on porosity and show a decreasing trend in porosity level as the pressure inside cavity increases.

Network Output(s) for Varied Input Biscuit thickness

[2]

1.5 1

[3]

0.5 0 1.516 1.535 1.554 1.572 1.591 1.610 1.629 1.647 1.666 1.685 1.704 1.723

[4]

Varied Input Biscuit thickness

[5]

Figure 10: Relationship between the level of porosity and the logarithm of a biscuit size. The results shown in Fig. 10 are interesting in the sense that they are conflicting to that available in the literature [2]. We can see a slight increase in porosity at the location E (in conflict to the literature) and a decrease in porosity at the location A (in accordance to the literature). It may have to do something with the thermal state of the die. It can also be interpreted as a different critical biscuit size requirement to reduce porosity at different locations. This finding is significant and requires further investigation.

[6]

[7]

[8]

6. Conclusions
In this paper an attempt was made to improve current modelling and understanding of porosity defects in HPDC machines. We have conducted conventional die casting tests with a neural network model of HPDC machine and compared the obtained results with the current understanding of the porosity defects. While most of our findings correspond well to established knowledge in the field, some of our finding are in conflict with the previous studies of die casting and as such are significant and require further investigation.

[9]

[10]

[11]

Acknowledgements
[12] Special thanks are due to Mr. Gary Savage, Nissan and CSIRO for providing data for e xperiments for this work. Figures 1-4 are courtesy of North American Die Casting Association (NADCA).

[13]

References
[1] Andresen W. T. and Guthrie B., Using Taguchi and Metflow to Determine Relationships Between Process Variables and

Porosity, 15th International Die Casting Congress and Exposition, St. Louis, MO. October 1989. Asquith, B. M., The Use of Process Monitoring to Minimize Scrap in the Die Casting Process, NADCA Transactions , T97063, 1997. Elkan, C ., The Paradoxical Success of Fuzzy Logic, Proceedings of the Eleventh National Conference on Artificial Intelligence, AAAI Press, pp. 698703 , 1993. Garber, L. W ., Filling of Cold Chamber during Slow-Shot Travel , Die Casting Engineer , July -August 36-38 1981. Garber, L. W., Theoretical Analysis and Experimental Observation of Air Entrapment during Cold Chamber Filling, Die Casting Engineer , May-June, 14-22, 1982. Huang J., Callau P. and Conley J.G., A Study of Neural Networks for Porosity Prediction in Aluminium Alloy A356 Castings, in B.G. Thomas and C. Beckermann, (Eds), Modelling of Casting, Welding, and Solidification Processes , VIII, TMS , June, 1998, pp. 1111 1118. Jain A. S. and Meeran S., A state-of-the-art review of job-shop scheduling techniques , Technical report, Department of Applied Physics, Electronic and Mechanical Engineering, University of Dundee, Dundee, Scotland, 1998. Kong L. X, Nahavandi S., and Baliga B., Defect analysis of high pressure die castings with artificial intelligence technology , Pacific Conference on Manufacturing , 506-511, Lawrence Technological University, USA, 2000. Plauchniak M. and Millage B. A., New Closed Shot Control System Features Total Integration, Die Casting Engineer , 1993. Pomerleau D. A, Neural Network Perception for Mobile Robot Guidance, PhD Thesis, T he Robotics Institute, Carnegie Mellon University, 1992. Rumelhart D., Hinton G., and Williams R., Learning Internal Representations by Error Propagation , in D. Rumelhart et al (eds), Parallel Distributed Processing 1 , MIT Press, 318-362, 1986. Thome M. and Brevick J. R., Optimal Slow Shot Velocity Profiles for Cold Chamber Die Casting, NADCA Transactions , 1995. Yarlagadda P. K. D. V. and Chiang E. C., A neural network system for the prediction of process parameters in pressure die casting, Journal of Materials Processing Technology, vol. 89-90, pp. 583-590, 1999.

Output(s)