About This Project
We are developing a physics-informed neural network (PINN) to reconstruct neutron sources from extremely low detector counts. By embedding the detector response and imaging physics directly into the model, we expect the PINN will be able to recover source structure in regimes that traditional methods fail. This project will validate the approach using Geometry and Tracking v4 (GEANT4) nuclear physics simulations, creating a faster reconstruction than currently available methods.
Ask the Scientists
Join The DiscussionWhat is the context of this research?
The project develops a physics-informed neural network (PINN) to reconstruct neutron source images in extremely low-count environments—situations where detectors record very few neutron events, making traditional imaging methods unreliable. Neutron imaging is important for nuclear safety, security, and nonproliferation because neutrons can penetrate shielding and reveal the presence of special nuclear materials [1]. However, prior field and laboratory systems have shown that current detectors often capture only sparse, noisy data, producing low-quality images [2],[3]. This project builds on that work by using a PINN that incorporates detector physics and detector space to image space inversion directly into its training process. This approach enables reliable source reconstruction even with very limited data. The goal is to move neutron imaging closer to practical use in real-world security scenarios by drastically reducing image computation time.
What is the significance of this project?
This project improves our ability to detect and locate materials that emit neutrons, which is crucial for nuclear security, nonproliferation, and emergency response. Current neutron imagers often fail when only a few neutron events are detected, producing blurry or unreliable images. The PI previously built a system designed for these low-signal conditions and published the results [1], but the imaging method required over 100 iterations per image. This work replaces that slow process with a physics-informed neural network (PINN) that uses built-in physical models of how neutrons interact with the detector. In simple terms, PINN can form a clearer image from very little data and do so much faster. This matters in situations such as scanning cargo at a port of entry or quickly assessing a suspicious radiation source during an emergency. The project will deliver an open-source PINN reconstruction tool and a benchmark simulation dataset that others can use to advance neutron imaging.
What are the goals of the project?
The goal of this project is to develop a physics-informed neural network (PINN) that can accurately reconstruct neutron source distributions from extremely low-count detector signals. The objectives are to (1) design and train a PINN that embeds detector-response physics directly into its loss function, (2) demonstrate that it outperforms traditional reconstruction methods in sparse-data regimes, (3) validate its performance using high-fidelity Monte Carlo–simulated detector data, and (4) release an open-source reconstruction framework. The end goal is to enable faster, more reliable neutron source localization for nuclear security and emergency response.
Budget
The requested items directly support completion of the research. I have already conducted preliminary tests using a free online GPU service to verify feasibility, but that platform does not provide the computational resources required for full-scale training and analysis. I selected the least expensive hardware that still meets the project’s needs to avoid placing unnecessary burden on supporters. The open-access publication cost is estimated based on my plan to submit to Nuclear Instruments and Methods in Physics Research Section A, which charges an Article Publishing Fee of USD 2,590 before taxes.
Endorsed by
Project Timeline
Over the course of the project, we will generate simulated neutron detector data, develop and train a physics-informed neural network (PINN) for source reconstruction, and validate its performance against detailed simulations. We will then release the open-source PINN framework along with documented tools and an open-access publication to ensure full reproducibility. Backers will receive early access to the paper draft as well as the trained AI model and the testing data.
Nov 26, 2025
Project Launched
Jan 26, 2026
Generate high-fidelity simulated neutron detector data.
Jan 26, 2026
Develop PINN architecture and set up training pipeline.
Mar 30, 2026
Train and optimize model on simulated data.
May 25, 2026
Validate model performance and compare to baseline methods.
Meet the Team
Team Bio
Adam Glick, Ph.D., is a nuclear physicist and software engineer with expertise in AI, Python, and Monte Carlo simulations for nuclear imaging and radiation therapy. Miles is an electrical engineer experienced in radar algorithms, multithreading, and real-time systems. Mustapha is a software engineer specializing in embedded systems, computer vision, and radiation detection. Most prior work is classified, but all bring deep technical and research expertise.Adam Glick
Adam Glick is a software engineer and nuclear physicist with over a decade of software development experience and more than three years as a professional engineer at PeopleTec, IERUS Technologies, TEAL Systems, and ScienceLogic. He specializes in AI for IT operations (AIOps), cloud technologies (AWS, Azure), and high-performance Python development for large-scale data acquisition and analysis, including memory-efficient parallel processing of terabytes of data.
He earned a Ph.D. in Nuclear Engineering from UC Berkeley, developing neutron detection and imaging systems to characterize cosmogenic background radiation using Monte Carlo simulations and Python-based acquisition software. His postdoctoral work at MD Anderson Cancer Center focused on preclinical FLASH radiation therapy, creating Python tools for data analysis, hardware communication, and radiation transport simulation with his Geant4Py toolkit.
Adam has led teams and mentored engineers as Scrum Master and team lead at IERUS Technologies, overseeing backend architecture, unit testing, and agile process adherence. He has also optimized industrial control systems at TEAL Systems and automated event engines at ScienceLogic. Independently, he has developed mathematical models to optimize cancer therapies using Python and advanced algorithms.
His research has been published in Nuclear Instruments and Methods in Physics Research Section A, Discrete and Continuous Dynamical Systems, and International Journal of Radiation Oncology. Proficient in Python (PyQT, WX, SciPy, NumPy, Pandas), C++, SQL, MATLAB, and DevOps tools, he delivers high-availability software integrating complex hardware and data systems.
Adam’s expertise in nuclear physics, AI/ML, and software engineering enables him to address complex scientific and technical challenges, bridging computational innovation with practical applications in nuclear security, medical physics, and industrial automation.
Examples of past work:
https://www.researchgate.net/p...
Miles O'Brien
Miles is an Electrical Engineer with a B.S. in Electrical Engineering from Missouri University of Science and Technology. He led the modernization of a 3-million-line codebase and developed radar signal processing algorithms, incorporating multithreading and machine learning techniques. Miles has contributed to the integration of radar models, real-time code development, and the creation of automated calibration tools. While most of his prior work has been on classified projects and cannot be publicly shared, his extensive experience in programming (C++, MATLAB, Python), algorithm development, and cross-functional collaboration uniquely positions him to successfully advance this project.
Mustapha Saad
Mustapha is a software engineer specializing in embedded systems. He has over 6 years of professional experience ranging from semiconductor fabrication and design, signal and image reconstruction, computer vision, and embedded systems. The past 5 years he has been employed by Northrop Grumman Corporation and has worked in semiconductor process refinement, computer vision, cloud computing, command and control software, and embedded software. Prior to working at Northrop Grumman Corporation, he interned at Lawrence Berkeley National Laboratory for a year and worked on the design and simulation of charge-sensitive preamplifiers for low-noise applications and writing software to automatically characterize the energy-dependent angular response of 4π radiation detection systems.
Examples of past work:
Additional Information
All data for this project will be generated via high-fidelity Monte Carlo simulations, ensuring realistic detector response without physical experiments. Funding will support GPU resources for model training and open-access publication. The project will produce an open-source PINN framework for neutron source reconstruction, enabling future application to real detector systems while ensuring reproducibility and transparency. The PI has professional experience in AI/ML development and has successfully completed prior AI-driven research projects. This project will leverage that expertise to reconstruct neutron source distributions from simulated data using physics-informed neural networks, ensuring feasibility, rigor, and efficient use of computational resources.
Project Backers
- 2Backers
- 22%Funded
- $1,100Total Donations
- $550.00Average Donation
