4th Annaal Summer Symposiam4tandards
Cooperative Investigation of Precision and Accuracy In Chemical Analysis of Silicate Rocks WILLIAJI G. SCHLECHT U. S. Geological Survey, Washington, D. C.
This is the preliminary report of the first extensive program ever organized to study the analysis of igneous rocks, a study sponsored by the United States Geological Survey, the Massachusetts Institute of Technology, and the Geophysical Laboratory of the Carnegie Institution of Washington. Large samples of two typical igneous rocks, a granite and a diabase, were carefully prepared and divided. Small samples (about 70 grams) of each were sent to 25 rock-analysis laboratories throughout the world; analyses of one or both samples were reported
‘T
IIE conventional procedure for the analysis of a silicate rock, as described by Hillebrand, Washington, and other writers, is designed for a rock having the approximate composition of common granite. I t is a happy circumstance that the procedure is suitable for such a rock; elements that interfere with the separations or are not provided for in the analytical scheme are rarely found in granites; The scheme has been found satisfactory for most silicate rocks, and has been tested by excellent, chemists over many decades. Because of its sloffness and its high cost, however, no large number of analyses were made of a single rock, from which the accuracy and precision of the met,hod could be estimated. There are no standard samples of common silicate rocks; in the absence of commercial motives, no one has been willing or able to pay for the work necessary to establish a standard sample. Except for very special kinds of rocks, analytical results are not critical in the control of any industrial process. The members of the Branch of Geochemistrjr and Petrology, C . S. Geological Survey, were therefore delighted when Harold IY. Fairbairn, Department of Geology, Massachusetts Institute of Technology, offered them a chance to join in a cooperative project with the institute and the Geophysical Laboratory, Carnegie Institution of Washington, to study the precision and accuracy of rock analyses. Felix Chayes of the Geophysical Laboratory had taken a large sample of a typical granite, to study variations in its composition, and the Geological Survey had prepared a large batch of a diabase rock, with the aim of eventually using it as a standard sample. Carefully mixed and divided samples of these two batches, weighing about 70 grams each, were distributed to 25 rock-analysis laboratories hhroughout the world, where they were analyzed by 34 chemists, all trained in rock analysis. Their first results show that, rock analyses are generally not as precise 8 s has been assumed; Figure 1 gives frequency diagrams to show the variability in reported results. The analyses are plotted a t intervals of 0.1%; each dot represents a single analyst. The “correct” results are not known a t this preliminary stage, but the patterns of results give some suggestion of the errors. They are large, in contrast to the beautiful agreements that Bright showed for the analysis of steel ( 1 ) . The patterns of the dist,ributionsare rather like those reported by Willits (IO),and their asymmetry shows, as did his results, the tendency for a given chemical procedure to give biased results. Some chemical processes are more likely to give low results than high, as shown by the silica determinations; others, like the
by 34 analysts in these laboratories. The results, which showed rather large discrepancies, are presented in histograms. The great discordance in results reflects the present unsatisfactory state of rock analysis. It is hoped that the ultimate establishment of standard samples and procedures will contribute to the improvement of quality of analyses. The two rock samples have also been thoroughly studied spectrographically and petrographically. Detailed reports of all the studies will be published.
alumina determination, give results that are more often too high than too low. ANALYTICAL PROCEDURES
After the samples were dissolved by fusion with sodium carbonate in 8 platinum crucible, silica was separated by dehydration in hydrochloric acid, followed by dehydration in hydrochloric or other acid. The silica determinations show evidence of a negative systematic error-that is, they are oftener low than high; so the actual composition is probably nearer to the high results than to the average. This is a common experience in chemical analysis, and Lundell ( 5 )gave interesting examples of collaborative analyses in which an extreme result is more nearly correct than the average. In most laboratories, the filtrate from the silica determination was neutralized with ammonia to separat,e the R20g group; alumina was calculated by subtracting Fe20s, titanium dioxide, and phosphorus pentoxide from the weight of RzO3. The alumina determination is indirect; errors in the other Rz03 constituents affect it. Total iron was determined in the R2Oa residue or in a separate sample by the Zimmerman-Reinhardt method, by reduction with hydrogen sulfide, or with a Jones reductor, and titration with permanganate or dichromate. Ferrous iron was determined similarly in a separate sample, dissolved in the absence of air; and Fe20awas found by subtracting the ferrous iron from the total iron, so the FezOa results were affected by errors in both determinations. Magnesium and calcium were determined in the filtrat,e from the RzO, separations, by precipitating the calcium as oxalate, filtering, and precipitating the magnesium from the filtrate with ammonium phosphate. The precipitates were ignited and weighed as calcium oxide and as magnesium pyrophosphate. Most analysts determined manganese by the periodate method; but eight used bismuthate. The aberrant result for the diabase was probably caused by a blunder in dilution or in calculation of dilution; the analyst used the periodate method. Alkalies were determined on a separate sample. Most analysts used the classical J. Lawrence Smith procedure of sintering with calcium carbonate and ammonium chloride. B few decomposed the sample with hydrofluoric and sulfuric acids; this seems to give just as good results as the Smith method. Most analysts determined the alkalies by weighing the mixed chlorides of sodium and potassium, then separating and weighing the potassium aa
1568
V O L U M E 23, NO. 11, N O V E M B E R 1 9 5 1 chloroplatinate. Several variat’ions were used, such as weighing the potassium as perchlorate or cobaltinitrite, weighing the sodium or the total alkali as sulfate, and weighing the sodium as uranyl zinc acetate. Four analysts det’ermined sodium and potassium with the flame photometer, with good results. All the analysts determined titania colorimetrically, in t,he R2O3 residue, by oxidizing with hydrogen peroxide to the amber compound, in the presence of phosphoric acid to reduce interference by iron. The t x o lowest results for the diabase were probably caused by blunders in dilution. \\-ith few exceptions, phosphate was separated from nitric acid solution as ammonium phosphomolybdate, and determined by weighing as the phosphoniolybdate or as magnesium pyrophosphate. >lost analysts determined total water by some version of the Penfield method, although three reported “loss on ignition,” and three by volatilization and absorpt,ion. “HzO-,” determined by loss in weight a t 105” to 110” C. was subtractedfromtotal~vaterto get the “H20+”figures. The summation of all the constituents determined is a t’raditional test,; its nearness to 100% is a necessary but not a sufficient condition for the correctness of an analysis. For example, analyses with summations near 10070 vary widely in silica and alumina, and some of the analyses with rather extreme summations have fairly good silica and alumina results. ERRORS IN ROCK ANALYSlS
Averages are not indicated in the results; they could have no meaning for the present preliminary results. Likewise tolerance ranges cannot be derived from the data statistically; these determinations are not compared with fixed standards, as are those described by Scribner (6) and by Willits (IO). Rather they are like those t,hat Bright described for steel ( 1 ) ; as he said, we cannot derive rigorous limits from them for the differences from “true content.” As Rernimont stated, “The precision of a test method cannot be stated as a single number” (9). In this set of first trials, the best description of the reliability found is the nhole set of results. For that reason, all the results will be ineluded in the complete report of this work ( 2 ) . In an>-Lase, as pointed out by \!-ernimont, the results would be hard to interpret statistically because the experiment could not be designed efficiently for that purpose. It was impossibie to control the pract,ices of the collaborators, and the aim of the first t,est was to find the variability under the conditions that now exist in rock-analysis laboratories. Rock analysts are not a closely organized group; they are far apart all over the world, and few of them are in close or constant communicat,ion. Their techniques and their personalit,iesare very individualistic. IVhat has been accomplished, then, is a reconnaissance survey of the present state of rock analysis. Some of the collaborators did not state what replication was done; several gave averages but not individual values. Few described their procedures in any detail. At this stage ranges to be expected for the various determinations cannot be given, even by using the “chemical judgment” that Bright spoke of using in judging standard samples. This work is indeed the first step in arriving a t ranges for the precision and accuracy of a rock analysis, but much work lies ahead before that aim is aceomplisbed. As Willits has stressed, collaborative work is a slow and laborious process; but Bright can testify that there is no royal road to establishing the composition of a sample without this labor. It is hoped that a follow-up to locate and remove the sources of discordance will result in the establishment of standard, improved analytical procedures, and of standard samples; if standards had been set up 50 years ago, it would now be possible for rock analysts to judge their work critically, and know what steps to take when it is not up to standard. Lundell (6) gives a short account of the kind of work necessary to approach the “true composition”
1569 of a standard; the main work is not statistical, but involves Che achievement of concordance by methods that are independentas different as possible in their chemical nature. The errors reported here are far greater than those traditionally expected in rock analysis. The tradition, however, is not backed by any extensive interlaboratory comparison, and its basis c a n only be inferred indirectly. -4s there has been no commercial demand for highly accurate rock analyses, the principal users have been research scientists-mostly geologists and petrologistswho have set rather personal, and perhaps sometimes wishful, standards. The analysts’ practice of reporting determinations to the nearest O.Olyoimplies a limit of accuracy of a few hundredths. There is a definite wish for such accuracy, among rock analysts m well as geologists and petrologists, and the fact that such accuracy was apparently established for analysis of related materials gave, added confidence that it could be attained for silicate rocks, T h e experience and judgment of competent analysts resulted in t h e set of requirements for good rock analyses published by Washington (8)and Hillebrand ( 4 )and applied in Table I to granite of t h e kind used in this study. For comparison, the table also s h o w t h e limits of error expected by a skilled modern rock analyst ( 8 ) . Even allowing for the fact that these requirements were nicarit to apply t o analyses in duplicate by the same analyst, it is doubtf u l that the best ten analysts of the present group could attain a corresponding accuracy under the conditions of the present comparison. It may well be that the analysts of Washington a n d Hillebrand’s time were as good as Washington’s limits suggest,; unfortunately we have no interlaboratory comparisons on nhich to establish the fact. If they were, the present study may stimulate modern analysts to do as well. -
Table I.
.4llow-ableVariation in Duplicate Determinations of Same Constituent in a Typical Granite
Constituent Si02 .%Oa Fez02 I‘eO CaO LlgO Sag0 Ill0 Hz0 T102 PzOa MnO
Limits of Error Washington Grovea
=to.1.5
10.10 10.05 zto.05 =t0.05 1 0 05 zto 05
10.02 10.02 10.03 10.03 1 0 02 zt0.02
1 0 01
+n
01 10.01
The limits of error shown in Table I were presumably bawd on the authors’ experience with replicate analyses made in their ov n laboratories. They were published without explicit statemcmt of their applicability to analyses made by different lahoratorics; but the implication is that such agreement should be striven f o r between laboratories. A t the time the limits were set up, analysts may not have been so conscious as they are no\y becoming t h a t differences in determinations carried out in parallel may give a misleading idea of the accuracy of the determinations. The term “duplicity” has been facetiously suggestrd for an unreliable Mtimate of error based on the agreement of duplicate deteminations made a t the same time by the same analyst; the agreement is often good because under these conditions the =me djsturb anceq can easily be introduced in both determinations. There can be a similar element of duplicity in the agreement of any analyses made by the same method; that is why a given analytical procedure cannot be considered to be completely validated until it is checked by chemically different methods (3, p. 236). It is not possible on the basis of present evidence t o decide how much of the error in rock analysis is caused by lack of skill in t h e analyst and how much is inherent in the procedures themselves. I t will require laborious and carefully designed experiments determine this. If the conventional method can give no b e t t a
ANALYTICAL CHEMISTRY
1570 GRANITE
..
. . . .... . .... . . ............. .. . . . ., . , 0
71
DIABASE
SI
... .... ..... ...... ..........
02
73
72
e
n e
..no..
,
) . . . . r 14
,
16
15
.. .. .... ..
13
.L*I q..
0 ,
15
17
16
19
. ...... ......... .. . * . . e * * . .
05
. ... ... ... ... .... ....... . .. ....... . ...... ........ ... .. ..... .... . .
0
e . ( *
1.5
1.0
65
7.5
70
8.5
80
e
Total Fs as
0 .
0 . .
1.0
30
,
0,
3
4
5
90
9.5
... .... . . .......... ,
T 10.0
SO
2t5
2.0
1.5
19
. .*e e
Fet03
,
2
I
.,
..-. . ....... ............
Fe 0
_ _ .
05
29
28
. )
~
2.0
0 . .
I8
... ". ...... . . ...."..... . .
Fe203
)
1.5
10
27
26
17
?
,ne.....
18
0
,
16
. .,.. . . .....
0
..e.
00,
15
n...
*
e
0. ....
.e.)
14
..
...oN
53
.. ..... ..... .. .......
n
W.
13
. f
52
51,
10.5
12.0
11.5
11.0
12.5
e
.e
0 .
i i..
c
,
.e.(
Q.0
0.5
0
4.5
1.0
.... .. ... ..... ...... 0 .
.e
CaO
.e.
1.0
1.5
.... .. .... ... ......
0.00
0.05
5.0
q 5.5
.. . .... . ..... ..e.
,
65
6.0
MnO
, 70
7.5
..... .... ...... .........
10.5
2.0
0 .
,
11,0
11,5
." ....... ,... ....".". . . .N
1
0+10
0.10
0,20
0.30
I
0.40
0.50
0,60
V O L U M E 23, NO. 11, N O V E M B E R 1 9 5 1
1571 DIABASE
GRANITE
.. .. .... .... ...... .........
L. 2 5 3.0
35
.... ... ..... ...... ,. ......... 0
N 020
15
4.0
.
25
2 0
.... .. ... ..... ..
3 0
e.
e.
,
0,O
.... .. .. ..
.
0.0
0.5
1.0
I5
.. .. ..
TI 08
, .**) 0 0
0.5
0 .
..... ...... - .
0 . .
? 0.5
1,5
1.0
..
.. .... e.
p2
LITERATLRE CITED
e. e .
OS
0 .
(1) Bright, H. .I.,A N ~ L .CHExf,
29..
,
0.0
0,5
0-,5
.. .. .. .. .. 0
.. . :: .... ...
.
*.
0 .
0 .
Y * 0.0
0 . .
080
0,5
. H2O t L . *
.. . ... ...... . ........ 100,5
. .. .... .(.. ..
RECEIVED August 11, 19.51
0
05
9
,
1.0
I 5
Figure 1 (Pages 1570 and 1571). Frequency Diagrams
.. .: *
Summation
........ . . :.:e
L...,*...,.
,
100.0
05
I
..e.
0.0
99.5
,
*e ..
H20-
0 .
101.0
99.5
100.0
100,5
23,
1544-7 (1951). (2) Fairbairn, 1%.IT-., et nl.. U. S. Geol. Survey. Bull. 980 (1951). (3) Groves, -4.JT., “Silicate Analysis,” 2nd ed., London, George Allen & Unwin, 1951. (4) Hillebrand, IT.F.,U. S. Geol. Survey, Bull. 700 (1919). ( 5 ) Lundell, G. E. F..IND. E s c . CHEM., AXAI..ED..5. 221-5 (1933). (6) Scribner, R. F., A s ~ L . CHEM.,23, 1548-52 (1951). (7) Shapiro, Leonard, and Brannock, JT. W..“Rapid analysis of silicate rocks.” U. S. Geological Survey circular (in preparation). (8) Washington, H. S., “The Cheni;cal .Inalysis of Rocks.” 4th ed., Sew York, John Kiley & Sons, 1930. (9) M-ernimont, Grant, ASAI.. CHEM., 23, 1572-6 (1951). (10) Willits. c‘. O., I h i d . , 23, 1565-7 (1951).
e.
..... .....
results than reported here, it niight perhaps be abandoned in favor of rougher rapid methods of the kind developed by Shapiro and Brannock in the U. S. Geological Survey, using volumetric and photometric techniques ( 7 ) . There is much reason to hope, however, that the conventional method can be made to give acceptable results. Many questions need t o be examined. The desire t o use a single sample for determinations of several elements may sometimes lead to unnecessary difficulty in analysis. John P. hfarble is accustomed to point out that a n analytical procedure giving accurate results for a given rock may be inaccurate unless modified for use with rocks of moderately different composition. With modern reagents, a direct determination of alumina in rocks ma>be practicable. Flame photometrj offers the possibility of increased accuracy and less labor in the determination of alkalies. With the establishment of satisfactory standard procedures in prospect, the next stage in this cooperative investigation should yield results that are amenable t o statistical planning and control, and ultimately the accuracy of rock anal\-ses can be estimated with a more realistic confidence than in the paqt,
101.0