Enhanced radiological nuclear inspection and evaluation

Abstract

A method of detecting a potentially radiologically hazardous material in a vehicle comprises inputting results of a radiological scan of a vehicle, performed by one or more detectors of a radiation portal monitor (RPM). The method further comprises inputting vehicle context information related to the vehicle. The vehicle context information includes information related to the vehicle other than the radiological scan results data. The method further comprises extracting a feature set from the radiological scan results data and the vehicle context information, and performing an investigative classification with an investigative classifier machine learning model based on the feature set. The investigative classification results in the vehicle being classified as recommended or not for further investigation for possible presence of a radiological hazard. The method further comprises generating an alert indicative of the result of the investigative and causing the alert to be output to a user.

Description

ENHANCED RADIOLOGICAL NUCLEAR INSPECTION AND EVALUATION STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH

[0001] The techniques introduced here were made with Government support under Contract No. DE-AC52-07NA27344 awarded by the United States Department of Energy. The Government has certain rights in the invention. CROSS-REFERENCE TO RELATED APPLICATION(S)

[0002] This application claims priority to U.S. Patent Application No. 63 / 694,680 filed on September 13, 2024, which is incorporated herein by reference in its entirety. TECHNICAL FIELD

[0003] The techniques introduced here generally pertain to radiological nuclear inspection and evaluation, and more particularly, to an improved technique for radiological nuclear inspection and evaluation of vehicles passing through an inspection portal. BACKGROUND

[0004] Radiation Portal Monitors (RPMs) have long been deployed to scan vehicles and cargo for the presence of radiological and nuclear materials. While an important tool for finding lost “orphan” sources, in recent decades large numbers of RPMs have been deployed to screen cargo shipments and vehicles for the presence of radiological and nuclear materials that could be used to cause catastrophic harm.

[0005] In operation, the sensitivity of an RPM is determined by its alarm threshold. A low alarm threshold increases sensitivity but also increase nuisance alarms. Nuisance alarms are primarily caused by cargo with benign but measurable radioactivity from Naturally Occurring Radioactive Material (NORM). Radionuclides used for medical and industrial applications can also trigger alarms and are typically required to be inspected to ensure a radiological hazard such as a radiation dispersal device (dirty bomb) is not present. The alarm threshold needs to be set such that the number of nuisance alarms is within what the system operator can process, typically withSecondary inspections are laborious, time consuming, costly, and they can impede the flow of commerce.

[0006] Reduction of nuisance alarms without compromising sensitivity is a technical challenge. Nuisance alarm reduction can be achieved with higher performance detector materials, but such materials are generally expensive and retrofits or replacement of the instruments can be disruptive and costly. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] One or more embodiments of the techniques introduced here are illustrated by way of example and not limitation in the figures of the accompanying drawings, in which like references indicate similar elements.

[0008] Figure 1A shows illustrates an example of an environment in which an Enhanced Radiological Nuclear Inspection and Evaluation (ERNIE) system can be implemented.

[0009] Figure 1B shows illustrates an example of an environment in which the ERNIE system can be implemented, with separate computer systems used for training and classification.

[0010] Figure 2 illustrates an example of the elements of the ERNIE system.

[0011] Figure 3 is a flow diagram illustrating an example of the overall classification process that may be performed by the classification subsystem in the ERNIE system.

[0012] Figure 4 illustrates an example of the outputs of radiation detectors and vehicle presence sensors (VPS) for a particular vehicle carrying a radiation source.

[0013] Figure 5 illustrates graphically an example of how a detected radiation profile can be evaluated against a background suppression template for vehicle to produce more reliable radiation source detection.

[0014] Figure 6A shows conceptually an example of a single decision tree with illustrative feature split thresholds and resulting class distributions, and illustrative values according to an embodiment of the ERNIE technique.

[0015] Figure 6B shows a more specific example of a decision tree learning process, with illustrative values according to an embodiment of the ERNIE technique.supervised learning (SSL) process for training a machine learning model according to an embodiment of the ERNIE technique.

[0017] Figure 8 illustrates an example of a training process for training a machine learning model according to an embodiment of the ERNIE technique.

[0018] Figure 9 illustrates an example of an overall classification process for classifying vehicles and radiation sources according to an embodiment of the ERNIE technique.

[0019] Figure 10 is a high-level block diagram illustrating an example of a computer system in which some or all of the ERNIE technique can be implemented. DETAILED DESCRIPTION

[0020] In this description, references to “an embodiment”, “one embodiment” or the like, mean that the particular feature, function, structure or characteristic being described is included in at least one embodiment of the techniques introduced here. Occurrences of such phrases in this specification do not necessarily all refer to the same embodiment. On the other hand, the embodiments referred to also are not necessarily mutually exclusive. 1. Overview

[0021] Introduced here is an improved technique for detecting a potentially radiologically hazardous material in a vehicle passing through an RPM. In at least some embodiments, the technique is implemented in software, using machine learning techniques. The technique provides increased sensitivity to radiological and nuclear materials of concern while reducing occurrences of nuisance alarms. The technique introduced here is alternatively referred to herein as Enhanced Radiological Nuclear Inspection and Evaluation (ERNIE), “ERNIE,” “the ERNIE technique,” “the detection technique” or “the technique.” A system that implements ERNIE is referred to herein as “the ERNIE system.”

[0022] In this description, to facilitate explanation, ERNIE is described in the context of using radiation detectors to detect a potentially hazardous radiological material. Note, however, that in other embodiments the core aspects of the ERNIEof detectors. For example, aspects of the ERNIE technique potentially can be used to detect chemical or biological hazardous materials.

[0023] The ERNIE technique breaks down the task of reducing nuisance alarms without sacrificing sensitivity by formulating the task as a machine learning based classification problem for a given vehicle passing through a set of fixed detectors of an RPM. The outcome of the classification is to classify the results of scanning the vehicle as either “investigate” or “release,” and this outcome or an indication thereof is output to a user / operator of the system. An outcome of “investigate” is an alarm condition, indicating that the scanned vehicle may be carrying a hazardous material and is recommended for additional inspection (e.g., with handheld detectors). An outcome of “release” means the scanned vehicle is unlikely to be carrying a hazardous material and is safe to release from the inspection point.

[0024] In some embodiments, the ERNIE technique uses vehicle context information in addition to radiological scan results, to perform the classification. The vehicle context information can include, for example, the type, size and / or speed of the vehicle, as determined by a vehicle presence sensor (VPS) subsystem system located at the RPM. The vehicle context information can be used, among other purposes, to identify an appropriate background radiation suppression profile (BRSP) (or “template”) of a particular vehicle type passing through the portal, and that BRSP can then be used in the classification process to subtract an estimate of the background so as to perform a more accurate classification.

[0025] In some embodiments, the classification process includes three separate classifications: an anomaly classification, a source type classification, and an “investigate” classification such as mentioned above. Anomaly classification is performed to produce an anomaly score indicative of whether the input data from the detectors and / or VPS subsystem is anomalous to the degree that any outcome of the overall classification process would be unreliable. The source type classification is used to identify the type of radiological source (if any is detected), since certain types of sources are more likely to be dangerous and / or indicative of “bad actors” than others. Indications of the outcomes of these additional classifications can also be output to the user.uncertainty associated with the source type classification by applying the source type classification to a source type uncertainty classifier machine learning model. The technique then determines that the source type classification cannot be determined uniquely and could be one of several likely source types, or cannot be determined at all. Uncertainty is determined from the uncertainty classifier output satisfying a specified criterion, e.g., when it exceeds a certain threshold.

[0027] In some embodiments, the investigative classifier machine learning model is trained on training data that include motion profile (e.g., speed as a function of time) information and vehicle type information associated with a plurality (preferably many) vehicles of different types. The motion profile information can be extracted from the full VPS measurements, one or more velocities extracted from the VPS measurements, other vehicle motion tracking methods (e.g. LiDAR or video), or based on historical distributions of motion of vehicles that have passed through the portal. The investigative classifier machine learning model can also be trained on training data that include injection of a plurality of source types into each of a plurality of vehicle types. Multiple source types of interest can be represented in the injected labeled data by semi- synthetic measurement-based simulations. Injection can be achieved by modeling the source emissions as a function of time, accounting for scattering and absorption through intervening materials as the distance and angle to the detectors change, and adding the expected number of counts (with a Poisson random distribution) to measured RPM scans of vehicles that are determined to be unlikely to include radioactive materials.

[0028] In some embodiments, the investigative classifier machine learning model initially is trained by labeled data corresponding to each of a plurality of source types of interest in a plurality of locations in the vehicle, and is subsequently trained incrementally in a plurality of additional training cycles with additional unlabeled data labeled by a semi-supervised learning process. In such embodiments, the initial labeled training source data may include only injected data (modeled source data added to measured background data), and do not include measured data with source emissions. Further, the injected data may be biased by position within a vehicle as a function of source type.

[0029] In some embodiments, the ERNIE technique trains and uses at least two separate machine learning models for the “investigate” classification, e.g., one for use when reliable VPS data are available and another for use when no reliable VPS data are available. For a given vehicle passing through an RPM, the ERNIE technique first determines whether VPS data are available for that vehicle and satisfies a specified reliability criterion. It then selects the appropriate machine learning model (VPS or no- VPS) for use in the investigative classification based on whether (reliable) VPS data are available.

[0030] In accordance with the above features, training of the investigate machine learning model for use with VPS data includes training with vehicle context information, including motion profile information and vehicle type information (from VPS data), for a plurality of vehicles. Many other features of the ERNIE system, according to various embodiments, will be apparent from the description which follows. 2. Environment for ERNIE Implementation

[0031] Figure 1A illustrates an environment in which the ERNIE system can be implemented. The ERNIE system 100 includes two major subsystems: a classification subsystem 101 and a training subsystem 102. The classification subsystem 101 performs the classification tasks described herein, whereas the training system 102 trains the machine learning models that are used by the classification system 101. The classification subsystem 101 and the training subsystem 102 can be, but are not necessarily, implemented on physically / functionally separate computer systems, which may communicate with each other over a network 103, as shown in Figure 1B. The network 103 (if any) may be, for example, the Internet, a wide area network (WAN), a local area network (LAN), or combination thereof. Accordingly, the machine learning models used by the classification subsystem 101 may be (but are not necessarily) trained in a separate computer system (i.e., training subsystem 102) located at a separate location from the classification subsystem. Alternatively, in some embodiments the classification subsystem 101 and the training subsystem 102 may be implemented in the same physical computer system.

[0032] In operation, a vehicle 104, which may be carrying a radiation source 105, passes through a monitoring portal 106. The portal 106 is equipped with an RPM including one or more radiation detectors 108 (eg gamma ray detectors and neutronare known in the art. The detectors 108 are communicatively coupled to provide their outputs to the classification subsystem 101. 3. ERNIE High-Level Methodology

[0033] The task of reducing nuisance alarms in RPMs without compromising sensitivity to sources of concern can be formulated as a classification problem. NORM can be differentiated from other sources by two key characteristics: spectral signatures and spatial extent. Spectral signatures are dominated by the presence of potassium (which contains40K), thorium (232Th), and natural uranium and its decay products (e.g.,226Ra). NORM materials have low specific activity but can produce measurable signatures that trigger alarms when considerable quantities of these materials extend over a large area. Thus, in RPM scans, NORM emissions are typically more spatially extended than other more compact sources. While significant, these differentiating signatures for NORM vary significantly. Further, the signatures from sources of concern also have large variations. While there are a limited number of radionuclides involved, including radiological materials (high activity192Ir,60Co,169Yb,75Se,99Mo, and137Cs), and fissile materials (particularly235U and239Pu), the spectral signatures from each of these radionuclides will change with the distribution of materials between the source and detectors. Intervening materials will absorb and scatter radiation from the source, and the observed signatures will depend strongly on the composition and distribution of these intervening materials. In cargo shipping, intervening materials include possible shipping containers for the radioactive source, other cargo materials, and the structure of the shipping container and vehicle, all of which can vary significantly. An alarm algorithm should be able to operate with the full range of NORM signatures found in the field and the full range of measurable signatures from sources of concern.

[0034] This presents a challenge for many approaches to alarm algorithm design. Alarm algorithms based purely on intensity (e.g., N-Sigma threshold) are not capable of detecting sources of concern with intensity lower than NORM sources. Those based on common spectral analysis techniques such as energy window ratios and spectral template matching are not able to account for all the variations and spectral overlaps between sources of concern and NORM. Heuristic approaches to the spatial distributions are complicated both by the large variation in the spatial profiles presentedmany different vehicle types and cargo loading shadowing the detectors from background radiation. These variations in background shadowing also make background subtraction problematic.

[0035] As a classification problem, the differentiation between NORM and sources of concern can be handled by use of machine learning techniques. In supervised machine learning analysis, a model is trained by exposure to large numbers of labeled examples of data, and the trained model is then applied to assess measurements in the field. The machine learning model is not limited to simple thresholds of intensity, intensity ratios, or matches to templates, but rather it can combine all these factors with many more potentially informative features of the data to return a reliable score. Previous works using machine learning have not had access to the huge variations in NORM signatures found in large numbers of RPM scans and have not employed a full distribution of signatures from threat material configurations. Both of those concerns are addressed by ERNIE.

[0036] ERNIE approaches RPM alarm analysis as a classification problem to differentiate between scans that need further investigation and those that can be released. Scans that need further investigation generally are those that may include nuclear or radiological materials that could pose a threat to health or safety. Such scans are referred to as potentially representing threat sources, or an “Investigate” class. Scans that can be released are those which only include background radiation (i.e., non-emitting radiation) and NORM, and they form the benign “Release” class. In some embodiments, sources used in medical and industrial applications may be included in the Investigate category, but the framework developed here can be applied to any categorization needed. For each scan, ERNIE estimates an “Investigate Score” that is compared against an alarm threshold, which is set using historical data to trade off between threat sensitivity and nuisance alarm rates. 3.1. Data Acquisition

[0037] The ERNIE technique can fuse data from many aspects of the measurements available. An RPM system can produce time histories of gamma-ray and neutron count rates. Some RPM systems provide gamma-ray counts in different energy bins. The ERNIE technique in one embodiment can be applied to systems withIn some embodiments, the ERNIE system is implemented as machine learning based software that runs on top of, or in communication with, RPM native (non-machine learning based) radiation source detection software. In some such embodiments, the ERNIE technique does not include neutron measurements in the machine learning analysis, since neutron alarms are infrequent in practice and would often require a full secondary inspection regardless of any additional processing. Instead, neutron alarms produced by the RPM native (non- machine learning based) detection software are monitored and included as an alarm category in the ERNIE system.

[0038] An RPM system can also include VPS typically including one or more pairs of an infrared (IR) beam emitter and sensor. In one embodiment the beam emitters are placed on one side of the portal and the sensors on the other, so that when a vehicle enters the portal, the IR beam is broken. An example of the VPS configuration is discussed further below in relation to Figure 3. The VPS provides the time and duration of the vehicle’s passage through (or occupancy within) the monitoring portal, and an estimate of the entry and exit speeds of the vehicle. When the full VPS data stream is available, the ERNIE system can use all of the sensor signal transitions available to estimate a more reliable initial velocity and acceleration. When more than two reliable transitions are available, the system can also estimate the jerk (the time rate of change of the acceleration) to provide a reliable third-order motion profile. The motion profile can include a vehicle length estimate, and when the full VPS data stream is available, the system can use the pattern of upper and lower beam breaks to classify the vehicle type as one of many categories, ranging from motorcycles to double semi-trailer trucks.

[0039] Standard analytic methods are not well suited to fusing data from many disparate measurements such as those described herein. Furthermore, variations in cargo, cargo loading, vehicle types, and vehicle motion all can impact measured data, and simple models of background and source signatures do not capture the full range of measurements. The ERNIE technique, therefore, employs a machine learning approach where a machine learning model is trained by exposure to examples. To achieve reliable results with significant nuisance alarm reduction and improved threat sensitivity, one should use training data that span the full range of measurements that are actually observed and those that could be observed. To capture sufficient variations in measurements that are actually observed, it may be desirable to select more thanbe included in a given ERNIE model without any selection due to measured radiation. It further may be desirable to use a separate SOC data sample of about 120,000 scans for training. When possible, it may be desirable to collect these training data from a full year of operations, so that seasonal variations in weather and commodities can be captured.

[0040] To capture sufficient variations in measurements that could be observed, the ERNIE technique relies on radiation models and simulated semi-synthetic measurements. It probably is not practical to measure all of the source types in all possible cargo loadings that could be created and encountered in the field. Also, measurements with different RPM units of the same model in different locations can produce different results due to different backgrounds, settings, and condition of the instrument. Therefore, the ERNIE technique uses semi-synthetic data created by injecting modeled source emissions into background-only measurements and can use available controlled source measurements (measurements of vehicles with sources intentionally included with known source location and shielding configurations) to validate models and simulated measurements. Additional controlled source measurements can be conducted with the ERNIE system deployed in the field to validate the full system operation. 3.2. Training -- Data Augmentation

[0041] Controlled source measurements can be augmented with simulated semi- synthetic measurements to create datasets that span the full range of measurements that could be observed. In one embodiment, the following set of training procedures and tests are used to ensure that the simulated semi-synthetic measurements accurately and comprehensively reflect measurements that could be observed:

[0042] 1) To calculate the radiation emissions, the system uses multiple modeling tools, including Gamma Detector Response and Analysis Software (GADRAS) (see, e.g., D. J. Mitchell, “Gamma Detector Response and Analysis Software (GADRAS),” Sandia National Laboratories, SAND88-2519, 1988) and MCNP (see, e.g., C. J. Werner, et al., “MCNP6.2 Release Notes”, Los Alamos National Laboratory, report LA- UR-18-20808, 2018) to create and validate the models. Radiation generated by radionuclides of interest can be modeled individually and in groups.scattering and absorption of the emerging radiation emissions as they pass through a wide range of intervening materials are then modeled, including materials that may be packed with the source (such as shipping containers, or other configurations of materials that shield the radiation), and also randomly selected cargo and vehicle wall materials.

[0044] 3) The modeled radiation signals measured by the detectors are injected into a non-emitting (background-only with background suppression from the cargo) vehicle measurement selected from the SOC samples. Source injection is achieved by calculating the mean number of source counts expected in the measurement, generating a random realization of that number with a Poisson distribution, and adding those gamma-ray or neutron counts to the measured data. This produces measurement-based semi-synthetic data that reflect field measurements with very high fidelity.

[0045] 4) The training subsystem randomly selects a location for the source in the vehicle and use an estimated motion profile from the VPS to calculate the line of sight between the source and each detector panel at each measurement time (typically collected at 10 Hz). The ERNIE system selects from a distribution of likely motion profiles for training and use the best estimate for testing, thus reducing overtraining. For some source classes, the location of the source within the vehicle can be biased to provide more semi-synthetic measurements in regions of the vehicle where that class of source would be more likely to appear. For example, medical sources are more likely to be located in the passenger region, and radiological and nuclear materials (which can be very heavy) are more likely to be found in the cargo area of the vehicle. To ensure full coverage of all measurements that could occur, all source classes can be made to have some probability of being injected in all locations within the vehicle.

[0046] 5) For the initial training, semi-synthetic measurements of every source class of interest is produced by injection. Measured data of sources are not used in the initial training. This helps ensure that machine learning is differentiating between source classes, not between semi-synthetic and unaltered measurements.

[0047] 6) Each non-emitting SOC scan is used multiple times in the training dataset, including once with no modifications and again with injection of each sourcesample with a single source class.

[0048] 7) SOC scans used for training are not used for testing and evaluation. To allow proper cross-fold validation of machine learning models, folds are designated such that each SOC scan used for injection and training only appears in one fold (subset) of data. Therefore, in at least some embodiments, testing and validation are never performed with SOC scans used in training, allowing for properly conservative evaluation and reducing the risk of model overfitting.

[0049] 8) A two-step training process can be employed. The initial training step uses the labeled training data where all source classes are represented by the semi- synthetic measurement-based simulations. The second step uses Semi-Supervised Learning (SSL) to incrementally label additional SOC data and then include the new labeled data in the subsequent model training cycles. This process mines the field measurements for additional informative signatures and patterns found in broad ranges of SOC measurements. The data used in SSL does not require labeling by human experts, making the inclusion of such data and training of well-informed models affordable.

[0050] 9) The system uses feature extraction to create the attributes presented to the machine learning algorithm. The features are variables calculated from the measured data that reflect characteristics of the source classes. The statistics and physics-based features are designed to focus on the radiation signatures of interest and avoid any signatures that could be used to differentiate between the semi-synthetic measurements and field measurements. This feature extraction step is useful to further ensure that machine learning, in its quest for the most performant models, does not leverage any “tells” that may unintentionally differentiate samples based on how they were generated rather than their actual source class. The features are also designed to have good statistical accuracy, as they sum large portions of the spectrum rather than relying on individual spectral bins. 4. System Description

[0051] Figure 2 illustrates the ERNIE system in greater detail, according to one embodiment. In the illustrated embodiment, the classification subsystem 101 includesVehicle / cargo data records 203 representing scans of vehicles passing through an RPM are input to the feature extraction module 201, which extracts features from the data records 203 and provides the extracted features to the ML prediction module 202. The ML prediction module 202 performs the classification analysis for each vehicle passing through the portal, using a trained machine learning (ML) model stored in ML model storage 206, and generates the user interface 204 that indicates to a user 205 the results of each scan.

[0052] The major components of the training subsystem 102 include a model injection module 221, a feature extraction module 222 (which includes the same code as feature extraction module 201), a feature table generation module 223, an ML training module 224, an ML diagnostics and evaluation module 225 and an ML prediction module 226 (which includes the same code as ML prediction module 202). The training subsystem 102 uses the same algorithms as the classification system for feature extraction, ML trained model and ML prediction. The ML training module 224 trains the ML models based on one or more feature tables stored in feature table storage 227, which have been generated by the feature table generation module 223 based on features extracted by the feature extraction module 222 and based on source models 230 injected by the model injection module 221. The ML diagnostics and evaluation module 225 uses the prediction module 226 to perform ML prediction based on the stored feature tables 227 and the ML model 206. The ML diagnostics and evaluation module 225 also evaluates and scores the performance of the ML models allowing the operator to determine the alarm rate and source sensitivity the model achieves, and to set an investigate score threshold. Performance reports produced by the ML diagnostics an evaluation module 225 are stored in ML performance reports storage 229.

[0053] As shown in Figure 2, the ERNIE system according to at least some embodiments uses field measurements that are augmented when needed by injection of source models 230. A feature table is generated from these data. Each row of the feature table represents the features from a single RPM scan of a particular vehicle. Each column represents a different feature, with additional columns being included for metadata about the vehicle. The system may use a combination of measurement datadata.

[0054] The feature table is then divided into disjoint sub-tables (folds) that are used to train, test, and validate a machine learning model, as described further below. In at least some embodiments, no data derived from a particular field measurement is present in more than one fold. Testing portions of the feature table are used to produce diagnostics and performance evaluations. In operation, measurements from each RPM scan are sent through the same feature extraction code and analyzed with the same machine learning prediction code and machine learning model used in training. Results are provided to the user via a user interface. The dashed line 232 from the user 205 to the data record classification 231 represents an optional feedback loop that allows the user 205 to label data found in operation and use it in retraining the ML models. This is particularly useful if the user finds examples in operation that were not well represented in the training developed from earlier measurements and / or at different locations.

[0055] Figure 3 shows an example of the overall classification process performed by the classification subsystem 101 in an embodiment of the ERNIE system. The process 300 may be performed by the ML prediction module 202 (Figure 2) in the classification subsystem 101.

[0056] For any particular RPM, and potentially for any particular vehicle, reliable VPS data may or may not be available. Accordingly, either of two types of feature sets can be provided as input to the process for any given RPM or, potentially, for any given scan: a feature set that was calculated using VPS data or a feature set that that was calculated without VPS data. Correspondingly, the ERNIE system 100 in an embodiment can include two types of ML models: a ML model trained based on feature sets that use VPS data and another ML model trained on feature sets that do not use VPS data. Accordingly, at step 301 the process 300 inputs a feature set and selects the corresponding ML model based on whether the feature set uses VPS data or not. Next, at step 302 the process applies the anomaly classifier, which generates an anomaly score, Ascore, for the input data and determines whether the Ascore exceeds a specified threshold, Athreshold. If the anomaly score exceeds the threshold, then the process 300 proceeds to step 308, in which the scan is found to be anomalous, and theand neutron alarm algorithms used prior to ERNIE installation, and the scan is flagged as an “Anomalous Scan” at step 308. An abnormally large number of anomalous or irregular scans can alert the operators to malfunctioning equipment or other operational problems.

[0057] If the scan is not found to be anomalous at step 302 (i.e., Ascore is below Athreshold), the features are then processed by an Investigate classifier at step 303. At step 303, if the score, Iscore, produced by the Investigate classifier exceeds the investigate threshold, Ithreshold, then at step 307 the detected source is determined to be a threat source and an investigate alarm is generated. The source class is then identified at step 304 as the source class with the highest source class score among the investigate classes (of particular interest is anything other than non-emitting or NORM). If the Iscore does not reach Ithreshold at step 303, the source is released at step 306 (deemed not to be a threat) with a source class based on the highest score within the release classes. The source class selection also includes input from the source class uncertainty classifier at step 305, which selects more than one source class output in cases where the data do not support a unique source class determination. In some cases, the data may not support any source class determination, in which case the uncertainty classifier will return a source class of “unknown.” The investigate / release decision of step 303 and the source class determination from step 304 and 305 are presented in the graphical user interface 204 (Figure 2), along with estimates of the source’s position within the vehicle determined by the source location features most consistent with the source class determination. 5. Detailed Methodology

[0058] In some embodiments, for each data record, ERNIE produces a feature vector that characterizes the energy, intensity and spatial distribution of any radioactivity within the cargo of a vehicle. Tables of the feature vectors are created for training and testing machine-learning models. In one embodiment, the approach to the machine learning process is to create two different sets of tables. The first set is the training set, which is used to determine the optimal decision criteria. This set is subdivided into folds (subsets) to allow cross-fold validation in which the machine learning algorithm can estimate its performance for parameter optimization. Cross-fold validation is used inuse all of the training data together to create the machine learning model. The second set of tables is the testing set, which is used to evaluate the performance of the final model. The two sets do not share any records, so that the evaluation will be unbiased. As described below, both sets share common methods, such as labeling and injection. 5.1. Training

[0059] In one embodiment, the ERNIE training set contains at least 150,000 archival RPM scans. The machine-learning model “learns” about typical encounters with non-radioactive cargo and shipments containing NORM from RPM scans of containers in the SOC. Potential threats and uncommon sources are trained using RPM scans that did not contain any radioactive emissions when measured, but which have since been augmented by injecting signatures from nuclear threats and other radioactive sources of concern.

[0060] In one embodiment, the ERNIE training process begins by screening RPM scans from the port or ports of interest to select about 10,000 archival scans without signs of radioactivity in the cargo, which are referred to as “non-emitting.” Each of these is then augmented by injecting a model of radioactivity from the four radioactive classes: nuclear, industrial, medical, and NORM. For single-source injections, this produces about 60,000 samples with uniform coverage of all classes (folds 1-10, described below). In the training table, injection levels are set high enough to ensure the injected samples are distinguishable from non-emitting records. In the real data, sources can exist at any level. Therefore, to ensure that the training is complete and to improve the coverage of the most commonly encountered cargo at each site (non-emitting and NORM), about 100,000 additional samples (fold 11, below) are drawn from the SOC without any selection other than to remove scans that are incomplete or that have other errors (irregular scans). An iterative process called “Semi-Supervised Learning” (SSL) is then applied to these mostly non-emitting and NORM stream-of-commerce samples. The SSL process provides a further improvement in performance by increasing the depth of training to include a wider range of these most commonly encountered sources. The injection and SSL process are described in greater detail below. In some cases, additional training is provided using data labeled based on alarm adjudication with Radioisotope Identifier (RIID) spectrometer instruments. In some cases, severalradioactive sources are included in folds 15-19 (see below).

[0061] Performance evaluation of the ERNIE system can be done using a completely separate testing set of records. The testing set also contains a set of unlabeled data and a set of injected records. However, the purposes for these two sets are different than those used for the training. The uninjected SOC set (fold 20) is used to determine the inspection rate that will be experienced at the port during typical operation. There are no screening criteria applied to this uninjected set. The single- source injected set of records (folds 32-34) is used to evaluate the detection sensitivity. Unlike the training set, these injected records are intended to challenge the system to find the limits of sensitivity. The injections therefore cover the full range from undetectable to detectable levels. A summary of the folds according to one embodiment is shown in Table 1. Fold Description r n r ) o5.2. Injection

[0062] Machine learning involves training so that the system can recognize the important classes of objects. Out of millions of cargo scans measured in the field, more than 99% of them are either not radioactive (non-emitting) or NORM. There are some measurements of industrial, medical, nuclear (fuel, oxide, UF6), and contamination sources of radioactivity, but not enough to span the full range of signatures these sources may produce. And there are other configurations of interest that may be completely absent from the data, such as smuggled fissile material or nuclear weapons (as far as is known). To provide training on all classes of sources, therefore, one can synthesize data, creating examples of important measurements that are not found in actual data. These synthetic measurements can be made as realistic as possible by “injecting” source models of interest into real measurements (semi-synthetic measurements).

[0063] To create an injected sample, one may first select a large set of scans with non-emitting cargo. This can be achieved by isolating records that have a peak intensity below those found in NORM and other samples, and also do not have any of the energy window ratios above a specified value. The selection of non-emitting samples can be done graphically with conservative limits to assure that virtually all of the samples are non-emitting.

[0064] Once non-emitting records are selected for injection, a model of a known signature can be injected. For consistency, in one embodiment, folds 1-10 of the training table are created using each non-emitting sample up to five times. It appears without injection labeled as non-emitting. It appears up to four more times, with injections and labeling for NORM, nuclear, industrial and medical. The primary training table is divided into 10 groups called fold 1 through fold 10. Each non-emitting base record has up to four injected versions are confined to a single fold. This allows ten-without overlap between training and testing. As described below, a large population of unaltered data containing mostly NORM and non-emitting can be used to fill in any difference between injected measurements and non-injected measurements. The NORM models can be created from background-subtracted measurements that have been verified to be NORM and from models tuned to match measured NORM samples. Contamination models can be created in a similar fashion. Training data with multiple sources are included in folds 101-110.

[0065] The nuclear models can be created using a spanning set of shielded fissile materials (sometimes called fissile models), for example, the LLNL Surrogate Nuclear Model (L-SNM). This set can use a complex series of randomly chosen shields surrounding a randomly chosen fissile material. The L-SNM models include signatures from configurations that could be used to smuggle Special Nuclear Materials (SNM) and other fissile nuclides, as well as signatures similar to nuclear weapons.

[0066] Industrial models can be created in a similar way using simpler shielding configurations around sources that are used in industry applications (e.g.60Co,137Cs,192Ir, etc.) and those that could be used to do harm as a radiological dispersals device or radiological exposure device. Finally medical source models can be created using a comprehensive compilation of radiopharmaceuticals with activity levels and frequency of encounter based on measurements and effective biological half-lives.

[0067] There is no way to know how a source might be placed in a cargo container. To ensure the full range of possibilities is included, a large number of injected samples can be created by placing sources randomly in different positions within a vehicle which is loaded with different configurations of attenuating cargo that are also randomly chosen. The attenuation model may include significant anisotropy due to variations both in bulk cargo, and in the outer shell of the vehicle. This anisotropy causes changes in the effective shielding as a vehicle passes through the portal. These changes in effective shielding affect both the overall count rate and spectral shape. Consequently, the shielding models can be tuned to match observations from the field.

[0068] In one embodiment, two attenuation models are used for radiation transport: in particular, the cabin portion of the vehicle uses a random surface model representing attenuation, and the cargo area uses both a bulk cargo model based oncreate anisotropy similar to those observed in the population as a whole. The same models can be applied regardless of the source injected. Thus, the subset of threat samples that are injected in the cabin can be treated the same as a medical source.

[0069] The radiation transport from the source through the attenuating cargo, and the detector response can be calculated using the deterministic and / or Monte Carlo radiation transport and detector response functions. For example, source model generation can also be developed using GADRAS (Mitchell, Dean, and & Mattingly, John. Gamma Detector Response and Analysis Software (GADRAS) v.16.0. Computer software. Vers. 01. USDOE. 24 Dec. 2009) or RadSim (K. Nelson et al., RadSim: a Modular Open-Source Gamma-ray Detector Simulator, LLNL-CODE-854710 and LLNL-CODE-855199 https: / / github.com / LLNL / RadSim / ) deterministic radiation modeling tools, and can be spot-checked using MCNP (M. E. Rising et al. MCNP® Code Version 6.3.0 Release Notes. Los Alamos National Laboratory Tech. Rep. LA- UR-22-33103, Rev.1. Los Alamos, NM, USA. January 2023) or GEANT (S. Agostinelli et al., Geant4—a simulation toolkit, Nucl. Instrum. Meth. A 506 (2003) 250, issn: 0168- 9002.) Monte Carlo radiation modeling tools to ensure the emission and shielding calculations are accurate. The range in emission levels will generally far exceed any discrepancies in the modeling. Thus, any error in an individual calculation is compensated for by the distribution of samples that span a wide range of source strengths and attenuation.

[0070] In addition to the variations in intensity due to the random cargo emplacement, the intensity of all the injection models can also be varied to avoid any biases associated with limitations in the model and in the cargo. Once the injection is complete, the intensities can be calculated for each time during the scan in each of the energy bins and in each of the detector panels. For the training tables, the peak intensity can then be extracted and compared with background from non-emitting scans. If the total measured counts due to the injection are too small to make a significant change in the signature, these injections can be excluded from training in folds 1-10. This can cause a small “gap” between the non-emitting and emitting sources which is filled in by the Semi-Supervised Learning (SSL). For testing, a separate set of injected test folds can be created, folds 32-45, that use non-emitting samples that were not used anywhere in the training set, and include samples with much lower emissionsa challenging set of single medical (fold 32), industrial (fold 33) and nuclear (fold 34) sources, and test of performance against mixed sources (folds 42-45). 5.3. Feature Extraction

[0071] As shown in Figure 4, a vehicle 404 carrying a radiation source 405 may pass through the portal 406. The outputs of the radiation detectors 408 are used to produce a radiation profile for the vehicle versus time. The outputs of the VPS 410 can be used to select a BRSP 412 for the vehicle 404, examples of which is shown in Figures 4 and 5.

[0072] The VPS typically includes one or more pairs 410 of an infrared (IR) beam emitter and sensor. The beam emitters are placed on one portal pillar and the sensors on the other, so that when a vehicle enters the portal 406, the IR beam is broken. The VPS emitters and sensors are grouped in pairs 410, where each pair can be positioned on opposite edges (leading or trailing) of the panel to capture the time needed for a vehicle to transit the panel width. Preferably, one pair is set near parallel to the ground at a position low enough to capture the wheels and undercarriage of the vehicle, and a second pair is positioned at a diagonal from a much higher location to capture the upper parts of the vehicle, e.g., the cab and trailer for a tractor-trailer rig. Ultrasonic sensors can also be employed. The VPS provides the time and duration of the occupancy, and an estimate of the entry and exit speeds. When the full VPS data stream is available, the ERNIE system can use all of the signal transitions available to estimate a more reliable initial velocity and acceleration. When more than two reliable transitions are available, the system can also estimate the jerk (time rate of change of the acceleration) to provide a reliable third-order motion profile. The motion profile can include a vehicle length estimate, and when the full VPS data stream is available, the system can use the pattern of upper and lower beam breaks to classify the vehicle type as one of many categories, ranging from motorcycles to double semi-trailer trucks.

[0073] Some RPM systems may not record the full VPS data stream but may report one or more speed estimates. For such systems, the ERNIE technique can employ an alternative motion profile extraction that uses the entry speed estimate (when reliable), the occupancy duration, and a model based on historical data from similar systems, to estimate the most likely exit speed and acceleration. More specifically, theand an exit speed vs. occupancy duration distribution, and then calculate the acceleration for those two speeds.

[0074] To avoid using the same motion profile for injection that is used in feature extraction analysis, the injector can select an exit speed randomly weighted by any indicated exit speed and historical distributions. An initial jerk term randomly weighted by the distribution of jerks found in full VPS motion profiles can also be used to provide more variation in motion profiles. The initial acceleration can then be calculated to provide the measured entry speed and selected exit speed with the selected jerk term. This approach provides a realistic motion profile for feature extraction of each scan, and a slightly different profile for injection even for samples with the same entry speed and occupancy duration.

[0075] In one embodiment, inputs to the ERNIE machine learning analysis are provided by feature extraction. In addition to isolating the analysis from any residual artifacts that could artificially differentiate between semi-synthetic and field measurements, physics-based features can be used that are designed to reflect characteristics of the source classes. These characteristics can be grouped into three types: intensity, spatial, and spectral. Intensity levels are key to identifying the higher intensity threat samples, and the background-only non-emitting scans. The spatial distribution of the intensity can help isolate the extended emissions from the distributed radioactivity in NORM sources, from the more point-like emissions from many of the threat sources. Also, spectral distributions, when available, can be used to further isolate different source classes.

[0076] Intensity features can use statistical methods, such as the mean background-subtracted count rate and the number of standard deviations above background (N-Sigma). Features that characterize the spatial distribution of the intensity can provide indications of how compact or spread out the emissions are, and can include the moments of the mean (standard deviation, skew, etc.) calculated directly on the background-subtracted number of measured gamma-ray counts, and on the position weighted by the number of counts. Positions can be determined from the time of the measurement converted to distance along the vehicle using the vehicle motion profile (speed as a function of time). Direct measures of the spatial extent of theand related statistics can also be used. In one embodiment, the Full Width at one Quarter Maximum (FWQM) and Full width at Three Quarters Maximum (FW3QM) can also be used as features and can be computed both by starting at the peak and working outward to the half-maximum location (“inside” FWHM) and starting at the beginning and end of the scan and searching inward (“outside” FWHM). The ratio of these can be used as features as well, which is close to one for most single sources, and can be far from one if there are spatially distinct peaks.

[0077] Location features also be used to provide spatial information. The source location in the lateral direction (distance from the center to the left or right of the conveyance), and the source height location above the ground can be estimated using trained models applied to sums and differences of the upper, lower, left and right-side detector panel gamma-ray counts as appropriate. The source location along the direction of travel can be estimated by finding the time of the maximum source signature, and then applying the motion profile to convert the time to the distance from the front of the conveyance.

[0078] Spectral features can use energy window ratios and correlations with spectral shapes of interest that are tuned to specific source classes. In one embodiment, iterative k-means clustering can be used to extract spectral archetype templates that span the spectral shapes found in both the field measurements and semi- synthetic measurements. The goodness of fit or correlation metrics of each scan versus each of the spectral archetypes can provide feature arrays that convey to the machine learning model indications when a sample matches more than one spectral shape archetype. A poor fit to one or more spectral archetypes may be as informative an indicator of source class as a good fit to a single archetype. Similarly, features based on energy window ratios can be used with different combinations of energy channels in the numerator and denominator.

[0079] To enable searching for single or multiple sources with unique spectral signatures, additional sets of features can be used. In one set, the statistical and spectral analyses can be applied to the largest excursion of the energy window ratios found in the scan, which does not always align with the maximum in background subtracted total counts used in the other statistical and spectral features. Another setbetween the measured data and each of the spectral archetypes is calculated, and a matched filter is constructed that matches the primary source spectrum found in the data. The matched filter can be used to remove this source, and a filter can be constructed for the next source. A joint analysis can then determine the intensity of each source needed to match the measured data. The location of the maximum intensity of each source can then be found, and statistical and spectral features can be calculated for each. This joint analysis can find two sources that are overlapping as well as those that are in different locations. If only one source is present, the intensity feature for the second source will be small, thus guiding the machine learning toward a one-source conclusion. The scans can also be split into two sections creating front and rear segments before and after the centroid location of counts for the full scan. All the statistical features can then be calculated on the peak region of the front section, and on the peak region of the rear section as well as the peak region found for the entire scan. 6. Background Suppression

[0080] During a scan of a large vehicle, the background radiation (also called “background” herein) is suppressed as the vehicle shades the detectors from the background radiation. The depth and shape of this background suppression effect varies tremendously from scan to scan, but for vehicles of the same type and configuration, the overall pattern of suppression is fairly consistent. When available, the ERNIE system 100 uses the VPS data to classify the vehicle type, and then applies a background radiation suppression profile (BRSP) (template) created from averaging data of similar vehicles out of many (e.g., thousands) of scans. The depth of the suppression will vary with the cargo loading, but the average provides a reasonable estimate, and it helps expose source signatures within the suppressed region that do not extend above the background level measured before and after the vehicle scan. This can be done by subtracting the estimated background from the selected BRSP template prior to feature extraction. An example of the background suppression is shown in Figure 5, in which the use of a BRSP 512 (represented by the dashed line) enables the ERNIE analysis to detect and locate the position of a source 505 within aabove the background measured before the vehicle 504 entered the RPM. 7. Machine Learning Approach

[0081] Any of various machine learning methods can potentially be employed in the ERNIE classification technique. In at least one embodiment, random forest analysis is employed, due to its transparency, efficiency, and effectiveness. Random forests are ensembles of highly interpretable decision trees, and as such, they can provide statistical metrics that can be used to assess reliability of each prediction on-the-fly, that is, at the same time when these predictions are made. Such real-time diagnostics can be very useful, especially in applications where tolerance to errors is very low. In other embodiments, other types of machine learning approaches may be employed, such as k-nearest neighbor, kernel density estimates, support vector machines, and variations of neural networks and deep learning.

[0082] In one embodiment, the ERNIE system 100 uses a random forest machine learning analysis such as shown applied to RPM source analysis schematically in Figures 6A and 6B. A decision tree is a collection of binary decisions that splits training data by values of the most informative feature at each decision point (i.e., each subsequent node) starting from the top. Figure 6A shows conceptually an example of a single binary decision within a tree with illustrative feature split thresholds and resulting class distributions, and with illustrative values. Figure 6B shows a simplified example of a decision tree. Each feature is labeled by number (e.g. “feature_973”) and below each feature is the feature threshold for that node. If the value of the feature is greater than or equal to the threshold, the analysis moves to the right where additional features may be analyzed. If the feature value is less than the threshold, the analysis moves to the left. This process extends until no further nodes are included in the model either because no further information gain was possible, or the maximum tree depth has been reached. The rectangles at the end of each branch (leaf nodes) show the number of samples in each class (NORM, RAD, INDUSTRIAL, SNM) found at that leaf. These values in the final leaf node found when descending the tree for a given sample are used to estimate the likelihood of each class. In a random forest, many trees are used, each trained with variations of the training data, and the final likelihood is determined by votes from all the trees. In one embodiment, 30 decision trees are usedlevels. Each decision node reflects a feature of data and its threshold value that is used to determine which way to move down the tree towards the decision point. Each of the bottom (“leaf”) nodes contains a class distribution of the training data points that fall into that leaf after being routed appropriately from the top node of the tree. This provides a multi-dimensional analysis with potentially very complex decision boundaries; in contrast to the relatively simple two-dimensional decision surfaces used by the simple thresholds used in current RPM detection systems. Each decision tree is seeded differently and provides a different “perspective” on the analysis. The final class prediction is based on the tally of outcomes from the multiple trees.

[0083] One can use a random forest with bootstrap aggregating (bagging) by training multiple decision trees. Bootstrap aggregating (bagging) is a standard practice in random forest construction. As the algorithm for determining the structure of a tree is deterministic, applying the decision tree creation algorithm to the same data will result in the same tree. To solve this difficulty so that a diverse set of trees is produced, each placing different emphasis on the features, one can force the tree-building algorithm to construct suboptimal trees. This can be done either by presenting a subset of the data or by presenting a subset of the available features as the tree is built. In the bootstrap approach, one can use subsets of the data for each tree. Although the quality of each individual tree will be suboptimal of the training set as a whole, the forest of trees is much more robust in their ability to generalize the decision boundaries. Each decision tree is trained on a different bootstrapped (random sampling with replacement) version of the training data. This bagging approach helps to improve performance by reducing the variance component of the prediction errors and is typically more robust to overfitting.

[0084] As mentioned above, in some embodiments a two-step training process can be employed to train the ML model. The initial training step uses the labeled training data where all source classes are represented by the semi-synthetic measurement- based simulations. The second step uses SSL to incrementally label additional SOC data and then include the new labeled data in the subsequent model training cycles.

[0085] The primary goal of SSL is to improve model performance by adding the measured unlabeled data to the original training dataset in an iterative fashion. Mostpoints from a second set of SOC data, designated fold 11 (see Table 1 above), to the original training set. Fold 11 is similar to fold 20, but drawn from different records not used anywhere else in the ERNIE process. The resulting model improves performance as it learns about previously underrepresented regions in feature space, and tunes the model to match the commodities and detector responses found in a given set of data.

[0086] Figure 7 illustrates an example of an SSL training subsystem that, in some embodiments, can perform the second step of training mentioned above, i.e., to incrementally label additional SOC data and then include the new labeled data in the subsequent model training cycles. The major components of the SSL training subsystem 700 include a model injection module 721, two feature extraction modules 722 and 723 (which can include the same code), an initial ML training module 724, an SOC feature table 725, an initial feature table 727, an expanded feature table 728, an iterative ML training module 729, and updated ML model storage 733, and an iterative labeling module 734. The initial feature table 727 is generated by the feature extractor 723 using semi-synthetic measurements produced with model injection 721 of source models 730 into non-emitting measurements 731. The initial feature table 727 is used by initial ML training module 724 to train the initial ML model 706. The SOC feature table 726 is generated by feature extraction 722 of SOC data 732. The iterative labeling module 734 uses the initial ML model 706 to determine an investigate score for the SOC feature table 725, and selects SOC samples with the most confident resolution (lowest and sometimes highest investigate scores) which it then removes from the SOC feature table 725 and adds them to the expanded feature table 728, which in the first iteration includes the initial feature table 727 samples. An updated ML model 733 is then trained by the iterative ML training module 729 using the expanded feature table 728. The updated ML model 733 is then used by the iterative labeling module 734 to determine new investigate scores for the remaining SOC samples from the SOC feature table 725 and again selects the most SOC samples with the most confident resolution, removes those samples from the SOC feature table 725 and adds them to the expanded feature table 728. This loop continues until an appropriate number of SOC samples have been included in the expanded feature table 728, or the remaining SOC samples are found by the iterative labeling module 734 to be indeterminant. The updated ML model 733RPM scans.

[0087] In some embodiments, another training stage is added as a final training stage, to further augment the model expanding further into underrepresented regions in feature space found in scans of cargo with radioactive contamination or distributed nuclear materials. The contamination and nuclear materials (CNM) initial training table, designated fold 12 (see Table 1 above) can include, for example, at least two additional classes of sources: 1) contaminated metal, waste products, and food; and 2) nuclear materials such as depleted uranium metal, uranium hexafluoride, uranium oxide, and nuclear fuels. These sources can be injected in a similar manner as the single and multiple sources (folds 1-10 and 101-110), both as small objects and as materials that can extend up to the full length of the cargo region of the vehicle. CNM are an important hazard source class (that warrants investigation) that is not fully represented in the single and multiple source initial training sets mentioned above (folds 1-10, and 101- 110 respectively). CNM sources are generally extended, and as such can overlap with NORM. CNM also includes radionuclides that are considered radiological hazards or used in industrial applications. These overlaps, particularly the overlap with the NORM source type (which produces nuisance alarms), make it difficult to build a training set with CNM sources that does not degrade crucial nuisance alarm suppression.

[0088] A supplemental CNM method of training the ERNIE system avoids these overlaps and maintains good nuisance alarm rejection performance while increasing hazard material detection performance. A training set (feature table) of extracted CNM features is created to include a broad spanning set of CNM sources regardless of overlaps with the other training data. It can be generated by extracting features from non-emitting measurements and injecting CNM source models. The above-described SSL-enhanced machine learning model created without CNM sources then can be used to probe the initial CNM training table to select samples that do not have overlaps with the results of previous (non-CNM) training but that do fill in feature space that should be flagged as investigate. These selected CNM samples then can be added to the initial training feature table or to the expanded feature table that resulted from the SSL if SSL is applied, to make a final more comprehensive training table with better hazard detection capabilities but without compromising the nuisance source rejection capabilities. An iterative approach can be applied using machine learning models tostarting with the initial ML model or the final SSL model if SSL is applied. In another embodiment, CNM samples are selected using kernel density estimates or other methods to select data from the CNM initial training table that fill in low density regions in feature space in the initial training table or the final SSL expanded training table if SSL is applied. 8. Overall Processes

[0089] Figure 8 illustrates an example of a training process 800 for training a machine learning model according to an embodiment of the ERNIE technique. Initially, at step 801 the process 800 accesses motion profile information and vehicle type information associated with a plurality of vehicles. At step 802, the process 800 trains a machine learning model on training data that include the motion profile information and the vehicle type information associated with the plurality of vehicles. The machine learning model is for use by a classifier program to determine whether a vehicle that has transited a monitoring portal potentially includes a hazardous material.

[0090] Figure 9 illustrates an example of an overall classification process 900 for classifying vehicles and radiation sources according to an embodiment of the ERNIE technique. At step 901 the process 900 inputs radiological scan results data representative of results of a radiological scan of a vehicle, where the radiological scan has been performed by one or more detectors arranged at a radiation portal monitor (RPM) during a passage through the monitored portal by the vehicle. At step 902 the process inputs vehicle context information related to the vehicle and corresponding to the passage through the monitored portal by the vehicle, wherein the vehicle context information includes information related to the vehicle other than the radiological scan results data. At step 903 the process extracts a feature set from the radiological scan results data and the vehicle context information. At step 904 the process performs an investigative classification by using an investigative classifier machine learning model on the feature set, as a result of which the vehicle is classified as recommended or not for further investigation for possible presence of a radiological hazard. At step 905 the process generates an alert indicative of whether the vehicle is classified as recommended for further investigation for possible presence of a radiological hazard, based on a result of the investigative classification. And, at step 906 the process causesdevice). 9. Computer System

[0091] Figure 10 is a high-level block diagram illustrating an example of a computer system 1000 in which some or all of the ERNIE technique introduced above can be implemented. Computer system 1000 can represent a system that implements the classification system, training system, or both, in Figure 1. The computer system 1000 includes one or more processors 1001, one or more memories 1002, one or more input / output (I / O) devices 1003, and one or more communication interfaces 1004, all connected to each other through an interconnect 1005. In at least one embodiment, the one or more processors 1001 include a plurality of GPUs, each with dedicated memory, that collectively execute the ERNIE technique described above. The processor(s) 1001 may control the overall operation of the computer system 1000, including controlling its constituent components. The processor(s) 1001 may be or include one or more conventional microprocessors, GPUs, programmable logic devices (PLDs), field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), etc. The one or more memories 1002 store data and executable instructions (e.g., software and / or firmware), which may include software and / or firmware for performing the ERNIE technique introduced above. The one or more memories 1002 may be or include any of various forms of random access memory (RAM), read-only memory (ROM), volatile memory, nonvolatile memory, or any combination thereof. For example, the one or more memories 1002 may be or include dynamic RAM (DRAM), static RAM (SDRAM), flash memory, one or more disk-based hard drives, etc. The I / O devices 1003 provide access to the computer system 1000 by human user, and may be or include, for example, a display monitor, audio speaker, keyboard, touch screen, mouse, microphone, trackball, etc. The communications interface 1004 enables the computer system 1000 to communicate with one or more external devices (e.g., one or more servers and / or clients of the computer system 1000) via a network connection and / or direct connection. The communications interface 1004 may be or include, for example, a Wi-Fi adapter, Bluetooth adapter, Ethernet adapter, Universal Serial Bus (USB) adapter, or the like. The interconnect 1005 may be or include, for example, one or more buses, bridges or adapters, such as a system bus,like. Examples 1. A method of training a machine learning model, the method comprising: accessing, by a computer system, motion profile information and vehicle type information associated with a plurality of vehicles; and training, by the computer system, the machine learning model on training data that include the motion profile information, the vehicle type information being associated with the plurality of vehicles and data representative of results of a plurality of scans of vehicles for a hazardous material, the machine learning model being for use by a classifier program to determine whether a vehicle that has transited a monitoring portal potentially includes the hazardous material. 2. The method of example 1, wherein the motion profile information comprises vehicle presence sensor (VPS) data. 3. The method of example 1, wherein the motion profile information comprises a speed as a function of time for a passage of a vehicle through the portal. 4. The method of example 1, wherein the motion profile information comprises data indicative of a time history of output signals from the one or more detectors. 5. The method of example 1, wherein the vehicle type information comprises a background suppression profile of a vehicle. 6. The method of example 1, further comprising: training a second machine learning model on training data that include a plurality of source types, the second machine learning model for use by a second classifier program to classify a source type of the hazardous material.training the machine learning model during a first phase by using labeled data corresponding to each of a plurality of source types of interest, wherein each of the plurality of source types of interest is a type of hazardous material; and training the machine learning model during a second phase, for use by a classifier to determine whether a hazardous source is potentially present, by incrementally performing a plurality of additional model training cycles with additional data labeled in a semi-supervised learning process, wherein data labeled during at least one of the plurality of additional training cycles is used to train the machine learning model in a subsequent model training cycle. 8. The method of example 7, wherein at least two of the plurality of source types of interest are types of radiation source. 9. The method of example 7, wherein the plurality of source types of interest are represented in the labeled data by semi-synthetic measurement-based simulations. 10. The method of example 1, wherein the machine learning model is trained with injected labeled data that include injection of signatures from hazardous materials into a plurality of vehicle types. 11. The method of example 10, wherein the training source data include only injected data and do not include measured data with source emissions. 12. The method of example 10, wherein the injected labeled data are biased by position within a vehicle as a function of source type. 13. The method of example 1, further comprising: training the machine learning model during an additional training phase by using contamination and nuclear materials (CNM) training data generated in part using CNM source model injection; andtraining data, to train the machine learning model. 14. A processing system comprising: at least one processor; and at least one storage facility storing instructions, execution of which by the at least one processor causes the processing system to perform a machine learning training process that comprises: accessing motion profile information and vehicle type information associated with a plurality of vehicles; and training a machine learning model on training data that include the motion profile information, the vehicle type information associated with the plurality of vehicles and data representative of results of a plurality of scans of vehicles for a hazardous material, the machine learning model for use by a classifier program to determine whether a vehicle that has transited a monitoring portal potentially includes the hazardous material. 15. The processing system of example 14, wherein the motion profile information comprises vehicle presence sensor (VPS) data. 16. The processing system of example 14, wherein the motion profile information comprises speed as a function of time for a passage of a vehicle through the portal. 17. The processing system of example 14, wherein the motion profile information comprises data indicative of a time history of output signals from the one or more detectors. 18. The processing system of example 14, wherein the vehicle type information comprises a background suppression profile of a vehicle.training process further comprises: training a second machine learning model on training data that include a plurality of source types, the second machine learning model for use by a second classifier program to classify a source type of the hazardous material. 20. The processing system of example 14, wherein the training comprises: training the machine learning model during a first phase by using labeled data corresponding to each of a plurality of source types of interest, wherein each of the plurality of source types of interest is a type of hazardous material; and training the machine learning model during a second phase, for use by a classifier to determine whether a hazardous source is potentially present, by incrementally performing a plurality of additional model training cycles with additional data labeled in a semi-supervised learning process, wherein data labeled during at least one of the plurality of additional training cycles is used to train the machine learning model in a subsequent model training cycle. 21. The processing system of example 20, wherein at least two of the plurality of source types of interest are types of radiation source. 22. The processing system of example 20, wherein the plurality of source types of interest are represented in the labeled data by semi-synthetic measurement- based simulations. 23. The method of example 14, wherein the machine learning model is trained with injected labeled data that include injection of signatures from hazardous materials into a plurality of vehicle types. 24. The processing system of example 23, wherein the training source data include only injected data and do not include measured data with source emissions.are biased by position within a vehicle as a function of source type. 26. The processing system of example 14, wherein training the machine learning model further comprises: training the machine learning model during an additional training phase by using contamination and nuclear materials (CNM) training data generated in part using CNM source model injection; and selecting samples from the CNM training data that do not overlap with previous training data, to train the machine learning model. 27. A machine-readable program storage medium tangibly embodying sequences of instructions, execution of which by at least one processor in a processing system causes the processing system to execute a machine learning training process comprising: accessing motion profile information and vehicle type information associated with a plurality of vehicles; and training a machine learning model on training data that include the motion profile information, the vehicle type information associated with the plurality of vehicles and data representative of results of a plurality of scans of vehicles for a hazardous material, the machine learning model for use by a classifier program to determine whether a vehicle that has transited a monitoring portal potentially includes the hazardous material. 28. The machine-readable program storage medium of example 27, such that the motion profile information comprises vehicle presence sensor (VPS) data. 29. The machine-readable program storage medium of example 27, such that the motion profile information comprises speed as a function of time for a passage of a vehicle through the monitoring portal.the motion profile information comprises data indicative of a time history of output signals from one or more detectors at the monitoring portal. 31. The machine-readable program storage medium of example 27, such that the vehicle type information comprises a background suppression profile of a vehicle. 32. The machine-readable program storage medium of example 27, such that the machine learning training process further comprises: training a second machine learning model on training data that include a plurality of source types, the second machine learning model for use by a second classifier in the detection system to classify a source type of the hazardous material. 33. The machine-readable program storage medium of example 27, such that the training comprises: training the machine learning model during a first phase by using labeled data corresponding to each of a plurality of source types of interest, wherein each of the plurality of source types of interest is a type of hazardous material; and training the machine learning model during a second phase, for use by a classifier to determine whether a hazardous source is potentially present, by incrementally performing a plurality of additional model training cycles with additional data labeled in a semi-supervised learning process, wherein data labeled during at least one of the plurality of additional training cycles is used to train the machine learning model in a subsequent model training cycle. 34. The machine-readable program storage medium of example 33, such that at least two of the plurality of source types of interest are types of radiation source.the plurality of source types of interest are represented in the labeled data by semi- synthetic measurement-based simulations. 36. The method of example 27, wherein the machine learning model is trained with injected labeled data that include injection of signatures from hazardous materials into a plurality of vehicle types. 37. The machine-readable program storage medium of example 36, such that the training source data include only injected data and do not include measured data with sources. 38. The machine-readable program storage medium of example 36, such that the injected labeled data are biased by position within a vehicle as a function of source type. 39. The machine-readable program storage medium of example 25, such that the machine learning training process further comprises: training the machine learning model during an additional training phase by using contamination and nuclear materials (CNM) training data generated in part using CNM source model injection; and selecting samples from the CNM training data that do not overlap with previous training data, to train the machine learning model.

[0092] Unless contrary to physical possibility, it is envisioned that (i) the methods / steps described herein may be performed in any sequence and / or in any combination, and that (ii) the components of respective embodiments may be combined in any manner.

[0093] The machine-implemented computational and control operations described above can be implemented by programmable circuitry programmed / configured by software and / or firmware, or entirely by special-purpose circuitry, or by a combination of such forms. Such special-purpose circuitry (if any) can be in the form of, for example,devices (PLDs), field-programmable gate arrays (FPGAs), system-on-a-chip systems (SOCs), etc.

[0094] Software or firmware to implement the techniques introduced here may be stored on a machine-readable storage medium and may be executed by one or more general-purpose or special-purpose programmable microprocessors. A “machine- readable medium”, as the term is used herein, includes any mechanism that can store information in a form accessible by a machine (a machine may be, for example, a computer, network device, cellular phone, personal digital assistant (PDA), manufacturing tool, any device with one or more processors, etc.). For example, a machine-accessible medium includes recordable / non-recordable media (e.g., read-only memory (ROM); random access memory (RAM); magnetic disk storage media; optical storage media; flash memory devices; etc.), etc.

[0095] Any or all of the features and functions described above can be combined with each other, except to the extent it may be otherwise stated above or to the extent that any such embodiments may be incompatible by virtue of their function or structure, as will be apparent to persons of ordinary skill in the art. Unless contrary to physical possibility, it is envisioned that (i) the methods / steps described herein may be performed in any sequence and / or in any combination, and that (ii) the components of respective embodiments may be combined in any manner.

[0096] Although the subject matter has been described in language specific to structural features and / or acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are disclosed as examples of implementing the claims and other equivalent features and acts are intended to be within the scope of the claims.

Claims

AMENDED CLAIMS received by the International Bureau on 03 March 2026 (03.03.2026)1. A method of detecting a potentially radiologically hazardous material in a vehicle, the method comprising: inputting radiological scan results data representative of results of a radiological scan of the vehicle, the radiological scan of the vehicle having been performed by one or more detectors arranged at a radiation portal monitor (RPM) during a passage of the vehicle through a monitored portal; inputting vehicle context information related to the vehicle and corresponding to the passage of the vehicle through the monitored portal, wherein the vehicle context information includes information related to the vehicle other than the radiological scan results data; extracting a feature set from the radiological scan results data and the vehicle context information; selecting one of a plurality of machine learning models, including a first machine learning model and a second machine learning model, to use as an investigative classifier machine learning model for performing an investigative classification, based on whether vehicle presence sensor (VPS) data are available and satisfy a reliability criterion, wherein the first machine learning model has been trained on features extracted from data including VPS data and the second machine learning model has been trained on features extracted from data that excludes VPS data; performing the investigative classification by using the selected investigative classifier machine learning model on the feature set, as a result of which the vehicle is classified as recommended or not for further investigation for possible presence of a radiological hazard; generating an alert indicative of whether the vehicle is classified as recommended for further investigation for possible presence of a radiological hazard, based on a result of the investigative classification; and causing the alert indicative of whether the vehicle is classified as recommended for further investigation for possible presence of a radiological hazard to be output to a user.

2. The method of claim 1 , wherein the vehicle context information comprises a vehicle type of the vehicle.

3. The method of claim 1 , wherein the vehicle context information comprises motion of the vehicle as a function of time associated with passage of the vehicle through the monitored portal.

4. The method of claim 1 , wherein the vehicle context information comprises data indicative of a time history of output signals from the one or more detectors.

5. The method of claim 1 , further comprising: using the vehicle context information to select a background suppression profile of the vehicle.

6. The method of claim 1 , further comprising: performing a source type classification on the feature set by using a source type classifier machine learning model on the feature set, to classify a source type of a radiological source associated with the vehicle; generating an alert indicative of results of the source type classification; and causing an alert indicative of the source type classification to be output to a user.

7. The method of claim 6, further comprising: determining a level of uncertainty associated with the source type classification by applying the source type classification to a source type uncertainty classifier machine learning model; and determining that the source type classification is indeterminate when the level of uncertainty satisfies a specified criterion.

8. The method of claim 1 , wherein the selected investigative classifier machine learning model has been trained on training data that include motion profile information and vehicle type information associated with a plurality of vehicles.

9. The method of claim 8, wherein the motion profile information is based on historical distributions of motion of vehicles through the monitored portal.

10. The method of claim 1 , wherein the selected investigative classifier machine learning model has been trained on training data that include injection of a plurality of source types into each of a plurality of vehicle types.

11. The method of claim 1 , wherein the selected investigative classifier machine learning model has been trained with injected labeled data corresponding to each of a plurality of source types of interest.

12. The method of claim 11 , wherein the training source data include only injected data and do not include measured data with source emissions.

13. The method of claim 11 , wherein the injected labeled data are biased by position within a vehicle as a function of source type.

14. The method of claim 1 , further comprising: determining whether the VPS data are available and satisfy a reliability criterion.

15. The method of claim 1 , further comprising: performing an anomaly classification on the feature prior to performing the investigative classification, by applying the feature set to an anomaly classifier machine learning model to produce an anomaly score; wherein the investigative classification is performed only when the anomaly score satisfies a specified criterion.

16. The method of claim 1 , further comprising: accessing VPS data associated with the vehicle; determining a vehicle type of the vehicle based on the VPS data; identifying a background suppression template that corresponds to the vehicle based on the determined vehicle type of the vehicle; andwherein using the investigative classifier machine learning model comprises using the identified background suppression template to subtract expected background radiation during the extracting of the feature set.

17. The method of claim 1 , wherein the investigative classifier machine learning model has been trained incrementally in a plurality of training cycles with data labeled by a semi-supervised learning process.

18. The method of claim 1 , wherein the investigative classifier machine learning model has been trained from contamination and nuclear materials (CNM) training data generated by using CNM source model injection and by selecting samples from the CNM training data that do not overlap with previous training data, to train the investigative classifier machine learning model.

19. The method of claim 1 , wherein the feature set extracted from the radiological scan includes a set of features each indicating a distribution of intensities found within the scan.

20. The method of claim 1 , wherein the feature set extracted from the radiological scan includes a set of features each indicating an estimated distance between a rise and a fall of measured radiation to key fractions of the maximum measured intensity.

21. The method of claim 20, wherein the features in the feature set include a distance between the rise and fall of the measured radiation to key fractions of the maximum measured intensity as found in a rise and a fall closest to the maximum intensity, and also found in a rise and a fall farthest from the maximum intensity.

22. The method of claim 21 , wherein the features in the feature set include a ratio of distances between a nearest and a farthest rise and fall to key fractions of the maximum measured intensity.

23. The method of claim 1 , wherein the feature set extracted from the radiological scan includes a set of features indicating a maximum value of key energy window ratios found by summing selected energy channels to form a numerator, and selecting other energy channels to form a denominator.

24. The method of claim 1 , wherein the feature set extracted from the radiological scan includes a set of features each indicating a degree to which the measured energy data from key portions of the scan match a set of spectral templates selected to span the range of sources that might be present.

25. The method of claim 1 , wherein the feature set extracted from the radiological scan includes a set of features calculated after a joint decomposition of the radiological scan into a plurality of scans, including a first scan having a most dominant source signal and a second scan having any signal present from a second source, wherein the features in the feature set include energy and intensity distributions from each of the plurality of scans.

26. A system comprising: a plurality of physical detectors arranged around a portal designed to allow transit of vehicles therethrough, the plurality of physical detectors being arranged to perform a scan of a vehicle for hazardous material while the vehicle is transiting through the portal; and a processing system communicatively coupled to the plurality of physical detectors, the processing system including; at least one processor, and at least one memory accessible to the at least one processor and individually or collectively having instructions stored therein, execution of which by the at least one processor causes the processing system to perform an evaluation process, the evaluation process including: inputting scan results data representative of results of the scan of the vehicle;extracting a feature set from the scan results data; selecting one of a plurality of machine learning models, including a first machine learning model and a second machine learning model, to use as an investigative classifier machine learning model for performing an investigative classification, based on whether vehicle presence sensor (VPS) data are available and satisfy a reliability criterion, wherein the first machine learning model has been trained on features extracted from data including VPS data and the second machine learning model has been trained on features extracted from data that excludes VPS data; performing the investigative classification on the feature set by applying the feature set to the investigative classifier machine learning model, as a result of which the vehicle is classified as recommended or not for further investigation for possible presence of a hazard; performing a source type classification on the feature set by applying the feature set to a source type classifier machine learning model, to classify a source type of a hazard source associated with the vehicle; and generating an alert indicative of the investigative classification and the source type classification; and causing the alert indicative of the investigative classification and the source type classification to be output to a user.

27. The system of claim 26, wherein the investigative classification is further based on vehicle context information related to the vehicle and corresponding to passage of the vehicle through the monitored portal.

28. The system of claim 27, wherein the vehicle context information comprises a vehicle type of the vehicle.

29. The system of claim 27, wherein the vehicle context information comprises motion of the vehicle as a function of time associated with passage of the vehicle through the monitored portal.

30. The system of claim 27, wherein the vehicle context information comprises data indicative of a time history of output signals from the plurality of detectors.

31. The system of claim 26, wherein the physical detectors are radiation detectors, and the scan is a radiological scan.

32. The system of claim 26, wherein the selected investigative classifier machine learning model has been trained on training data that include motion profile information and vehicle type information associated with a plurality of vehicles.

33. The system of claim 32, wherein the motion profile information is based on historical distributions of motion of vehicles through the monitored portal.

34. The system of claim 26, wherein the selected investigative classifier machine learning model has been trained on training data that include injection of a plurality of source types into each of a plurality of vehicle types.

35. The system of claim 26, wherein the selected investigative classifier machine learning model has been trained incrementally in a plurality of training cycles with data labeled by a semi-supervised learning process.

36. he system of claim 26, wherein the selected investigative classifier machine learning model has been trained with injected labeled data corresponding to each of a plurality of source types of interest.

37. The system of claim 36, wherein the training source data include only injected data and do not include measured data with source emissions.

38. The system of claim 36, wherein the injected labeled data are biased by position within a vehicle as a function of source type.

39. The system of claim 26, such that the evaluation process further comprises: determining whether VPS data are available and satisfy a reliability criterion; and selecting one of a plurality of machine learning models to use in the performing the investigative classification, based on whether the VPS data are available and satisfies the reliability criterion.

40. The system of claim 26, wherein the selecting comprises: in response to a determination that the VPS data are available and satisfies the reliability criterion, selecting a first machine learning model to use as the investigative classifier machine learning model in the performing the investigative classification, the first machine learning model having been trained on a feature set that is based on VPS data; or in response to a determination that the VPS data are not available or fails to satisfy the reliability criterion, selecting a second machine learning model to use as the investigative classifier machine learning model in the performing an investigative classification, wherein the second machine learning model has been trained on a feature set that is not based on VPS data.

41. The system of claim 26, further comprising: determining a level of uncertainty associated with the source type classification by applying the source type classification to a source type uncertainty classifier machine learning model, wherein the source type classification is deemed valid only when the level of uncertainty satisfies a specified criterion.

42. The system of claim 26, further comprising: performing an anomaly classification on the feature prior to performing the investigative classification, by applying the feature set to an anomaly classifier machine learning model to produce an anomaly score; wherein the investigative classification is performed only when the anomaly score is below a specified threshold.

43. The system of claim 26, further comprising: accessing VPS data associated with the vehicle; determining a vehicle type of the vehicle based on the VPS data; identifying a background suppression template that corresponds to the vehicle based on the determined vehicle type of the vehicle; and wherein using the investigative classifier machine learning model comprises using the identified background suppression template to subtract expected background radiation during the extracting of the feature set.

44. The system of claim 26, wherein the investigative classifier machine learning model has been trained from contamination and nuclear materials (CNM) training data generated in part by using CNM source model injection and by selecting samples from the CNM training data that do not overlap with previous training data, to train the investigative classifier machine learning model.

45. A machine-readable program storage medium tangibly embodying sequences of instructions, execution of which in a processing system causes the processing system to execute a process comprising: inputting radiological scan results data representative of results of radiological scan of a vehicle that passed through a portal monitored by a radiation portal monitor (RPM); inputting vehicle context information related to the vehicle and corresponding to passage through the portal by the vehicle, wherein the vehicle context information includes information related to the vehicle other than the radiological scan results data; extracting a feature set from the radiological scan results data; performing an investigative classification on the feature set by applying the feature set to an investigative classifier machine learning model, as a result of which the vehicle is classified as recommended for further investigation for possible presence of a radiological hazard or not recommended for further investigation for possible presence of a radiological hazard, wherein the investigative classifier machine learning model initially has been trained by injected labeled data corresponding to each of a plurality of sourcetypes of interest, and wherein the investigative classifier machine learning model subsequently has been trained incrementally in a plurality of additional training cycles with additional data labeled by a semi-supervised learning process, wherein the performing the investigative classification includes performing one of; when vehicle presence sensor (VPS) data are available and satisfies a reliability criterion, selecting a first machine learning model to use as the investigative classifier machine learning model in the performing the investigative classification, the first machine learning model having been trained on a feature set that is based on VPS data, or when VPS data are not available or fails to satisfy the reliability criterion, selecting a second machine learning model to use as the investigative classifier machine learning model in the performing an investigative classification, wherein the second machine learning model has been trained on a feature set that is not based on VPS data; performing a source type classification on the feature set by applying the feature set to a source type classifier machine learning model, to classify a source type of a radiological source associated with the vehicle; and generating an alert indicative of the investigative classification and the source type classification; and causing the alert indicative of the investigative classification and the source type classification to be output to a user.

46. The machine-readable program storage medium of claim 45, such that the vehicle context information comprises a vehicle type of the vehicle, motion of the vehicle as a function of time associated with a passage of the vehicle through the portal, and a time history of output signals from one or more detectors.

47. The machine-readable program storage medium of claim 45, such that the vehicle context information comprises data indicative of a background suppression profile of the vehicle.

48. The machine-readable program storage medium of claim 45, such that the investigative classifier machine learning model has been trained on training data that include motion profile information and vehicle type information associated with a plurality of vehicles.

49. The machine-readable program storage medium of claim 48, such that the motion profile information is based on historical distributions of motion of vehicles through the portal.

50. The machine-readable program storage medium of claim 49, such that the investigative classifier machine learning model has been trained on training data that include injection of a plurality of source types into each of a plurality of vehicle types.

51. The machine-readable program storage medium of claim 45, such that said process further comprises: determining a level of uncertainty associated with the source type classification by applying the source type classification to a source type uncertainty classifier machine learning model, wherein the source type classification is deemed valid only when the level of uncertainty satisfies a specified criterion.

52. The machine-readable program storage medium of claim 45, such that said process further comprises: performing an anomaly classification on the feature prior to performing the investigative classification, by applying the feature set to an anomaly classifier machine learning model to produce an anomaly score; wherein the investigative classification is performed only when the anomaly score is below a specified threshold, further comprising: accessing vehicle presence sensor (VPS) data associated with the vehicle; determining a vehicle type of the vehicle based on the VPS data; identifying a background suppression template that corresponds to the vehicle based on the determined vehicle type of the vehicle; andwherein using the investigative classifier machine learning model comprises using the identified background suppression template to subtract expected background radiation during the extracting of the feature set.

53. The machine-readable program storage medium of claim 45, such that the investigative classifier machine learning model has been trained from contamination and nuclear materials (CNM) training data generated in part by using CNM source model injection with selected samples from the CNM training data that do not overlap with previous training data, to train the investigative classifier machine learning model.

Images