LIQUIDUS TEMPERATURE DETERMINATION IN MOLTEN SALTS

S. Rolseth ¹⁾, P. Verstreken ²⁾ and O. Kobbeltvedt¹⁾

1) Electrolysis group, SINTEF Materials Technology, N 7034 Trondheim, Norway
²⁾ Heraeus Electro-Nite Int. N.V., B-3530 Houthalen, Belgium

Abstract

Three different ways of determining the liquidus temperature of Hall-Héroult baths are discussed:

- bath analysis and application of liquidus equations

- thermal analysis in laboratory furnaces

- using liquidus probes or sensors

These methods are compared in both synthetic and industrial baths. Basically, all three methods make use of thermal analysis. Taking bath samples for chemical analysis only introduces more errors and seems to be the least reliable method.

Due to the high temperatures and the corrosive nature of the cryolitic baths used in the Hall-Heroult process for aluminium electrolysis, it is difficult to monitor important process parameters like bath composition and temperature. Because of the corrosiveness of the bath the process is based on the existence of a stable ledge of frozen bath to protect the sidelining of the cells. The stability of this sideledge is determined by the heat flux through the sidelining, which means that the temperature difference between the bulk of the bath and the liquidus temperature of the bath is an important process parameter with respect to the thermal balance of the cells. This temperature difference, called the bath’s superheat, is not readily measurable, because one has to rely on bath sampling and subsequent chemical analysis or direct thermal analysis (in a specially designed furnace) to determine the liquidus temperature. In both cases one has to wait for the analytical results. It is obvious that this makes superheat determination less useful as a diagnostic tool because any decision-making will be delayed by this analytical lead time.

One factor to consider is the inherent uncertainty associated with using bath analysis to calculate the liquidus temperature. Several equations for calculating the liquidus temperature from bath composition data have been published over the last 30 years [1] , [2] , [3] , [4] , [5] , [6] . For some bath compositions the discrepancy between the various equations is quite large, and even for the SINTEF equation [6] , which in our opinion is the most reliable, one must expect an uncertainty of about 2 °C in the calculated liquidus temperature for AlF₃-rich bath compositions (AlF₃ excess > 10 wt-%). The contribution from uncertainty in bath analysis data is probably more serious. Experience has shown that analysis of Hall-Heroult baths is very difficult, and there is a large scatter in the analytical results when the same bath is analysed by various plant laboratories. In addition one should expect a lowering of the liquidus temperature of about 1 - 3 °C due to other bath components not accounted for in the liquidus equation, for example the various contaminants in industrial baths (and LiF and MgF₂ when not analysed).

With these factors in mind it seems obvious that direct measurements of the liquidus temperature would be advantageous compared to the most common practice of to-day, i.e. calculating the liquidus temperature from bath analysis data. It may seem surprising that very few attempts have been reported on developing a sensor for in situ liquidus measurement in Hall-Heroult baths. Gan et al. [7] used thermocouples embedded in spherical shaped metal (Cu/Ni/stainless steel) probes of 8 to 30 mm diameter. The determination of the liquidus temperatures was based on the temperature time relationship obtained when the cold probe was immersed in the molten bath. A break in the curve was observed when the temperature reached the bath’s liquidus temperature due to the disappearance of the last traces of frozen bath from the surface of the probe. It was claimed that the accuracy with respect to liquidus temperature measurement was within 2 °C. However, it was difficult to obtain distinct breaks in AlF₃-rich baths. To some extent this method will be sensitive to the local bath flow rate as the heat transfer from the bath to the probe will increase with the flow rate in the vicinity of the probe.

A similar approach has been reported by Grimsey et al. [8] . They used cylindrical probes of copper or nickel and a fast computer based logging system to record the heating profile when the probe was immersed in a molten salt. As for Gan’s probe the liquidus temperature was determined as the temperature where the remaining frozen salt slipped away from the probe, manifested as an increase in the heating rate of the probe. This probe was tested in pure salts (NaCl and NaNO₃) and in iron slag systems. In the pure salts systems the measured liquidus temperatures fell within ± 5° C of the respective true values. In molten salt mixtures one can raise the question if the last traces of frozen bath disappears from the metallic surface at the liquidus temperature because the "quenched" bath adhering to the probe will start melting at a temperature well below its liquidus temperature.

The instrument used in the experiments in this work is based on the principle of classical thermal analysis where a sample of the bath is cooled in a crucible containing a thermocouple. The temperature decay is monitored and the liquidus temperature is determined as the temperature where cryolite starts to solidify, manifested as a break in the temperature curve. Some of the essential features and the guide lines for the development of the instrument are explained below.

The task of developing a liquidus probe was undertaken on the basis of Heraeus Electro-Nite expertise in making thermocouples for the steel and aluminium industry. The use of type S thermocouples in a miniaturised sensor is one of the characteristic features of the probe. Due to its low thermal mass a stable reading is obtained before the protective materials are attacked. The Cry-O-Therm liquidus sensor was developed with the aim of obtaining fast direct measurements in an electrolysis cell. The measuring principle is that of traditional thermoanalysis, i.e. the thermal response is recorded when the bath sample is cooled. When small samples are used, supercooling is a source of error that should be avoided. Supercooling can normally be eliminated by stirring and by addition of seeds to the sample. In this case it was solved by vibrating the whole miniaturised crucible compartment with the thermocouple. In addition the inner wall of the crucible which was made of copper, was roughened to ease nucleation during cooling of the sample. The design of the sensor is shown in Figure 1. Cardboard is used at the carrier tube material since it serves well as an insulator during the measuring period. Due to its moisture content is creates convection in the bath when the probe is immersed for sampling. This is regarded as beneficial for obtaining representative bath samples and increasing the heat transfer to the probe. The probe is connected to a PC-controlled data logger that samples with 0.1 sec intervals. A more detailed description of the sensor is given in [9] .

Figure 1. Design of lower part of the liquidus probe showing the carrier tube (cardboard), crucible (copper) and thermocouple (type S) protected by copper-sprayed quartz tubes

A measuring cycle begins with immersion of the sensor in the bath. The sensor is connected to a hand-held lance, and when a pre-set temperature is reached, a motor starts vibrating the whole lance. When a stable bath temperature plateau is reached a sound and light signal is given for the operator to remove the probe from the bath. The cooling starts immediately when the probe is removed from the bath and the "end" signal is given when the measurement is terminated. The whole measuring cycle normally takes about 40 seconds. The data is immediately processed by the PC-based unit. Plateau and curve determination is made by various mathematical algorithms. The determined temperatures are immediately displayed on the monitor screen. The data is automatically stored in the computer’s memory and the curves can be recalled for viewing later. Altogether 300 curves can be stored in this way.

The development work was mainly carried out in conjunction with measurement on industrial prebake cells. The measured values were compared to the calculated liquidus temperature from bath analysis data, and in many cases the discrepancy was unacceptably large. As discussed in the Introduction there are several uncertainties associated with the calculated liquidus temperature, and it was then decided to do some control laboratory measurements. Several series of such laboratory experiments have been carried out.

A bath sample of 6.0 kg bath was prepared from pure bath components: 5220 g Greenland hand-picked cryolite, 360 g sublimated AlF₃, 300 g p.a. CaF₂ and 120 g technical grade Al₂O₃. The bath was melted in a 20 cm id. graphite crucible. A calibrated type S thermocouple protected by a stainless steel tube was immersed in the bath. The bath could be stirred by a graphite impeller. The SINTEF equation [6] gives a calculated liquidus temperature of 980.5 °C for this composition. The bath was kept molten for 9 hours, and over this period six measurements were taken with the liquidus probe. One 200 g bath sample was also taken for measurement in a separate apparatus specially made for liquidus measurements. This is the apparatus used in the work were the SINTEF equation was developed [6] . Three times during this period the liquidus temperature was also determined by a so-called in-situ measurement where the power to the furnace was shut off and the temperature measured as a function of time while the bath in the 20 cm crucible was stirred by a graphite impeller. A distinct break in the cooling curve showed the liquidus temperature. The results are shown in Figure 2.

It is evident that the bath composition changed during the experiment. This could be attributed to a minor leak in the water cooled tube at the top of the furnace. The leak was hardly visible but one could expect that some of the humid air from the top of the furnace got access to the bath through several openings in the lid of the furnace. At this bath composition the reaction: 3H₂O + 2AlF₃ ® Al₂O₃ + 6HF will decrease the liquidus temperature of the melt ( i.e. the liquidus temperature will be more dominated by the increase in Al₂O₃ content than the decrease in the AlF₃ content). If one extrapolates the measured values back to time zero the results also appear to be fairly consistent with the calculated liquidus temperature based on the composition given by the amounts of weighed-in pure bath components. In this case the deviation is also about 1 °C.

Figure 2. Liquidus temperature measurements in a 20 cm id. graphite crucible. Liquidus temperature measurements plotted vs the time the bath was kept molten. Measurement in liquidus furnace (square) is based on bath sampled at the time indicated in the graph

Due to the high cost by using pure (sublimated) AlF₃ the subsequent series of measurements were performed with industrial grade AlF₃. This means that we lost the possibility to do a reliable comparison with the calculated liquidus temperature as we could not rely on our bath analysis data. 6500 g bath was prepared from 5395 g hand-picked Greenland cryolite, 677 g industrial grade aluminium fluoride, 325 g calcium fluoride (p.a) and 103 g "industrial" grade alumina. Assuming that cryolite and aluminium fluoride contained 0.25 % and 4.5 % alumina respectively, this gave a bath composition of 2.25 % Al₂O₃, 5 % CaF₂, 10 % AlF₃ and 82.5 % Na₃AlF₆. From the SINTEF equation [6] a liquidus temperature of 968.3 °C was calculated. Over a period of about three hours three series of measurements with the liquidus probe were taken at three different bath temperatures. As there was no electrolysis and no addition of alumina or AlF₃ during the experiment, it was assumed that the bath composition was constant as the duration of the experiment was relatively short (the leak on top of the furnace had been fixed).

An in-situ liquidus measurement was performed after three hours. After a break in the cooling curve was observed and before the bath was completely frozen, the power was switched on again and the bath was reheated to about 983 °C. Approximately 200 g was then sampled for measurement in the special liquidus furnace. The results are shown in Table 1

The liquidus probe and the "in situ" measurement agreed reasonably well (±1 °C) with the liquidus temperature of 968.3 °C calculated from the estimated bath composition. The liquidus temperature measured in the liquidus furnace was about 3 °C lower than the other values. The recording in itself looked very reliable, but one explanation for this discrepancy is that the bath was sampled before all the bath had been completely re-melted after the "in situ" measurement. This emphasizes the fact that selecting a representative bath sample always is of major importance in this type of investigations.

Table 1. Liquidus temperature measurements with the liquidus probe in laboratory cells compared to liquidus temperature measurements measured "in situ" by cooling the whole bath and compared to the liquidus temperature of 200 g bath sample measured in specially liquidus furnace.

Three minor bath samples (» 20 g) were also taken for analysis during this run. The excess AlF₃ was analysed by the ISO 4277 wet chemistry method and the alumina by a LECO instrument. The CaF₂ content was not analysed however, as the content was assumed to be equal to the weighed-in amount. The average bath composition calculated from these data was: 11.4 % AlF₃, 3.39 % Al₂O₃ and 5 % CaF₂. The calculated liquidus temperature for this composition is 955.8 °C. This is more than 10 °C below the measured values, and as both the Al₂O₃ and AlF₃ contents are higher than the composition calculated from the weighed-in amount, this data was discarded. This emphasized the point that obtaining reliable analysis data is a very difficult task, and that actually measured liquidus data tend to be superior to calculated liquidus temperatures based on bath analysis data. If we assume that the deviation is due to systematic errors in the bath analysis, the data shows there was no significant shift in the bath composition during this series of measurements.

At the end of the experiment a batch of 300 g of industrial grade AlF₃ (containing 4.5 % Al₂O₃) was added to the melt. On a mass balance basis a new bath composition was calculated : 13.84 % AlF₃, 2.35 % Al₂O₃ and 4.77 % CaF₂, which gives a calculated liquidus temperature of 949.5 °C. The average of four measurements with the liquidus probe gave a liquidus temperature of 950,0 ± 1.0 °C. A new "in situ" measurement in the graphite crucible showed a liquidus temperature of 950.5 °C. A liquidus measurement in the liquidus furnace of a 200 g bath sample taken prior to the in situ measurement gave 952.4 °C, and another 200 g sample carefully split from the remaining bath in the cooled crucible after the experiment showed a liquidus temperature of 951.3 °C. As for the experiments before the addition of AlF₃, the temperature calculated from bath analysis data gave values about 10 °C below the measured values.

Some additional experiments were also performed with "industrial" bath collected from a pre-bake cell. This stemmed from about 300 kg excess bath tapped from a point fed cell in one of the Hydro Aluminium plants. Altogether four runs were made with this bath. A sample of about 6 kg bath was collected and melted in the 20 cm i.d. graphite crucible. In spite of the fact that the bath was tapped from the same cell, one should expect some variation in composition between the various runs due to segregation of the bath after tapping. The pieces that constituted the melt were picked arbitrarily from the batch of tapped bath. One of the main purposes with these experiments were to throw some light on the reproducibility of the liquidus probe. Liquidus measurement (3 to 13) were made with the probe in each run at various bath temperatures. In two of the runs 200g bath was sampled for liquidus measurement in the special liquidus furnace and in two runs an in-situ measurement was performed. In a few cases the experiments were disturbed by heavy construction work in the vicinity of our laboratory. Even if the computer failed to find the liquidus temperature it was still possible to find the break by visual observations on the screen or by studies of the printed curves, as shown in Figure 3.

In these runs with industrial bath the measurements with the liquidus probe were on the low side compared to the in-situ measurements and the measurements in the liquidus furnace. A deviation of about 1 - 3 °C is still small compared to the uncertainties one must accept if one has to rely on the calculated liquidus temperature from bath analysis data.

In between Run 2 and Run 3 a new version of the algorithm for determining the liquidus temperature was installed in the PC-based control unit. The tables below indicate that this was an improvement as the experimental scatter was reduced.

Figure 3 - Print-out of a temperature recording of a measurement. The recording represents measurement 4 in Run 2 (shown in Table 3). A bath temperature of 978 °C and liquidus temperature of 956 °C are read from the curve.

Table 2 - Run 1 (97-04-23). Measurement with 6 kg industrial bath in 20 cm id graphite crucible.

Table 3 - Run 2 (97-05-23). As in Table 2.

Table 4 - Run 3 (97-05-26). As in Table 2.

Table 5 - Run 4 (97-05-27). As in Table 2.

In connection with Run 2 (Table 3) several bath samples were taken using a small graphite crucible. The samples, after cooling, were crushed and the 40-50 g bath samples were distributed to representatives from different smelters for chemical analysis. The results are shown in Table 6 together with the liquidus temperatures calculated from the SINTEF-equation [6] .

Table 6 - Analysis data received from different smelters on the bath which was used in the experiment 1997-05-23. The liquidus temperatures are calculated by inserting the analysis data in the SINTEF equation. In the two lowest rows are shown the liquidus temperature determined from the liquidus probe and an in-situ measurement, respectively.

In spite of the fact the sampling technique as well as the milling procedure might affect the analytical results the scatter in Table 6 is still too large to explain the discrepancy between the measured and calculated liquidus temperatures. This supports the notion that chemical analysis of the same bath can give significant variation depending on the analytical procedure used by the different smelters. The consequence is that liquidus determination based on bath analysis is highly dependent upon how the bath is analysed.

During test with the probe at different smelters bath samples have also been taken for liquidus determination in the specially designed liquidus furnace. The samples (» 200 g) were taken in the same positions where the probe was dipped into the bath and within the time interval of a series of three measurement or more. The results are given in Table 7. For the industrial measurements the scatter is given as the standard deviation, while an uncertainty of about ± 0.5 °C is estimated for laboratory measurement based on duplicate runs with the same (200g) sample. For comparison liquidus temperature recordings obtained by the two different methods are shown in Figure 4.

Figure 4. Liquidus temperature measurement in the SINTEF liquidus furnace [6] of a bath sample taken from an industrial cell (H75 in Table 7) in comparison to a direct measurements with the liquidus probe in the cell.

For most of the cells excellent agreement was found between the two types of measurements, i.e. a deviation of less than 1 °C. The largest deviation was found for cell No 2, Smelter G, where also the scatter in the liquidus probe measurements were large compared to the other series. If this reflects variations in bath composition (unstable cell?), the same uncertainty will be present for the bath sampled in this cell. For cell 466 a deviation of about 3 °C was found. For most industrial applications this is within an acceptable range of errors. As only a limited number of cells were investigated it is difficult to tell whether these deviations are atypical or not. This will be investigated further as similar measurements will be performed on other industrial cells.

Table 7 - Liquidus temperature measured in industrial cells compared to liquidus temperatures measured in the SINTEF liquidus furnace using bath samples from these cells.

Figure 5, 6 and 7 show results from an experiment where 100 kg AlF₃ was added to a prebake cell in order to throw some light on the "dissolution kinetics" of AlF₃ in Hall-Heroult baths. The fluoride was added near the side channel between two anodes at a distance of two anodes away from the end channel. The time needed to add all the fluoride was 2-3 minutes. The liquidus temperature and thus the superheat was measured in two positions in the same side of the cell as the feeding hole. The positions , A and B, were at a distance of three and six anodes from the hole. In the course of 20 minutes prior to and 70 minutes after the AlF₃-addition 9 measurement were made in each position. Several bath samples were collected in the same positions and later analysed with respect to Al₂O₃, AlF₃ and CaF₂. Figure 5 shows the bath and liquidus temperatures as a function of time after the AlF₃ addition, while the analytical results are displayed in Figure 6. The results show how the bath temperature decreases as the cold fluoride is added and how the liquidus temperature decreases as the AlF₃ is dissolved.

Figure 5. Measurements with the liquidus probe in a prebake cell. Bath and liquidus temperature as a function of time after addition of 100 kg AlF₃. Pos. A is three anodes away from the feeding hole, pos. B is six anodes away.

Figure 6. Bath analysis data from the same experiments as shown in Fig. 5.

Figure 7. The measured superheat shown as a function of the time after the AlF₃ addition (same experiment as shown in Fig. 5)

Figure 7 shows that the superheat increased from 5-6 °C to 11-12 °C in a period of 30 - 40 minutes after addition. Figure 6 shows that in the same period an increase in the AlF₃ and Al₂O₃-concentrations was observed (the fluoride contains alumina). The fact that the change was simultaneous in both positions shows that the mixing of dissolved AlF₃ in the bath was good. About 45 minutes after the addition it can be seen that the superheat as well as the AlF₃ and Al₂O₃ concentrations start to fall off. This can be explained by melting of the side ledge caused by the increased superheat. As the side ledge contains little AlF₃ and Al₂O_3, those bath components will be diluted when this melting occurs.

When the bath analysis data was inserted in the SINTEF liquidus equation [6] a liquidus temperature of about 957.5 was calculated for bath sampled prior to the addition and about 939.5 for bath sampled about 40 minutes after addition. Thus, from the bath analysis data a decrease in liquidus temperature of 18 °C was estimated, while the measured drop was only 12 °C as shown in Figure 5. All the other data in this paper suggest that the measured data is the most reliable in this respect. A dynamic heat flux simulations where the change in superheat is too large by about 50%, will obviously lead to erroneous results. This again stresses the importance of measuring the superheat instead of relying on bath analysis data. In this case we also got the measured values immediately during the measurements while we had to wait for several days for the analytical data.

Liquidus determination of the bath in reduction cells can in principle be made in two different ways. The common industrial practice is to chemically analyse a bath sample and to determine the liquidus temperature by means of an empirical equation. Alternatively, the liquidus temperature can be determined by direct thermal analysis of a bath sample. By summing up the sources of error which is attached to each of these two methods a comparison of the total errors for each of the methods can be made.

Since in the prior method an empirical equation is necessary to convert the analysis results into a liquidus temperature, this method will include errors which come from the making of the liquidus equation. All these equations are purely empirical and thermal analysis of baths with well known compositions are the basis for these equations. The thermal analysis in itself introduces one source of error. In addition there comes a source of error in how the data is represented in the model, i.e. inaccuracy in the interpolation between the measuring points.

When the equation is to be used a bath sample for chemical analysis has to be taken. This implies the error of collecting a representative bath sample. If the bath acidity is to be analysed by means of XRD, which is the common practice by most smelters, the rate of cooling of the sample is important since this determines the amount of the different phases which can be recorded by XRD Therefore, the cooling rate introduces another source of error. The bath sample is normally analysed for three components: AlF₃, Al₂O₃ and CaF₂ and one source of error is attached to each of the three analysis. Finally, as mentioned in the Introduction, contamination by impurities may represent a significant source of error.

In the latter method (direct thermal analysis) one still has a source of error connected to collecting representative bath samples in addition to the direct error in the thermal analysis. Nevertheless, compared to chemical analysis one eliminates six sources of error by applying direct thermal analysis in the liquidus determination.

Normally it is the superheat of the bath one seeks when the liquidus temperature is to be determined. When applying bath analysis in the liquidus determination a new source of error is introduced when calculating the superheat, namely inaccuracy in the bath temperature measurement. When the liquidus temperature is determined by direct thermal analysis by means of a sensor, the bath temperature is measured by the same thermocouple which is used in the thermal analysis. With this method, therefore, any systematic error in the thermal reading becomes insignificant for the superheat determination.

In the experiment in the present work where chemical analysis was performed by different smelters, it was seen that the spreading in the analytical results from the different smelters gave rise to significant variation in the calculated liquidus temperature (see Table 6). Therefore, the problem of analysing correctly probably represents the greatest source of error, while the source of error connected to the liquidus equation is of far less importance. This assertion is supported by the results in the first series of liquidus measurements. In this experiment the liquidus temperature determined by inserting the weighed-in bath composition in the SINTEF equation was in good agreement with the liquidus temperature determined by all three methods of direct thermal analysis (see Figure 2). The liquidus temperature calculated from bath analysis data showed to deviate more than 10°C from that which was measured.

In the experiment where the liquidus temperature was followed after a large addition of AlF₃ to the cell, the change in liquidus temperature calculated from the rise in bath acidity determined by bath analysis deviated clearly from that which was measured by direct thermal analysis. This result shows that direct thermal analysis is superior to indirect determination via bath analysis for this kind of cell dynamic studies. Furthermore, when applying direct thermal analysis the results is available immediately after each measurement. This will make it easier to adapt the experimental scheme while running a series of measurements.

Direct measurements of the liquidus temperature in industrial cells by the use of the liquidus probe is clearly advantageous compared to the alternative methods for determining the liquidus temperature. The greatest advantage from an operator’s point of view is probably the speed of the measurement. The almost instantaneous reading of the bath and liquidus temperatures, also gives superheat directly. The duration of a measurement is about 1 minute, which makes it possible to run frequent measurements. Hence the probe can also be a useful tool in cell dynamics studies.

The experimental scatter was found to be less than 1° C, and in control experiments where the average liquidus temperature measured with the probe were compared to the liquidus temperature determined by traditional thermal analysis in laboratory furnaces, the deviation in liquidus temperature was less than 3 ° C. This is within an acceptable limits for industrial measurement.

[1] E.W. Dewing, "Liquidus Curves for Aluminium Cell Electrolyte". J. Electrochem. Soc. 117, (6) 781 (1970).

[2] S. S. Lee , K-S. Lei , P. Xu and J. J. Brown jr., "Determination of melting point temperatures and Al2O3 solubilities for Hall-Heroult cell electrolyte compositions". Light Metals 1984, pp. 841-855.

[3] A.T. Tabereaux, "Phase and chemical relationships of electrolytes for aluminum reduction cells". Light Metals 1985, pp. 751-761.

[4] G. L. Bullard, and D. D. Przybycien, "DTA determinations of bath liquidus temperatures : Effect of LiF".Light Metals 1986, pp. 437-444.

[5] R. D. Peterson and A. T. Tabereaux, "Liquidus curves for the cryolite - AlF₃ - Al₂O₃ system in aluminum cell electrolytes ". Light Metals 1987, pp.383-388.

[6] A. Solheim, S. Rolseth, E. Skybakmoen, L.I. Støen, Å. Sterten, and T. Støre, "Liquidus temperature for primary crystallization of cryolite in molten salt systems of interest for aluminium electrolysis". Met. and Mater. Trans. B., 27B, pp. 937-744 (1996).

[7] Y.R. Gan, Z.S. Gao, A.L. Zhang, and J. Thonstad, "Multifunctional sensor for use in aluminium cells". Light Metals 1995, pp.233-241.

[8] E.J Grimsey, N. Li, Y. Yan, A. Warczok, and T. Utigard, "An in-bath liquidus measurement for molten salts and slag", Light Metals 1996, pp. 1149-1154.

[9] P.Verstreken and S. Benninghoff, "Bath- and liquidus sensor for molten salts", Light Metals 1996, pp. 437-444.

Return to article index