Enabling Port Security using Passive Muon Radiography. Nicolas - PowerPoint PPT Presentation

Enabling Port Security using Passive Muon Radiography. Nicolas Hengartner Statistical Science Group, Los Alamos National Laboratory Bill Priedhorski, Konstantin Borozdin, Alexi Klimenco, Tom Asaki, Rick Chartran, Larry Shultz, Andrew Green, Richard Shirato.

Nuclear smuggling is a clear and present danger Materials Interceptions “Law enforcement officials in the US seize only 10 to 40% of the illegal drugs smuggled into the country each year Russia stops from 2 to 10% of illegally imported goods and illegal immigrants on the border with Kazakhstan” Total = 1.13 IAEA “significant quantities” (8 kg Pu or 25 kg of U 235 in HEU)

Active radiography is an established inspection technique To date, 2001 radiography has Inspection of truck with American Science and depended on Engineering backscatter x-ray system artificial sources of radiation, which bring with them a risk-benefit tradeoff 1895 First x-ray image (Mrs. Roentgen’s hand)

Passive Source Radiography: Cosmic Radiation No artificial radiation means: 1. Cars and trucks inspection without evacuating the driver significant time factor 2. Deployment abroad without local regulatory complications Detection at point of origine 3. No radiation signal to set off a 1. Neutrons salvage trigger 2. Neutrinos Minimizes inspection risks. 3. Electrons 4. Muons 5. Etc.

Cosmic-ray muons • As cosmic rays strike our upper atmosphere, they are broken down into many particle components, dominated by muons. • Muons have a large penetrating ability, being able to go through tens of meters of rock with low absorption. • Muons arrive at a rate of 10,000 per square meter per minute (at sea level).

How Muons Interact with Material L Muons are Charged either Two modes of interaction: Positive and negative Absorption Coulomb Scattering High energy: Median 3MeV

History: absorption muon radiography Luis Alvarez, 1950

Muon mapping of Chephren’s Pyramid Science , 167 , p. 832 (1970) “Search for Hidden Chambers in the Pyramids” Luis W. Alvarez et al. Alvarez et al. used only absorption, not scattering Successful experiment - existence of hidden chamber ruled out actual image with no simulated image with hidden hidden chamber chamber like the one in Cheops’ pyramid

Shadowgrams (from scattering) Possible to get shadowgrams from scattering instead of absorption Proton radiography

Basic Concept of Multiple-Scattering Muon Radiography • Track individual muons (possible due to modest event rate). • Track muons into and out of an object volume. • Determine scattering angle of each muon. • Infer material density within volume from data provided by many muons.

Scattering is Material Dependent 40 80 Mean Square Scattering (mrad 2 /cm) . for 3 Gev muons 70 Radiation Length (cm) 30 60 50 20 40 30 10 20 10 0 0 r c e ) ) ) ) ) ) e 3 6 9 2 4 2 i t t t e 1 2 2 8 7 9 s a r a = = = = = = W c Z Z Z Z Z Z l n P ( ( ( ( ( ( o m n r d n m C e o a e u p u r e t n s i I p n L g i o m a n C r u u U l T A

Prototype Los Alamos instrument Muons Tungsten Block Chamber 1 Chamber 2 Chamber 3 Chamber 4 Scintillator (temporary trigger)

Reconstruction – Localizing Scattering • Assume multiple scattering Incident track occurs at a point • Find point of closest h approach (PoCA) of incident and scattered tracks Assumed point of scatter l Assign (scattering angle) 2 to • Actual multiple scattered track voxel containing PoCA • Since detectors have known position uncertainty, signal may be spread over voxels Scattered track θ relative to PoCA uncertainty. scat • Simply add localized scattering signals for all rays.

Maximum Likelihood Image Reconstruction Maximum Likelihood Image Reconstruction Use single layer probability model to calculate the contribution of voxel j to the observed displacement of ray i. L ij : path length of ray i Develop a model of the unknown through cell j object that maximizes the likelihood that we would observe what we Λ 1 λ 2 … … λ j actually observed. Δθ M f(x,y) E-M works well: Δθ 2 Δθ 1 … λ N -1 λ N … Can handle large voxalization Compute as data comes in Δθ i Δθ i+1

First Muon Radiograph

Radiograph of another object

Clamp in z-projections

Tomographic Maximum Likelihood Tomographic Maximum Likelihood Reconstruction (20 x 20 x 20 voxels) Reconstruction (20 x 20 x 20 voxels) Objects ML reconstruction ML reconstruction 1x1x1 m3 Fe box (3 mm walls) 1 minute exposure; 1 minute exposure; Two half density Fe spheres with U sphere No U sphere (automobile differentials) Shielding of SNM works to our advantage!

Maximum Likelihood Tomographic Reconstruction Maximum Likelihood Tomographic Reconstruction 28x28x64 voxelation, 1 minute simulated data 28x28x64 voxelation, 1 minute simulated data Top View 3-D Perspective View Side View U in empty container U in distributed Fe U and car differentials Calculation time: ~2 min on a 3 GHz single-processor Windows PC

Real data from drift tubes. Δ T μ + V e - R ≈ v d Δ T T 0 T min Cylindrical Drift Tube Geometry Representative Anode Signal • High E field at 20 μ m wire causes gas • Low count rate (~kHz) and multiplicity avalanche multiplication ⇒ Relatively large cell size allowed: • e - Drift Time ≅ 20 ns/mm × R in gas: 0 ≤ Δ T ≤ 500 ns D ~ 2 inch • Larger cell size ⇒ fewer channels • Radius of closest approach given by Δ T and saturated drift velocity v d . • Spatial resolution goal ≤ 0.4 mm

Drift Tubes Bonded into Modules RCS 9/21/04

Drift tubes for muon tracking X1 μ • Potentially low cost • No fancy materials • Detector built from: Δ Z aluminum tubes tungsten wire argon gas X2

Modules combined into Muon Tracker • Drift tube detectors • 4 x-y planes • 128 tubes per x or y • 1024 channels total • Reconfigurable 3.66 m EOY 2004 Goal: 40 modules, 64” x 64” active area with good solid angle

RCS 9/21/04 Large Muon Tracker

Momentum Estimation • Measuring particle momentum increases confidence in material inference. Object Object measurement measurement • One method is to estimate area area momentum from scattering through known material. With 2 plates Δ p/p is about • 1 Plates of Plates of 50%. Momentum Momentum known known 2 N measurement measurement thickness & thickness & area area composition composition With N measurements Δ p/p • approaches:

Bonus Material

Absorbtion ⎧ Z i = 1 Absorbed Data: ⎨ Problem: ⎩ 0 Not Different physics for stoppage ∫ S = ρ ( γ ( s )) ds Than scattering. Can Stoppage We really combine data? γ P [ Z = 1| S = s , E = e ] = G ( s − e ) Model Are planning experiments to estimate H ∫ P [ Z = 1| S = s ] = G ( s − e ) F ( de ) = H ( s ) Nice little inverse problem

Secondary particle polution Physics for electron-matter interaction different from μ muon-matter interaction. + e − Drift tubes detected charged e - particles, not type. R ≈ v d Δ T e − Sources of electron: 1. Knock-off (delta-rays) Knock off electrons and 2. Bremstrallung Bremsstrallung confuses 3. In-flight decay the drift tubes (~5%)

Modeling Muon Scattering Data from scattered muons: Change in position Change in angle L Inverse problem with the signal in the [ ] [ ] Δ θ = Δ = 0 E E x variance Material specific parameter λ [ ] 1 L Δ θ ∝ Var 2 p L rad Momentum (unknown)

Point of Closest Approach (PoCA) Point of Closest Approach (PoCA) Original Approach (2003) Original Approach (2003) Assumes that the scattering took place at the point where the incoming and outgoing paths come closest

Slices through reconstructed volume

Ray-crossing algorithm cuts clutter 3 uranium blocks No contraband (20 kg each) 30 second exposure 120 second exposure 10 tons of distributed iron filling the container

Clustering algorithms to automatically Clustering algorithms to automatically search for dense objects search for dense objects • Look at significantly scattered muons • If high-Z object present, inferred locations of scattering will “cluster” • Cluster centroids are considered the candidate locations for a threat object, and passed to a classifier Input to simulation: Shipping container full of automobile differentials & one uranium sphere Identified clusters, including the real one

Candidate clusters can be tested w ith a Candidate clusters can be tested w ith a “machine machine - -learned learned” ” algorithm algorithm “ Breakthrough: Algorithm has found a good set of features based on statistics of a local, 27-voxel cube Result: Low error rates for two-minute exposures 100 95 ) 90 % ( y 85 c a r 80 u c c 75 A 70 65 0 15 30 45 60 75 90 105 120 Exposure time (sec)

Model path as an integrated Brownian motion