Paper 1 — Research Plan
Proposed title
RAMAL-EBX: Rapid Machine-Learning Dose Prediction for Shielding of 5–7.5 MeV Industrial X-Ray LINAC Facilities Using PHITS-Generated Data
Document relationship
This folder is one research package. Its evidence and decisions flow in one direction:
review protocol → searches and screening → evidence review
→ research plan → proposed methodology
The Evidence Review Protocol defines the
prospective search and screening method. The Evidence Review
records the accessible-source results and scope decision, while the
Annotated Bibliography and Literature Review
retain source-level annotations and the narrative synthesis. This plan defines
the proposed study, and the Proposed Methodology defines how
it would be performed. A located source does not by itself establish a decision
or validate the proposed model.
Lawfully retained source files live under references/industrial-eb-x-ray-dose-prediction/files/.
The project owner's docs/human/paper1.md remains
authoritative for the broader research intent. The accessible-source evidence
review supports the narrower Paper 1 scope, and the project owner approved that
scope as D1 on 2026-09-04. D2 remains approved. D4 was amended on
2026-09-11 to make a vertically downward beam the primary room
configuration while retaining horizontal delivery as an application-supported
alternative outside the primary Paper 1 family. D3 and D5 were reopened
on 2026-09-10 after the owner required industrial radiation-processing
representativeness; their earlier compact circular and static-source decisions
are retained only as historical component-benchmark definitions.
D6 approves a conditional JENDL-5 photonuclear reference, ENDF/B-VIII.1 and
built-in-model comparisons, and neutron retention in the primary combined field.
Data installation, isotope coverage, and component verification remain required.
The owner has confirmed that institutional Scopus, Web of Science, INSPEC, and
IEEE full-text access is unavailable. This is a declared coverage limitation
rather than a blocker. Resolution of the review's awaiting-full-text records
remains outstanding.
Background
Industrial X-ray irradiation facilities generate bremsstrahlung by directing an electron beam onto a high-atomic-number converter. Their shielding calculations must account for the converter, photon energy spectrum and angular distribution, beam power, shielding materials, room geometry, scattered radiation, penetrations, and possible photoneutron production.
Published studies address industrial 5 MeV and 7.5 MeV X-ray converters, source spectra, concrete attenuation, maze scattering, photoneutron production, PHITS benchmark calculations, and machine-learning surrogates for related radiation-transport problems.
Research problem
Repeated Monte Carlo calculations are required when incident-electron energy, room dimensions, barrier thicknesses, or calculation location changes. Deep-penetration shielding calculations can require substantial computation time because only a small fraction of particle histories contribute to dose estimates outside thick barriers.
A machine-learning surrogate could reduce the time required for repeated preliminary calculations. Its predictions must nevertheless be evaluated against PHITS reference calculations, and final shielding designs must still be confirmed using radiation-transport calculations and professional assessment.
Research gap
The accessible literature does not establish a PHITS-based machine-learning surrogate for the complete conventional-LINAC source-to-shield problem over a specified range of 5–7.5 MeV industrial X-ray shielding configurations. The proposed contribution is this combination of electron-to-bremsstrahlung conversion, facility shielding, PHITS reference calculations, and prediction for complete shielding configurations excluded from the training data.
The contribution is not the individual use of PHITS, machine learning, photon-spectrum prediction, or industrial X-ray shielding. Each already exists in related work.
Aim
To develop and evaluate a machine-learning surrogate for rapid prediction of a specified dose quantity in 5–7.5 MeV industrial X-ray LINAC shielding configurations using PHITS reference calculations.
Research objectives
- Define a reproducible PHITS model for the selected electron beam, converter, shielding materials, geometries, and dose quantity.
- Verify the PHITS model against applicable benchmark calculations and validate it against relevant measurements where available.
- Determine the effects of incident-electron energy, room dimensions, and primary and secondary barrier thicknesses on the specified dose quantity for the fixed D3 source and D4 materials.
- Assess the neutron contribution and its sensitivity to the photonuclear data library and reaction model, particularly near
7.5 MeV. - Generate a traceable simulation Dataset with recorded calculation conditions and statistical uncertainties.
- Develop a surrogate model and compare it with interpolation and conventional regression.
- Evaluate prediction error, classification relative to any predefined design dose criterion, range of applicability, and computation time relative to PHITS.
Research questions
RQ1 — Overall performance
Can a machine-learning surrogate trained on PHITS reference calculations accurately predict the specified dose quantity for 5–7.5 MeV industrial X-ray LINAC shielding configurations excluded from the training data, while requiring substantially less computation time than PHITS?
RQ2 — PHITS verification and experimental validation
How well do the PHITS source and shielding calculations agree with applicable benchmark calculations and measurements of bremsstrahlung spectra, angular distributions, and shielding attenuation?
RQ3 — Dependence on shielding parameters
How do incident-electron energy, room width, length and height, and primary and secondary barrier thicknesses affect the specified dose quantity within the fixed D3 source and D4 material contract?
RQ4 — Photoneutron contribution
What is the neutron contribution to the specified dose quantity, particularly near 7.5 MeV, and how sensitive is the calculated contribution to the photonuclear data library and reaction model?
RQ5 — Prediction error and design-criterion classification
What are the frequency and magnitude of dose underestimation and overestimation, and how often do these errors cause incorrect classification relative to any predefined design dose criterion?
RQ6 — Comparison with simpler methods
Does the machine-learning surrogate improve prediction accuracy or computation time relative to interpolation and conventional regression over the same parameter ranges?
RQ7 — Computational performance
What reduction in computation time does the surrogate provide relative to PHITS when inference time, PHITS runtime, training-data generation, model training, and hardware are reported separately?
Scope
| Aspect | Included |
|---|---|
| Application | Radiation shielding for industrial irradiation facilities |
| Accelerator | Conventional linear accelerator |
| Source mode | D3/D5 amendment target: broad 100 cm x 10 cm planar Ta-water-steel converter; normally incident circular Gaussian local spot with 2 cm FWHM and declared 3 cm support; normalized one-dimensional scan-position distribution along the 100 cm direction; a uniform continuous line scan approximated by trapezoidal quadrature is the initial explicit reference assumption |
| Converter | D3 amendment target: planar rectangular 100 cm x 10 cm natural-Ta active plate with explicit water and Type 304 stainless-steel downstream layers; final Ta, coolant, and cover thicknesses require the revised component and sensitivity Study before acceptance |
| Incident-electron energy | Declared monoenergetic value in 5–7.5 MeV |
| Reference method | PHITS electron and photon transport, with photonuclear and neutron assessment |
| Predicted quantity | D2-approved three-dimensional ambient dose equivalent $H^*(10)$ field, PHITS -200, intrinsically normalized in pSv/primary-source-electron; operational mSv/h is derived separately |
| Geometry | D4-approved rectangular closed room: W=4–10 m, L=6–16 m, H=3–6 m; centred source with +y vertical, beam downward along -y, scan along x, and the second horizontal/conveyor axis along z; 100–180 cm primary floor; one shared 50–150 cm thickness for roof and four side walls. Horizontal beam delivery remains available in RAMAL-EBX but is outside the primary Paper 1 geometry family. |
| Shielding medium | Fixed PNNL ordinary NBS-03 concrete at 2.35 g/cm³; interior and exterior PNNL dry air at 0.001205 g/cm³ |
| Intended use | Preliminary calculations and evaluation of shielding alternatives followed by PHITS confirmation |
The PHITS reference model begins with primary electrons at the upstream converter face and includes the converter and all secondary particles. A converter phase-space source is permitted only after verification against the coupled calculation and normalization to the original primary-electron count.
Limitations
Unless separately investigated and supported, Paper 1 does not cover:
- direct-electron-beam irradiation;
- Rhodotrons, Dynamitrons, or other non-LINAC accelerator types;
- incident-electron energies outside
5–7.5 MeV; - scanner laws, widths, spot shapes, or current weightings other than the fixed and verified D3/D5 source archetype;
- product absorbed dose;
- shielding materials or geometries outside the specified parameter ranges;
- untested maze or penetration arrangements;
- activation or dose after shutdown;
- experimental validation of configurations for which no relevant measurement is available;
- regulatory approval or commissioning surveys; or
- replacement of PHITS calculations and professional shielding assessment.
Proposed methodology
The complete protocol is maintained in the Proposed Methodology.
Its staged design is:
- complete a reproducible evidence review and confirm the Paper 1 scope;
- approve the physical and numerical specification;
- verify component transport and limiting cases;
- perform preliminary PHITS calculations, verify the existing one-dimensional profile workflow, and predefine inclusion criteria;
- validate the physical model against applicable measurements where available;
- generate PHITS reference calculations over the approved domain;
- assign complete configurations to training, model-selection, and untouched final-test partitions;
- compare interpolation, conventional regression, and surrogate candidates;
- finalize the selected Model before final-test evaluation;
- report overall and subgroup error, underprediction, overprediction, classification relative to any predefined design dose criterion, and computation time; and
- compare selected difficult cases and the largest underpredictions with independent PHITS calculations.
Each PHITS Result must retain the code version, input, material compositions and densities, physics settings, nuclear-data libraries, source definition, normalization, tallies, conversion coefficients, variance reduction, histories, random seed, and statistical uncertainties. Each scientific parameter must be identified as a literature-supported fact, inference, or study-design choice. Published values from one facility must not be treated as universal. Derived machine-learning data must preserve Result traceability, and an unscored tally must not be treated as zero dose.
Paper 1 reporting will reuse the primary three-dimensional VTK field rather than create a second scientific Result format. RAMAL-EBX already supports axis-aligned one-dimensional profiles with voxel values and relative errors. Before final Dataset generation, Analyze must also expose publication-ready XY, XZ, and YZ slices from that same immutable field, with selectable physical slice position, photon/neutron/combined pages, logarithmic or linear colour scale, comparable fixed ranges across Cases, relative-error maps, and image plus tabular export carrying Case and Result provenance. A sectioned 3D viewport is useful inspection, but is not sufficient evidence for the paper by itself. Analyze now provides the first reusable XY, XZ, and YZ voxel-slice view over a verified Result, including physical slice selection, particle-page reuse, linear/log10 values, and the aligned relative-error field. This is sufficient for exploratory inspection of the preliminary calculation. D7 and publication or cross-Case map comparison remain blocked until fixed comparable colour ranges and provenance-bearing image and tabular exports are implemented and verified.
The first such calculation is recorded in
Preliminary Results. Its 2026-09-10 endpoint-spill
Result completed successfully and verified the Case-to-Result pipeline, but it
is explicitly excluded from D3 acceptance, D7 acceptance, and the production
Dataset. It establishes the next paired no-spill comparison and Q21/Q41/Q81
source-quadrature study; it does not establish converter efficiency or final
statistical thresholds.
Decisions still required
- Set grid, statistical, reportable-coverage, repeated-run, completion, and 1D/2D reporting criteria from preliminary PHITS calculations, including the predefined profile lines, slice planes, coordinates, colour ranges, and RSE presentation used for cross-Case comparison.
- Define exact verification geometries, reference values, and comparison criteria.
- Identify applicable measurement evidence and its uncertainty.
- Predefine Dataset grouping and training, model-selection, and final-test partitions.
- Predefine surrogate metrics and acceptable performance by subdomain.
Expected contributions
- A PHITS reference model verified against applicable benchmarks, with validation against relevant measurements where available.
- A traceable simulation Dataset containing source, geometry, particle, dose, normalization, and statistical-uncertainty information.
- A rapid surrogate evaluated using complete shielding configurations excluded from training and compared with simpler prediction methods.
- A defined range of applicability stating where surrogate predictions may be used and where a PHITS calculation is required.