Muon Data Analysis Workshop: An Introduction to the Practicalities

Transcription

1 Muon Data Analysis Workshop: An Introduction to the Practicalities Francis Pratt (ISIS) ISIS Muon Training Course March 2012

2 Outline A. Acquiring the data B. Location and format of the data C. Data analysis software D. Basic msr data analysis 1) Setting up 2) Analysing the relaxation 3) Modelling the fitted parameters 4) Plotting results E. Analysis of complex rotation spectra 1) Fourier transform 2) All-poles transform 3) Maximum entropy method 4) Time domain fitting

3 A. Acquiring the data Currently we have two different acquisition systems: SECI Windows system (MUSR, EMU, HiFi) MACS/EXP Windows/Linux system (ARGUS)

4 SECI (MUSR,EMU and HIFI) Main control is by command line and file-based scripting

5 MACS/EXP EXP MACS Hybrid Linux/Windows system used on ARGUS

6 MACS GUI control and dynamic scripting

7 B. Location and Format of Muon Data Finding the Muon Data: ARGUS From the ISIS network: \\ndw656\macs\nexus\argus nxs EMU From the ISIS network: \\emu\data\emu nxs MUSR From the ISIS network: \\musr\data\musr nxs HIFI From the ISIS network: \\hifi\data\hifi nxs Temperature Logs: NeXus files store the T log data internally (other logs may also be stored, e.g. fields)

8 Taking Data Away Copy to own laptop or USB stick Burn CD using PC in instrument cabin Send to remote system back home pushing is easier than pulling! Fetch from home using the data access portals: Old data up to Aug 2010: data.isis.rl.ac.uk Data after March 2011: data.isis.stfc.ac.uk

9 Formats for ISIS Muon Data All new data is currently saved in NeXus format SECI NeXus-1: hierarchical, extendable format (.nxs), soon Nexus-2 also RAL binary format (.ral) for older data MACS Nexus-2 (.nxs) plus RAL binary format (.ral) Typical run file is kb in size Data compression can reduce the size by up to a factor of 7 (bzip2 is the most efficient zip algorithm for muon data)

10 The NeXus Data Format Based on the NCSA Hierarchical Data Format (HDF) The NeXus hierarchical data format has three components: A set of subroutines to make it easy to read and write NeXus files A set of design principles to help people understand what is in them A set of instrument definitions to allow the development of more portable analysis software Example part of data structure: NeXus webpage: HDF webpage: ISIS NeXus muon definitions:

11 msr Data Formats in Use Worldwide RAL/RIKEN-RAL (UK).nxs (Nexus-1 and Nexus-2) PSI (Switzerland).bin (standard binary) PSI-root (based on CERN format).nxs (Nexus-2, eventually) TRIUMF (Canada).tri (VMS binary).mud (hierarchical) JPARC (Japan) JPARC-root (based on CERN format)

12 C. Muon Data Analysis Software Software used at ISIS: VMS Legacy: UDA RUMDA MESA (VMS, General Purpose) (VMS, General Purpose) (VMS, Maximum Entropy for TF Studies) Current PC based software: WiMDA Mantid (Windows, time and frequency domain) (based on multiplatform ISIS toolkit, time domain) Many research groups have also developed their own programs over the years

13 Both WiMDA and Mantid will be available for use with the practical sessions WiMDA can be downloaded from: Mantid can be downloaded from:

14 D. Basic msr Data Analysis Main Stages: 1. Preparing the data to be analysed, setting up 2. Fitting the measured asymmetry to a chosen relaxation function; analysing 3. Assessing the fitted relaxation parameters, which may involve a further stage of fitting these parameters to an appropriate model; modelling 4. Preparing plots of the results of analysing and modelling the data, plotting

15 D1. Setting Up the Data a) Checking the time origin t 0 and the time of the first and last good data points b) Defining the detector grouping c) Correction for counting loss due to counter deadtime d) Correction for steady background count rate e) Choice of binning

16 a) Checking the time origin t 0 and the time of the first and last good data points

17 b) Defining the detector grouping e.g. for LF/ZF : Forward group 1-16 Backward group Notes: for TF : Group1 1-4, Group2 5-8, Group3 9-12, Group , ARGUS has 192 detectors MUSR and HIFI have 64 detectors EMU has 96 detectors (In WiMDA the detectors are usually pregrouped into 32 histograms)

18 How many groups to use for TF analysis? A dephasing effect will reduce the asymmetry of TF data if not enough groups are used: N=4 N=8 Dephasing factor = sin(p/n) / (p/n) TF Groups Dephasing Factor % 8 98 % 4 90 % 2 64 % i.e. 8 TF groups are sufficient for most purposes

19 c) Correction for counting loss due to counter deadtime characterised by a deadtime t for each detector channel, typically t ~ 10 ns deadtimes for particular instruments are obtained from high statistics calibration runs using Ag data rate correction to the observed rate r ob is applied to give the true rate r the simplest form of correction is r = r ob / (1- r ob t) Less important for instruments with large numbers of detectors

20 d) Correction for steady background count rate can be included as part of the fitting procedure Not usually needed for ISIS data analysis

21 e) Choice of binning: The standard raw time bin for ISIS data is 16 ns. It is often useful to choose to increase the bin size for data analysis bin width : trade-off between number of points and fitting speed allows separate focus on fast and slow parts of the relaxation fixed/variable: variable binning compensates for the deteriorating signal-to-noise at longer times best to keep to fixed binning for weakly damped oscillations, e.g. TF studies

22 D2. Analysing the Relaxation Try fitting to possible alternative relaxation functions Look for systematic deviations of the fit from the data are additional relaxation components needed? Use the reduced chi-squared c r 2 to judge the quality of the fit and appropriateness of the model c r N i 1 y y 2 i 1 2 i ( x ; p, p.. p ) i m 2 N m N fitting points and m fit parameters N-m = n is the number of degrees of freedom

23 Expected standard error of c r 2 is (2/n) 1/2 High n Low n

24 Expected standard error of c r 2 is (2/n) 1/2 Quality of fit Poor fit High n Low n

25 Expected standard error of c r 2 is (2/n) 1/2 Quality of fit Good fit High n Low n

26 Expected standard error of c r 2 is (2/n) 1/2 Quality of fit Too good fit! High n Low n

27 Expected standard error of c r 2 is (2/n) 1/2 Quality of fit Difficult to distinguish High n Low n

28 D3. Modelling Fitted Parameter Sets A further stage of fitting involves modelling the fitted parameters for a related set of runs: e.g. following the temperature dependence of a precession frequency within a magnetic phase or fitting the field dependence of a relaxation rate to an appropriate model Online analysis and modelling allows feedback to the experimental data taking process helps in getting more efficient and complete data sets

29 D4. Plotting the Results Close integration with the fitting process is desirable for rapid feedback on: the data quality the state of the analysis the progress of the experiment GLE (Graphics Layout Engine) Graphics scripting system closely integrated with WiMDA and used in the data analysis workshop for making plots Origin and Excel are also available for further processing and plotting of results

30 E. Analysis of Complex Rotation Spectra 1. Fourier transform 2. All Poles transform 3. Maximum Entropy spectral analysis 4. Time domain analysis versus frequency domain analysis

31 Fourier and All-Poles Transforms FFT (Fast Fourier Transform) is the standard way to convert from time domain to frequency domain. FFT assumes frequency spectrum is well represented by array of evenly spaced points, which works well for spectra containing broad spectral features. However, if the spectrum contains very narrow features, other types of frequency transform can work better. The All-Poles (maxent) transform method is one such method, which makes an expansion of the data in terms of a series of sharp frequencies See Press et al, Numerical Recipes, CUP for further details of the All-Poles transform Problems with Transform Methods in msr All transform methods assume that the data error is independent of time, clearly not the case for msr data. Data filtering (apodization) is essential before transforming, but this produces significant spectral broadening.

32 Fourier and All-Poles Transforms Optimal filtering time constant for a single undamped test frequency

33 Fourier and All-Poles Transforms A close pair of undamped test frequencies

34 The Maximum Entropy Method Avoids the noise problem and need for filtering; takes data errors fully into account Iterative procedure for constructing the frequency spectrum with the minimum structure (i.e. maximum entropy) that is consistent with the measured data Entropy here is determined from the frequency spectrum p k S k p k log b p b k The procedure involves maximising S l c 2, where l is a Lagrange multiplier See Rainford and Daniell, Hyperfine Interactions 87, 1129 (1994) for a detailed discussion of using Maximum Entropy in msr for a general reference see: Maximum Entropy in Action, Buck and Macaulay, OUP (1991)

35 The Maximum Entropy Method Demonstration of MaxEnt using the test data for the transforms

36 Organic Superconductor Example Maximum Entropy Spectra Characteristic field distribution due to vortex lattice

37 Comparison of Frequency Domain Methods CdS Data Central frequency plus split pair Reconstruction methods Transform methods

38 Time Domain Analysis versus Frequency Single Frequency Domain Analysis Freq (MHz) Width (MHz) Test Data Time domain fit (1) 0.001(1) Maximum Entropy Pair of Frequencies Freq (MHz) Width (MHz) Test Data , , Time domain fit (1) (3) 0.003(3) 0.004(3) Maximum Entropy

39 Time Domain Analysis versus Frequency Domain Analysis Transforms are good for determining a qualitative picture of data: FFT best for spectra containing relatively broad features All-poles transform best for spectra composed of sharp features Iterative Maximum Entropy Method gives an unbiased view of the data but Time Domain Fitting gives best ultimate accuracy, provided the correct model is being used. CONCLUSION A combination of Frequency Domain and Time Domain analysis usually works best!

40 Next: Hands-on part of the Data Analysis Workshop